sensors Article

A Compact Energy Harvesting System for Outdoor Wireless Sensor Nodes Based on a Low-Cost In Situ Photovoltaic Panel Characterization-Modelling Unit Diego Antolín, Nicolás Medrano *

ID

, Belén Calvo and Pedro A. Martínez

Group of Electronic Design (GDE), Aragón Institute for Engineering Research (I3A), Departamento de Ingeniería Electrónica y Comunicaciones, Universidad de Zaragoza, C/ Pedro Cerbuna 12, Zaragoza 50009, Spain; [email protected] (D.A.); [email protected] (B.C.); [email protected] (P.A.M.) * Correspondence: [email protected]; Tel.: +34-876-55-33-58 Received: 26 June 2017; Accepted: 2 August 2017; Published: 4 August 2017

Abstract: This paper presents a low-cost high-efficiency solar energy harvesting system to power outdoor wireless sensor nodes. It is based on a Voltage Open Circuit (VOC) algorithm that estimates the open-circuit voltage by means of a multilayer perceptron neural network model trained using local experimental characterization data, which are acquired through a novel low cost characterization system incorporated into the deployed node. Both units—characterization and modelling—are controlled by the same low-cost microcontroller, providing a complete solution which can be understood as a virtual pilot cell, with identical characteristics to those of the specific small solar cell installed on the sensor node, that besides allows an easy adaptation to changes in the actual environmental conditions, panel aging, etc. Experimental comparison to a classical pilot panel based VOC algorithm show better efficiency under the same tested conditions. Keywords: maximum power point tracking; solar panel; artificial neural network; PV modelling; energy harvesting; wireless sensor network; cyber-physical systems

1. Introduction Energy harvesting (EH) is a main research field in cyber-physical systems reliability, especially in portable and unattended environmental sensing applications, such as wireless sensor networks (WSN) deployed on extensive outdoor areas monitoring ambient conditions as temperature, light, pressure, and humidity or air pollutants. Figure 1 shows a typical power consumption profile for a sensor node within a WSN intended for environmental monitoring. Most of the time the nodes are set to a low power state (sleep mode in Figure 1), being periodically wake up to acquire the desired environmental parameters, which are then conveyed to the network host. The slow variation of environmental conditions allows programming sleep times of tens of minutes, while the power-hungry measurement and radio frequency (RF) transmission processes take a few seconds or less. Depending on the network configuration, some of these wireless nodes can be configured to receive data from the neighbouring devices, thus extending the surveyed area using intermediate routers for data transmission in a multi-hop architecture, with an additional cost in energy due to the RF transmission. According to the current energy available in the node, a suitable power management requires the use of an optimized cross-layer framework [1] for a precise adjustment of several main parameters in the sensor node as duty cycle, RF power transmission, node data routing and energy harvesting from the environment. Actually, a suitable node sensor power management together with an efficient energy harvesting technique is critical to increase the sensor network autonomy while reducing the required maintenance operations.

Sensors 2017, 17, 1794; doi:10.3390/s17081794

www.mdpi.com/journal/sensors

Sensors 2017, 17, 1794 Sensors 2017, 17, 1794

2 of 18 2 of 18

these WSNmonitoring monitoringapplications applications in in natural natural areas, to to In In these WSN areas, reviewing reviewingthe thedifferent differenttechniques techniques harvest energy fromthe theenvironment, environment,solar solar energy energy is ofof itsits availability harvest energy from is the the preferential preferentialchoice choicebecause because availability and predictability [2–6]. The solar power is converted to electric energy through low-cost and lowand predictability [2–6]. The solar power is converted to electric energy through low-cost and size photovoltaic (PV) panels. Thus, to successfully design an efficient EH system, one first key issue low-size photovoltaic (PV) panels. Thus, to successfully design an efficient EH system, one first having an accurate PV model to reliably predict the panel behaviour under the expected working keyis issue is having an accurate PV model to reliably predict the panel behaviour under the expected conditions, with the objective of optimizing the energy transfer from the panel to the energy storage working conditions, with the objective of optimizing the energy transfer from the panel to the energy device, by selecting the suitable electric load connected to the panel as a function of the maximum storage device, by selecting the suitable electric load connected to the panel as a function of the generated power, which in turn depends on the physical solar cell characteristics and ambient maximum generated power, which in turn depends on the physical solar cell characteristics and working conditions. ambient working conditions.

Figure 1.1.Typical for an an environmental environmentalwireless wirelesssensor sensornode. node. Figure Typicalpower powerconsumption consumption profile profile for

The behavioural description description ofoflarge-size high-power PV panels and PV arrays typically based The behavioural large-size high-power PV panels and PV is arrays is typically on simulation/modelling algorithms using Matlab and/or Simulink [7–10], specific high level based on simulation/modelling algorithms using Matlab and/or Simulink [7–10], specific high level mathematical simulation models [11,12], or electrical–based models [13] that extract the required mathematical simulation models [11,12], or electrical–based models [13] that extract the required parameters from themanufacturer manufacturerdatasheet, datasheet, from from experimental experimental PV data (this one parameters from the PVcharacterization characterization data (this one being the optimum choice) or from a combination of both. A PV panel characterization system must being the optimum choice) or from a combination of both. A PV panel characterization system measure the power-to-voltage characteristic under different irradiance and temperature values, at must measure the power-to-voltage characteristic under different irradiance and temperature values, different working conditions obtained by changing the electric load connected to the panel. Thus, a at different working conditions obtained by changing the electric load connected to the panel. Thus, complete PV panel characterization station requires the use of bulky and expensive instruments a complete PV panel characterization station requires the use of bulky and expensive instruments (solar simulators, pyranometers, programmable power loads), so that these infrastructures are only (solar simulators, pyranometers, programmable power loads), so that these infrastructures are only worthy for large power photovoltaic panels. worthyIn for large power photovoltaic panels. the case of low-cost small PV cells, such as those usually used in wireless sensor nodes, models In the case of low-cost are small PV cells, such as those usuallywithout used inanwireless nodes, provided by manufacturers limited to an approximate behaviour accurate sensor description, models provided by manufacturers are limited to an approximate behaviour without an accurate thus jeopardizing the related efficiency in the power management and transmission system. To description, jeopardizing related efficiency in the power management transmission system. overcome thus this limitation, thisthe work first introduces a low-cost and compact PVand panel characterization To system, overcome this limitation, this work first introduces a low-cost and compact PV panel characterization designed to be included as a part of the sensor node. The acquired experimental data are then system, designed to abePV included as a part based of the on sensor node. The acquired experimental data are then used to complete panel modelling the use of artificial neural networks (ANN). With used complete ait PV panel modelling based on artificial neural (ANN). With itthis thistoapproach, is possible to characterize in the situuse theofphotovoltaic cell,networks and re-characterizing approach, it is possible to characterize in situ the photovoltaic cell,working and re-characterizing it periodically, periodically, upgrading the panel model according to the actual parameters, and including effects of efficiency lost due to aging, solar radiation variations related to seasonality dust of upgrading the panel model according to the actual working parameters, and including or effects deposition theaging, activesolar area.radiation The finalvariations goal is the implementation of aormaximum power over pointthe efficiency lost over due to related to seasonality dust deposition tracking based Voltage Open Circuit algorithm thatpoint models the solar panelbased to active area.system The final goalonisathe implementation of(VOC) a maximum power tracking system reproduce its open-circuit behaviour under the real environmental working conditions in order to on a Voltage Open Circuit (VOC) algorithm that models the solar panel to reproduce its open-circuit speed up the load of an energy storage system in low form factor wireless sensor nodes, maintaining behaviour under the real environmental working conditions in order to speed up the load of an energy as objectives low cost and minimization of visual storage system maximum in low formefficiency, factor wireless sensor nodes, maintaining as impact. objectives maximum efficiency,

low cost and minimization of visual impact.

Sensors 2017, 17, 1794

3 of 18

Sensors 2017, 17, 1794

3 of 18

The The paper is organized as follows: Section 2 presents thethe compact paper is organized as follows: Section 2 presents compactand andportable portablePV PVpanel panel characterization system designed for locally accomplish parameter extraction in low-cost solar panels, characterization system designed for locally accomplish parameter extraction in low-cost solar utilizing a small form-factor device that wireless sensor node. Section addresses the3 panels, utilizing a smallPV form-factor PVpowers devicea that powers a wireless sensor 3node. Section modelling of the photovoltaic panel, with a multilayer neural network as neural fitting tool usingas addresses the modelling of the photovoltaic panel,perceptron with a multilayer perceptron network the data collected fromthe the previous characterization system. Section 4 presents the energy harvesting fitting tool using data collected from the previous characterization system. Section 4 presents the energy harvesting system for the sensor nodebased mounting panel,boost basedconverter on a DC-DC boost converter system for the sensor node mounting the panel, on athe DC-DC with a maximum with a maximum power point tracking VOC that relies on the developed power point tracking VOC algorithm that relies onalgorithm the developed PV modelling approachPV to modelling track the to track the open circuit voltage. Experimental resultsof confirm the achievement of better openapproach circuit voltage. Experimental results confirm the achievement better efficiency compared to a efficiency compared to a classical VOC technique based on the use of a pilot panel, thus allowing classical VOC technique based on the use of a pilot panel, thus allowing faster charge of the energy faster charge of the energy storage device. storage device. In Situ Photovoltaic Panel Characterization 2. In 2. Situ Photovoltaic Panel Characterization The proposed panel characterizationsystem system follows follows the approach, adapted for lowThe proposed PV PV panel characterization theclassical classical approach, adapted for cost solar solar panels panels and allows one to to characterize thethe PVPV cell low-cost and itit owns ownsthe thecharacteristic characteristicofofportability: portability:it it allows one characterize in the same site where the device is deployed, thus providing a local accurate behavior model that cell in the same site where the device is deployed, thus providing a local accurate behavior model that considers theexternal real external parameters affecting thepanel PV panel operation (light, temperature), as well considers the real parameters affecting the PV operation (light, temperature), as well as as the real voltage and current provided by the panel at different programmable loads. the real voltage and current provided by the panel at different programmable loads. It consists of three main elements (Figure 2): a compact variable load connected to the device It consists of three main elements (Figure 2): a compact variable load connected to the device under test; a control and acquisition system that selects the load value connected to the PV cell at under test; a control and acquisition system that selects the load value connected to the PV cell at each moment and reads the data provided by the PV cell at actual temperature and light conditions; each moment and reads the data provided by the PV cell at actual temperature and light conditions; and the system software, which coordinates the full system operation. The PV panel selected is a MCand the system software, which coordinates the full system operation. The PV panel selected is a SP0.8-NF-GCS from Multicomp (Premier Farnell, Leeds, UK); its main parameters are: 800 mW MC-SP0.8-NF-GCS from Multicomp (Premier Farnell, Leeds, UK); its main parameters are: 800 mW maximum power, 4.8 V open circuit voltage and a temperature operation range from −40 °C to +85 °C. maximum power, 4.8 V open circuit voltage and a temperature operation range from −40 ◦ C to +85 ◦ C.

Figure 2. Proposed PV PV panel characterization system. Figure 2. Proposed panel characterization system.

Sensors 2017, 17, 1794

4 of 18

2.1. Variable Load Module The custom variable load consists on a set of 24 fixed value resistors (Table 1), which can be selected by means of three 1:8 demultiplexers that enable or disable a set of ultra-low on-resistance MOSFET transistor switches in series with the loads, connecting the PV panel to each selected resistor at a time (Figure 2). The selection of resistor values attends to the PV panel behaviour, covering from near open circuit for high resistors, to short circuit for the lowest resistor value. Table 1. Resistor values in the variable load module. SEL(1:3)

RMUX1 (Ω)

RMUX2 (Ω)

RMUX3 (Ω)

000 001 010 011 100 101 110 111

10,000 8200 7500 6200 5100 4300 3300 2200

1800 1200 1000 820 680 430 220 130

62 33 16 8.2 3.9 2 1 0.1

Their overall real values have been measured using a Keysight (Santa Rosa, CA, USA) 34410A 6 12 digital multimeter (DMM) using four-wire technique. In this way, the accurate knowledge of the resistor values allows estimate the value of the power transferred to the load at a certain voltage without requiring the current measurement, simplifying the data acquisition. The operation of this module is controlled by a low-cost microcontroller that selects the panel load and reads the measurements, locally storing the values before being sent to a host system for its processing. Note that a finer load variation resolution can be programmed. However, this possibility added high complexity at computing and managing level, while a very low increase of accuracy in the panel modeling was obtained and was therefore withdrawn. 2.2. Control and Acquisition System The core of this block is a 16 bits MSP430FR5739 microcontroller from Texas Instruments (Dallas, TX, USA). It selects at each moment the resistor to be connected from the load array to the PV module and acquires the main variables involved in the photovoltaic panel characterization: voltages in the panel output, temperature and illuminance. The microcontroller clock rate is set to 24 MHz, letting assume that both temperature and illuminance keep constant while sweeping the 24 resistors in the load module, thus simplifying the measurement process. In addition, this microcontroller includes 10-bit analog-to-digital (ADC) converters. Environmental temperature is measured by means of a TMP112 digital sensor (Texas Instruments, Dallas, TX, USA). Its measurement range spans from −40 ◦ C to +125 ◦ C, which covers the temperature modelling range (−10 ◦ C to +70 ◦ C) at a suitable accuracy (from 0 ◦ C to 65◦ C the sensor accuracy is ±0.5 ◦ C, so sensor readings keep this accuracy in almost the full working range, only giving some additional (acceptable) error at the upper and lower limits) and its bias voltage is compatible with the microcontroller requirements. Considering the measurement of incident solar light, systems designed for industrial solar panel characterization monitor irradiance using pyranometers. Irradiance is a physical magnitude that measures incident solar power per unit area and considers all incident wavelengths. However, pyranometers are bulky and expensive instruments, what make them unsuitable for compact and low-cost systems. Hence, the proposed system incorporates a compact and low-cost MAX44009 illuminance sensor from Maxim Integrated (San Jose, CA, USA). Unlike irradiance, illuminance is a photometric parameter which depends on the human eye sensitivity to colors and has no proportional relation to irradiance. In local measurements, light spectral composition can be assumed as constant

Sensors 2017, 17, 1794

5 of 18

during the most of the daylight duration and the relation between photometric and ratiometric magnitudes can considered fixed [14]. Therefore, if the energy harvesting system is calibrated Sensors 2017, 17,be 1794 5 of 18 for a specific location, illuminance offers a suitable measurement of the relative incident solar power for calibrated a specificilluminance location, illuminance offers a suitable measurement of 0.045 the relative that place. Thefor proposed sensor presents an input range from lux toincident 188,000 lux, solar power for place. The for proposed illuminance presents an input from 0.045 lux widely covering thethat illuminance an average sunnysensor day. In addition, bias range voltages are compatible to 188,000 lux, widely covering the illuminance for an average sunny day. In addition, bias voltages with those of the microcontroller. Data are provided to the microcontroller in a 12-bits floating point are compatible with those of the microcontroller. Data are provided to the microcontroller in a representation via I2C protocol. 12-bits floating point representation via I2C protocol.

2.3. System Software

2.3. System Software

The The control andand acquisition tobe becontrolled controlled external Before control acquisitionsystem systemisisdesigned designed to byby an an external host.host. Before its its deployment, a first characterization can be performed using a computer as host system, though deployment, a first characterization can be performed using a computer as host system, though after after placement in itsindefinitive location thethe system is is designed nodewhere where it is placement its definitive location system designedtotobe bemanaged managed by by the the sensor sensor node connected. Figure 3 shows dataflow in the microcontroller. First, the microcontroller it is connected. Figurethe 3 shows theimplemented dataflow implemented in the microcontroller. First, the microcontroller configures its peripherals and initializes variables, next goingstate to a low-power state configures its peripherals and initializes variables, next going to a low-power until the host device until the host device sends an interrupt request through the serial port. The first interrupt commands sends an interrupt request through the serial port. The first interrupt commands the microcontroller to the measurement microcontrollerprocess, to run which the measurement process, which starts acquiring temperature andpanel run the starts acquiring temperature and illuminance. Next, the illuminance. Next, the panel output voltage measurements are taken, consecutively selecting the 24 output voltage measurements are taken, consecutively selecting the 24 resistors in the variable load. resistors in the variable load. Considering switching, voltage settling time and ADC conversion time Considering switching, voltage settling time and ADC conversion time using the microcontroller using the microcontroller at a 24 MHz working frequency this process takes less than 20 ms, so that at a 24 MHz workingand frequency this process takes less thanalong 20 ms, that both process. temperature and both temperature illuminance can be considered constant the so measurement illuminance can be considered constant along the measurement process.

Figure 3. System software dataflow. Figure 3. System software dataflow.

Figure 4 shows a photographofof the the characterization characterization board connected to the solarsolar panel.panel. Figure 4 shows a photograph board connected to the Illuminance and temperature readings are acquired from the corresponding sensors located in thein the Illuminance and temperature readings are acquired from the corresponding sensors located upper-right side of the PV panel. The power transmitted by the PV panel is estimated from the output upper-right side of the PV panel. The power transmitted by the PV panel is estimated from the output voltage readings over the 24 selectable resistors in the variable load (which are distributed in the PCB, voltage readings over the 24 selectable resistors in the variable load (which are distributed in the PCB, see Figure 4). Figure 5 shows the output voltage changes in the solar panel due to variations in the see Figure 4). Figure 5 shows the output voltage changes in the solar panel due to variations in the connected load value (green channel). The PV output voltage changes with the resistor load as shown

Sensors 2017, 17, 1794

6 of 18

connected load value (green channel). The PV output voltage changes with the resistor load6as shown Sensors Sensors 2017, 2017, 17, 17, 1794 1794 6 of of 18 18 in this figure, where the non-linear region corresponds to resistor values in the range of few kilo-ohms in where the region to in range of kiloto ohms, in figure, the local irradiance and temperature conditions wherevalues the system in this this figure, where the non-linear non-linear region corresponds corresponds to resistor resistor values in the the was rangetested, of few fewbeing kilo- this ohms to in local and temperature where system was region directly related theirradiance maximum power point conditions area of interest. The full measurement time is ohms to ohms, ohms, in the theto local irradiance and temperature conditions where the the system was tested, tested, being being region directly related to the maximum power point area of interest. The full measurement time giventhis as the time between cursors a and b in Figure 5, shown as the high level in the upper magenta this region directly related to the maximum power point area of interest. The full measurement time is as time between aa and Figure shown as level magenta is given given as the the time between cursors and b b in in Figure 5, 5, shownchannel) as the the high high level in in the the upper magenta signal. Though the first 12 ms cursors the measured voltage (green seems to be upper constant, this is an signal. Though the first 12 ms the measured voltage (green channel) seems to be constant, this is an the first 12 scales; ms the measured (green seems to be constant, an effectsignal. due toThough the oscilloscope the valuesvoltage measured bychannel) the characterization systemthis areisdifferent effect due to the oscilloscope scales; the values measured by the characterization system are different effect due to the oscilloscope scales; the values measured by the characterization system are different and provide valid data for thethe panel modelling. The non-monotonicity non-monotonicity visible as two glitches in the and and provide provide valid valid data data for for the panel panel modelling. modelling. The The non-monotonicity visible visible as as two two glitches glitches in in the the non-linear region of the load signal appears random and occasionally. These errors can be avoided non-linear non-linear region region of of the the load load signal signal appears appears random random and and occasionally. occasionally. These These errors errors can can be be avoided avoided including a small additional delay between each measurement. However,However, because these measurements including a small additional delay between each measurement. because including a small additional delay between each measurement. However, because these these are collected to be are used as training data fortraining an artificial neural network to model the panel, measurements collected to data an neural network to model the measurements are collected to be be used used as as training data for for an artificial artificial neural network to PV model the and PV panel, and considering the robustness of these modelling techniques to local data errors, in order considering the robustness of these modelling techniques to local data errors, in order to ensure PV panel, and considering the robustness of these modelling techniques to local data errors, in order that to environmental (light temperature) are constant we preferred environmental parameters (lightparameters and temperature) constant we preferred thosekeep times as to ensure ensure that that environmental parameters (light and andare temperature) are have constant we have havekeep preferred keep those times as low as possible. Additionally, this measurement times allows using 0.5 W power those times Additionally, as low as possible. Additionally, this measurement times allows using 0.5 W power low as possible. this measurement times allows using 0.5 W power dissipation resistors, resistors, below the maximum power that can (0.8 without integrity dissipation resistors, below thethe maximum power that the the panel can provide provide (0.8 W) W) withoutrisk, integrity belowdissipation the maximum power that panel can provide (0.8panel W) without integrity damage reducing damage risk, reducing cost and size. damage cost and size.risk, reducing cost and size.

Figure 4. Characterization system connected to the thesolar solarpanel panel (control host not shown). Figure (control Figure 4. 4. Characterization Characterization system system connected connected to to the solar panel (control host host not not shown). shown).

Figure Figure 5. 5. Panel Panel output output voltage voltage values values (green) (green) in in the the characterization characterization process process sweeping sweeping resistor resistor values values

Figure 5. Panel output voltage values (green) in the characterization process sweeping resistorthe values in in the the PV PV panel panel load. load. Rising Rising and and descent descent slope slope in in purple purple signal signal indicates indicates the the start start and and end end time time in in the in themeasurement. PV panel load. Rising and descent slopecursor in purple signal indicates the start and end time in the measurement. Time Time span span is is below below 20 20 ms ms (see (see cursor measures). measures). measurement. Time span is below 20 ms (see cursor measures).

Sensors 2017, 17, 1794

7 of 18

After acquisition, data are locally stored waiting for a transmission command from the host. If a new measurement instruction is received from the host, previous data are overwritten. 2.4. Host Control The panel characterization system operation is designed to be controlled by considering as host device the sensor node where the panel is connected after its deployment in the area to be monitored. In this manner, the periodic PV panel characterization can be included as an additional task in the sensor node operation, thus achieving an updated and accurate local model of the solar panel, with an in situ periodical model recalibration using real temperature, illuminance and transferred power. The re-characterization period can be set according to the user requirements, programming its periodicity from the WSN host by means of the corresponding command. In addition, defining warning conditions a non-periodical characterization can be triggered either by the sensor node itself or by the network host. The host operation for the panel data acquisition process is the following: after configuration of the serial port that connects to the characterization system, the host sends a command to start the measurement process and waits for a time t to the end of this operation. In this time (stated 1 s), the characterization system performs the measurements of temperature, illuminance and power transferred to the resistors in the load array as indicated in Figure 3. Next, the host device sends a data transmission command to the control and acquisition system, receiving the data. This process is recurrently executed with a period of T s, which can be modified according to the requirements, for a sufficiently long time to ensure that measures are available with adequate variability in temperature and illuminance to construct a suitable panel model. 3. ANN-Based Photovoltaic Panel Modelling Using the data collected by means of the proposed characterization system, a numerical model of the solar panel used in this work has been developed. The behavioral model obtained is locally valid under the ambient conditions in which measurements have been taken, and can be updated by collecting new measurements with the characterization system previously presented. This choice simplifies the model, allowing its implementation in compact low-power systems where a microcontroller is the computing core. The PV panel model has been developed using an artificial neural network (ANN) as fitting technique. ANNs are suitable tools for system modelling in tasks where enough data are available, and no accurate relationship between input and output magnitudes in the modelled system is known. In fact, ANNs applied to power solar panel modelling have been recently reported in the literature [15–17]. Among the variety of ANN models, the multilayer perceptron (MLP) is the most common in system modelling due to its universal function estimator characteristic in non-linear fitting tasks, thanks to the use of sigmoid (nonlinear) output functions in its processing nodes [18], and the use of advanced parameter adjustment techniques, as the Levenberg-Marquardt algorithm [19]. Data used to perform the PV model were collected for a week. The ANN input and output data were normalized in the range [−1, +1], considering that environmental inputs (temperature and illumination) as well as output voltage are limited to those shown in Table 2, corresponding to the environmental limits in the region where the panel is working, and the useful panel voltage range. Measures have been acquired every minute, i.e., with a period T = 60 s. Table 2. Range of parameters used in the solar panel modelling. Parameter

Upper Limit

Lower Limit

Light (lux) Temperature (◦ C) Voltage (V) Panel load (Ω)

120,000 70 5 10,000

10,000 −10 2.8 0.1

Sensors 2017, 17, 1794

8 of 18

Sensors Sensors 2017, 2017, 17, 17, 1794 1794

88 of of 18 18

Figure 6 shows the experimental panel output voltage as a function of the load resistor for three different environmental conditions of illuminance and temperature and its corresponding simulated Figure 6 shows the experimental panel output voltage as a function of the load resistor for three Figure 6 shows experimental panel output loadMLP, resistor for three values provided by athe Levenberg-Marquardt trainedvoltage 3 inputas– a6 function hidden –of 1 the output giving very different environmental conditions of illuminance and temperature and its corresponding simulated different environmental conditions of illuminance and temperature and its corresponding simulated similar results. values provided by a Levenberg-Marquardt trained 3 input – 6 hidden – 1 output MLP, giving very values provided by a Levenberg-Marquardt trained 3 input – 6 hidden – 1 output MLP, giving very similar results. similar results.

Figure 6. Real and simulated voltage-to-load curves at three different light and temperature conditions. Figure and simulated voltage-to-load curvescurves at threeatdifferent light and temperature conditions. Figure 6.6.Real Real and simulated voltage-to-load three different light and temperature conditions.

Once verified the correct fitting, the MLP has been tested to obtain the power-voltage curves Once verified the correct fitting, the MLP has been tested to obtain the power-voltage curves that characterize the behaviour of a solar panel. For this, the environmental MLP model inputs Once verifiedthe thebehaviour correct fitting, the MLP hasFor been tested obtain the power-voltage curves that characterize of a solar panel. this, the to environmental MLP model inputs (illuminance and temperature) are fixed at the desired values, sweeping the range of the ANN input that characterize the behaviour a solar panel. Forvalues, this, the environmental model (illuminance and temperature) areoffixed at the desired sweeping the rangeMLP of the ANNinputs input that represents the panel load to obtain the voltage values from the ANN output. Power is calculated (illuminance and temperature) are fixed at the desired values, sweeping the range of the ANN input that represents the panel load to obtain the voltage values from the ANN output. Power is calculated as as the output voltage over the corresponding input load value. Figure 7 shows the obtained results. thatoutput represents theover panelthe load to obtain theinput voltage values the7ANN Power results. is calculated the voltage corresponding load value.from Figure showsoutput. the obtained Note Note that in this modelling, the voltage shows a coupled dependence with illuminance and as the voltage over the corresponding input load value. Figure 7 shows theand obtained results. that inoutput this modelling, the voltage shows a coupled dependence with illuminance temperature. temperature. This is because, unlike the models developed in laboratory, where temperature and Note that in this modelling, the voltage shows a coupled dependence with illuminance This is because, unlike the models developed in laboratory, where temperature and irradiance canand be irradiance can be independently controlled, the experimental models based on in-situ environmental temperature. This is because, the models developed laboratory, where temperature and independently controlled, the unlike experimental models based oninin-situ environmental measurements measurements account for coupling in temperature and illuminance (i.e., the more illuminance, the irradiance be independently controlled, the experimental models based onthe in-situ environmental account forcan coupling in temperature and illuminance (i.e., the more illuminance, more temperature). more temperature). Therefore, these simpler models are not general, but locally valid and suitable for measurements account for coupling in temperature and illuminance (i.e., the more illuminance, the Therefore, these simpler models are not general, but locally valid and suitable for its use in the proposed its use in the proposed application: the development of a solar energy harvesting system based on a more temperature). Therefore,ofthese simpler models are not general, buton locally valid and suitable for application: the development a solar energy harvesting system based a local behavioural model local behavioural model to control the maximum power point tracking. its use in the proposed application: the development of a solar energy harvesting system based on a to control the maximum power point tracking. local behavioural model to control the maximum power point tracking.

Figure 7. 7. P-V P-V curves curves obtained obtained from from the the ANN-based ANN-based solar Figure solar panel panel model. model.

Figure 7. P-V curves obtained from the ANN-based solar panel model.

Sensors 2017, 17, 1794 Sensors 2017, 17, 1794

9 of 18 9 of 18

4. Solar Energy Harvesting System Architecture 4. Solar Energy Harvesting System Architecture Figure 8 shows the block diagram of a typical solar-based energy harvesting system. A photovoltaic device converts solar to electrical The direct-current-to-direct-current (DC-DC) Figure 8 shows the block diagram of a typicalenergy. solar-based energy harvesting system. A photovoltaic converter generates from the PV device a stabilized voltage, that feds the powered system, device converts solar to electrical energy. The direct-current-to-direct-current (DC-DC) converter comprising an the energy storage system (battery, and the electronic systemanitself. A generates from PV device a stabilized voltage,supercapacitor), that feds the powered system, comprising energy maximum power transfer circuit (called maximum power point tracking system, MPPT) tracks storage system (battery, supercapacitor), and the electronic system itself. A maximum power transfer environmental and electrical magnitudes from the PV system, modifies in real time the DC-DC circuit (called maximum power point tracking system, MPPT)and tracks environmental and electrical architecture to keep an optimum energy transfer. The monitored variables are selected according to magnitudes from the PV system, and modifies in real time the DC-DC architecture to keep an optimum the implemented MPPT technique. energy transfer. The monitored variables are selected according to the implemented MPPT technique.

Figure 8. Generic solar energy harvesting block diagram.

Figure 8. Generic solar energy harvesting block diagram.

The DC-DC topology depends on the input and output voltages, provided by the solar panel and The by DC-DC topology depends on the input voltages,aprovided the solar panel required the powered system, respectively. The and solaroutput panel provides maximumbyvoltage of 3.85 V. and required by the powered system, respectively. The solar panel provides a maximum voltage of The energy storage device selected is a supercap, due to its simple control electronic requirements; 3.85 V. Thebatteries energy require storagecomplex device control selectedelectronics is a supercap, due to its simple control electronic conversely, to avoid overload and deep discharges that requirements; conversely, batteries require complex control electronics to avoid overload and deep can damage the load device, as well as higher loading currents, that demand higher panels, unsuitable discharges can damage the load as well higher loading currents, thatbe demand higher for the smallthat form factor designs used device, in wireless CPS.asThe selected supercap, as will detailed next, panels, unsuitable for the small form factor designs used in wireless CPS. The selected supercap, as has a maximum input voltage of 5.5 V, suitable for the powered environmental nodes [20]. Under will be detailed next, has a maximum input voltage of 5.5 V, suitable for the powered environmental these conditions, the DC-DC architecture has to be a boost converter. nodes [20]. Under these the converter DC-DC architecture hasWhen to be aswitch boost converter. Figure 9 shows the conditions, boost DC-DC architecture. S1 is ON, the diode is Figure 9 shows the boost DC-DC converter architecture. When switch S1 is in ON, diode isSC, in in reverse mode, thus the sensor node (Rload ) is powered by the energy stored thethe supercap reverse thusL1 theis sensor node (Rloadthe ) isPV powered the energy in the SC, while while themode, inductor energized from output.by When switch stored S1 is OFF, thesupercap output voltage V1 the inductor L1 is energized from the PV output. When switch S1 is OFF, the output voltage from from the PV panel and the energy stored in the inductor load up the supercap voltage (V2),V1 which the PV panel anddirectly the energy stored in the inductor up the supercap voltage (V2), which provides provides energy to the load. Capacitor Cinload filters abrupt changes in the output voltage V1 energy directly to the load. Capacitor C in filters abrupt changes in the output voltage V1 due to due to variations in the working conditions of PV panel. The component values (Table 3) were variations in the working PV panel. The component values (Table 3) were estimated using estimated using classical conditions techniquesof[21,22], assuming the system works in continuous conduction classical techniques [21,22], assuming the system works in continuous conduction mode (CCM) [23]. mode (CCM) [23]. The value value of of the the supercap supercap must must be be selected selected to to provide provide the the required required energy energy to to the the electronic electronic The system (the sensor node) for a minimum operating life, even in the absence of light. In our case, the the system (the sensor node) for a minimum operating life, even in the absence of light. In our case, selected device device is is aa 1.5 1.5 F–5.5 F–5.5 V supercap from from Panasonic Panasonic (Osaka, (Osaka, Japan), Japan), with with electrical electrical selected V EECF5R5U155 EECF5R5U155 supercap characteristics compatible with the low dropout (LDO) regulator used in the wireless nodes characteristics compatible with the low dropout (LDO) regulator used in the wireless nodes presented presented in [20]. The 1.5 F capacitance value guarantee a reduced node lifetime (about 12 h for the in [20]. The 1.5 F capacitance value guarantee a reduced node lifetime (about 12 h for the aforementioned aforementioned node), but has been selected to complete allow a fast and completeofcharacterization of the node), but has been selected to allow a fast and characterization the proposed energy proposed energy harvesting system. The energy storage capacity can be increased just increasing the harvesting system. The energy storage capacity can be increased just increasing the capacitance value. capacitance value.

Sensors Sensors2017, 2017, 17, 17, 1794 1794

10 10of of18 18

Figure 9. Boost DC-DC circuit. In red, a dual solar panel MPPT control block diagram. Figure 9. Boost DC-DC circuit. In red, a dual solar panel MPPT control block diagram. Table 3. Component values for the DC-DC boost converter. Table 3. Component values for the DC-DC boost converter. Component Component Cin Cin L1 L1 Diode Diode Switch Switch

Value Value >25 >25 µFµF >6.7 mH >6.7 mH HSMS-2800 HSMS-2800 IRFML8422 (MOSFET transistor) IRFML8422 (MOSFET transistor)

4.1. Maximum MaximumPower PowerPoint PointTracking Tracking 4.1. The maximum maximumpower powertransfer transferisisgiven givenat at impedance impedancematching, matching,i.e., i.e.,when whenthe the DC-DC DC-DCregulator regulator The presents an input impedance equal to the output impedance of the solar panel. Because the output presents an input impedance equal to the output impedance of the solar panel. Because the output impedance in in the the solar solar panel paneldynamically dynamicallydepends depends on onits itsintrinsic intrinsiccharacteristics characteristicsand andenvironmental environmental impedance conditions, impedance fitting able to to adapt thethe DC-DC topology. For For this, conditions, fittingrequires requiresaadynamic dynamictechnique technique able adapt DC-DC topology. energy harvesting systems makemake use ofuse maximum powerpower point point transfer (MPPT) algorithms [24], which this, energy harvesting systems of maximum transfer (MPPT) algorithms [24], monitor severalseveral different electrical and environmental parameters to come maximize the power which monitor different electrical and environmental parameters to to come to maximize the transfer at all times. power transfer at all times. The Voltage Voltage Open is is a widely used MPPT algorithm that tracks the voltage at which The OpenCircuit Circuit(VOC) (VOC) a widely used MPPT algorithm that tracks the voltage at the power transfer is maximized. Experimentally, it is known that this is proportional to the which the power transfer is maximized. Experimentally, it is known thatvoltage this voltage is proportional system at open according to: to the system at circuit, open circuit, according to: Vmpp = k · Voc (1) = ∙ (1) where Vmpp is the voltage at maximum power point transfer, Voc is the voltage at open circuit (which is the voltage at maximum power point transfer, Voc is the voltage at open circuit (which where Vmpp depends on the incident light and temperature), and k is a constant. The value of k depends on the depends on the incident light and temperature), and k is a constant. The value of k depends on the panel model; its value is typically limited to 0.7–0.8 [25–27], and can be experimentally estimated by panel model; its value is typically limited to 0.7–0.8 [25–27], and can be experimentally estimated by characterizing the solar panel (see Figure 7). In our case, k = 0.8. characterizing the solar panel (see Figure 7). In our case, k = 0.8. Therefore, in the VOC MPPT algorithm the tracking of the open circuit voltage is required. For this, Therefore, in the VOC MPPT algorithm the tracking of the open circuit voltage is required. For there are two different strategies: (1) to add an extra identical solar panel without load to act as a this, there are two different strategies: 1) to add an extra identical solar panel without load to act as a reference or (2) to periodically disconnect the solar panel from the converter and read its voltage at reference or 2) to periodically disconnect the solar panel from the converter and read its voltage at open circuit. Both options present drawbacks: in the first case, an additional cost and increase of open circuit. Both options present drawbacks: in the first case, an additional cost and increase of system size; in the second, the DC-DC is periodically leaved without power for the time the voltage is system size; in the second, the DC-DC is periodically leaved without power for the time the voltage measured, losing efficiency. is measured, losing efficiency. In the dual panel VOC technique (Figure 9), the control system monitors the open-circuit voltage In the dual panel VOC technique (Figure 9), the control system monitors the open-circuit voltage in the reference PV panel (Voc ), estimating the maximum power transfer voltage Vmpp (Equation (1)), in the reference PV panel (Voc), estimating the maximum power transfer voltage Vmpp (Equation (1)), which is compared to the working PV panel voltage V1. When V1 is lower than V , the switch S1 is which is compared to the working PV panel voltage V1. When V1 is lower than Vmpp mpp, the switch S1 is turned off, increasing the load impedance, and thus increasing V1. Otherwise, if V1 > V , the switch turned off, increasing the load impedance, and thus increasing V1. Otherwise, if V1 > Vmpp mpp, the switch

Sensors 2017, 17, 1794 Sensors 2017, 17, 1794

11 of 18 11 of 18

is on, reducing the load impedance. To avoid fast and continuous changes in the switch state, the comparator module presents hysteresis. is on, reducing the load impedance. To avoid fast and continuous changes in the switch state, the comparator module presents hysteresis. 4.2. Maximum Power Point Tracking Technique Based on ANN Modelling 4.2. Maximum Tracking Based on PV ANN Modelling As it was Power stated Point before, the useTechnique of an additional panel to control the switching in the MPPT algorithm presents several drawbacks, mainly cost, size and the need to accurately place both panels As it was stated before, the use of an additional PV panel to control the switching in the MPPT in the same orientation to allow right behaviour comparison. In this work, we propose to replace the algorithm presents several drawbacks, mainly cost, size and the need to accurately place both panels extra solar panel by the model obtained from an in-situ characterization (Figure 10) achieved using in the same orientation to allow right behaviour comparison. In this work, we propose to replace the the methodology shown in Section 3. In this way, the proposed tracking method allows continuous extra solar panel by the model obtained from an in-situ characterization (Figure 10) achieved using energy transfer from the solar panel to the energy storage without the need of additional PV device. the methodology shown in Section 3. In this way, the proposed tracking method allows continuous This model will be implemented in a low-power microcontroller incorporated to the sensor node energy transfer from the solar panel to the energy storage without the need of additional PV device. where the panel is connected, and it must be capable of simulating the full neural-based model fast This model will be implemented in a low-power microcontroller incorporated to the sensor node where enough. the panel is connected, and it must be capable of simulating the full neural-based model fast enough.

Figure 10. Proposed ANN-based VOC MPPT. Figure 10. Proposed ANN-based VOC MPPT.

Following similar similar considerations considerations to to that that presented presented in in Section Section 3, 3, the the neural neural model model selected selected is is aa Following MLP. In In this this case, case, in in order order to to be be implemented in aa low-cost low-cost microcontroller, microcontroller, several several constrains constrains next next MLP. implemented in detailed have been applied to the number of input parameters and non-linear function computational detailed have been applied to the number of input parameters and non-linear function computational complexity, so so as as to complexity, to simplify simplify the the model model while while preserving preserving accuracy. accuracy. The simulator simulator in in the the Control Control block block (Figure (Figure 10) of the panel from the readings of The 10) estimates estimates the the V Vmpp mpp of the panel from the readings of the sensors of illuminance and temperature; this limits the number of inputs to two. two. Thus, Thus, the the PV PV panel panel the sensors of illuminance and temperature; this limits the number of inputs to behaviour is modelled by a 2-2-1 MLP with non-linear output function in the hidden layer and linear behaviour is modelled by a 2-2-1 MLP with non-linear output function in the hidden layer and linear operation in in the the output output layer. layer. This This neural neural network network has has been been implemented implemented in in the the same same microcontroller microcontroller operation used in the panel characterization system (Section 2), a MSP430FR5739 (Texas Instruments, Dallas,TX, TX, used in the panel characterization system (Section 2), a MSP430FR5739 (Texas Instruments, Dallas, USA), which which includes includes aa 32 32bit bithardware hardware multiplier. multiplier. In Inthis thisway, way, the the computational computational resources resources are are shared shared USA), with the solar panel characterization module. Weight values are represented in 16 bit floating point, with the solar panel characterization module. Weight values are represented in 16 bit floating point, and data are normalized in the [ − 1, +1] range, limiting the training data to the values shown in Table and data are normalized in the [−1, +1] range, limiting the training data to the values shown in Table 2.2. The training training data data set set is is generated generated from from the thecomplete completeANN ANN panel panelmodel modelpreviously previouslyshown shownin inSection Section3,3, The just sweeping the load value for each illuminance-temperature couples (in the valid ranges of the panel panel just sweeping the load value for each illuminance-temperature couples (in the valid ranges of the location). Therefore, the maximum voltage Voc is obtained for each pair of illuminance—temperature location). Therefore, the maximum voltage Voc is obtained for each pair of illuminance—temperature values and, and,hence, hence,the thecorresponding correspondingVVmpp straightforwardly derived. mpp values is is straightforwardly derived. The MLP has been trained in off chip mode, using hyperbolic tangent as as non-linear non-linear operation operation in in The MLP has been trained in off chip mode, using hyperbolic tangent the hidden hidden layer, layer, leaving leaving only only the the execution execution phase phase for for its its electronic electronic implementation. implementation. Figure Figure 11 11 shows shows the the MLP architecture with the resulting weights after training. the MLP architecture with the resulting weights after training.

Sensors 2017, 17, 1794

12 of 18

Sensors 2017, 17, 1794

12 of 18

Due to its inherent computational complexity, the most important limitation when a neural model Sensors 2017, 17, 1794 12 of 18 Due to its inherent computational complexity, the most limitation In when neural is electronically implemented is the non-linear operation in theimportant hidden processors. this acase, the model is electronically implemented is the non-linear operation in the hidden processors. In this case, hyperbolic has been replaced by acomplexity, five sectionsthe piecewise polynomial approach given Due totangent its inherent computational most important limitation when a by: neural the hyperbolic tangent has been replaced by a five sections piecewise polynomial approach given by: model is electronically  implemented is the non-linear operation in the hidden processors. In this case, − 1, x ≤ − 5.3   −1, ≤ −5.3 the hyperbolic tangent  has been replaced by a five sections piecewise polynomial approach given by:   0.0127 ∙· x3 + 0.1435 ·∙ x2 + 0.9 >>x > 5.3  0.0127 + 0.1435 +0.5234 0.5234· ∙x −−0.3840, 0.3840,−−0.9 >− −5.3 −1, ≤ −5.3 3 y ==f ( x ) == (2) −0.2358 ·∙ x + 0.9 ≤ ≤ 0.9 0.9 (2) + 0.9834 0.9834· ∙x, , − −0.9 ≤x ≤ −0.2358  0.1435 ∙ · x2++0.5234 ∙· x− 0.3840, −0.9 >x >>0.9 −5.3 ∙ · x+  3−  0.0127 0.0127 0.1435 0.5234 + 0.3840, 5.3 >  0.0127 ∙ − 0.1435 ∙ + 0.5234 ∙ + 0.3840, 5.3 > > 0.9 = = (2)  −0.2358 ∙ + 0.9834 ∙ , −0.9 ≤ ≤ 0.9 + 1, x ≥ ≥5.3 5.3 +1, 0.0127 ∙ − 0.1435 ∙ + 0.5234 ∙ + 0.3840, 5.3 > > 0.9 +1, ≥ 5.3

Multilayer perceptron perceptron weight weight values values obtained obtained to to modelling modelling the the solar solar panel behavior. Red Figure 11. Multilayer values correspond to bias weights. Figure 11. Multilayer perceptron weight values obtained to modelling the solar panel behavior. Red values correspond to bias weights.

And implemented minimizing And that that has has been been implemented minimizing the the operations, operations, as as shown shown in in Figure Figure 12, 12, to to reduce reduce the the computing time. The accuracy of this polynomial approach is shown in Figure 13. Left side shows the computing time. of thisminimizing polynomialthe approach is shown in Figure 13. 12, Lefttoside shows And that hasThe beenaccuracy implemented operations, as shown in Figure reduce the representation of a tanh(x) computed by Matlab running on a personal computer overlapped to the the representation of a tanh(x) computed by Matlab running on a personal computer overlapped to computing time. The accuracy of this polynomial approach is shown in Figure 13. Left side shows the proposed polynomial approach in Figure 12 implemented on the used microcontroller. Right side in the proposed polynomial implemented on the used microcontroller. Right representation of a tanh(x) approach computedinbyFigure Matlab12running on a personal computer overlapped to the Figure presents relative error this error approach, which is on lower than the 2% zero crossing). side in13Figure 13 the presents the relative of this approach, is(except lower in than (except proposed polynomial approach in of Figure 12 implemented thewhich used2% microcontroller. Right side in in Finally, the execution time has been experimentally estimated and compared to the time required for the zero crossing). Finally, the execution time has been experimentally estimated and compared Figure 13 presents the relative error of this approach, which is lower than 2% (except in the zero crossing). computing tanh(x), using the standard C mathematical library. have Both been to the time required for computing the tanh(x), using the standard C Both mathematical Finally, the the execution time has been experimentally estimated and compared tooperations the timelibrary. required for implemented in the microcontroller, measuring the execution time switching an output pin from logical operations implemented the microcontroller, measuring the Both execution time switching an computing have the been tanh(x), using the instandard C mathematical library. operations have been ‘1’ to logical ‘0’. Figure 14 shows the operation time using the tanh(x) (a) and polynomial approach (b) for output pin from logical ‘1’ to logical ‘0’. Figure 14 shows the operation time using the tanh(x) (a) and implemented in the microcontroller, measuring the execution time switching an output pin from logical three different input values, corresponding to input the ranges shown Figure 12.input The approach tanh(x) polynomial approach (b)shows for three values, to the rangesfunction shown ‘1’ to logical ‘0’. Figure 14 the different operation timeinput using thecorresponding tanh(x) (a)inand polynomial (b) for requires a constant time to perform each of the three operations ( = = = 2.4 ms), while the in Figure 12. The tanh(x) function requires a constant time to perform each of the three operations three different input values, corresponding to the input ranges shown in Figure 12. The tanh(x) function 0 polynomial implementation aeach variable time of operations = 740 µs (worst if(.)ms), (T =Tconstant the polynomial implementation requires variable ofinTFigure = 74012), µs requires timewhile to requires perform of the three ( =acase, =second =time 2.4 the A = TB a C = 2.4 ms), Cwhile 0 0 = 500 µs (else condition) and = 33 µs (best case, first if(.)), always improving the execution (worst case,implementation second if(.) in Figure 12),a variable T = 500time µs (else and T case, = 33second µs (best case, firsttime. if(.)), polynomial requires of condition) = 740 µs (worst if(.) in Figure 12), B

A

always the execution = 500improving µs (else condition) and time. = 33 µs (best case, first if(.)), always improving the execution time.

Figure 12. Pseudocode representing the piecewise polynomial implementation of equation. Figure 12. Pseudocode representing the piecewise polynomial implementation implementation of of equation. equation.

Sensors 2017, 17, 1794 Sensors 2017, 17, 1794 Sensors 2017, 17, 1794

13 of 18 13 of 18 13 of 18

(a) (a)

(b) (b) Figure 13. 13. (a) (a) Hyperbolic Hyperbolictangent tangentcomputation computationusing usinga aastandard standard C library library tanh(x) function (red) and Figure Hyperbolic tangent computation using standard tanh(x) function (red) CC library tanh(x) function (red) andand the the piecewise approach (blue) on a microcontroller; (b) Relative error of the polynomial fitting the piecewise approach on a microcontroller; (b) Relative error of the polynomial fitting piecewise approach (blue)(blue) on a microcontroller; (b) Relative error of the polynomial fitting compared compared to the the library function. function. to the library function. compared to library

(a) (a) Figure 14. Cont.

Sensors 2017, 17, 1794 Sensors 2017, 17, 1794

14 of 18 14 of 18

Sensors 2017, 17, 1794

14 of 18

(b) (b)different input values using (a) C library-based Figure 14. Computing time measurement for three Figure 14. Computing time measurement for three different input values using (a) C library-based tanh(x) operation and (b) time polynomial approach; Times correspond, respectively, to calculate the 14. Computing measurement for three different input values using (a) C library-based tanh(x)Figure operation and (b) polynomial approach; Times correspond, respectively, to calculate ′ tanh(x)for operation andmeet: (b) polynomial correspond, to calculate 5.3 > abs(x) > 0.9 ( respectively, , ); and abs(x) ≤ 0.9 (0 the, ′ ). operations inputs that abs(x) ≥ 5.3approach; ( , ′ );Times 0 the operations for inputs that meet: abs(x) ≥ ′ 5.3 (TA , TA ); 5.3 > abs(x) > 0.9 (T B , T ); and ); 5.3 > abs(x) > 0.9 ( , ′ ); and abs(x) ≤ 0.9 ( ,B ′ ). operations for inputs that meet: abs(x) ≥ 5.3 ( , abs(x) ≤ 0.9 (TC , TC0 ).

Figure ANN-basedVOC VOCMPPT MPPT routine routine flow Figure 15. 15. ANN-based flowdiagram. diagram.

Figure 15. ANN-based VOC MPPT routine flow diagram.

Sensors 2017, 17, 1794

15 of 18

Figure 15 shows the flow diagram of the VOC MPPT routine with ANN modelling of the PV panel. The Control block microcontroller enters every time t in the interrupt routine. If the time after Sensors 2017, 17, 1794 15 of 18 the last Vmpp estimation is higher than a set time T (>t), the system updates its value by computing the new VFigure the ANN model using of current illuminance and with temperature readings. 15 shows the flow diagram the VOC MPPT routine ANN modelling of Next, the PVif the mpp from panel. The Control microcontroller enters every time t in the interrupt routine. If the time after voltage V1 at the panelblock output is higher than the V , the control switch is turn on reducing the input mpp the last V mpp estimation is higher than a set time T (>t), the system updates its value by computing the impedance and, hence, reducing the voltage V1. Otherwise, the switch turns off. new Vmpp from the ANN model using current illuminance and temperature readings. Next, if the voltageComparison V1 at the panel output is higher than the Vmpp, the control switch is turn on reducing the input 4.3. MPPT impedance and, hence, reducing the voltage V1. Otherwise, the switch turns off.

Figure 16 shows the electronic implementation of the energy harvesting system based on the VOC 4.3. technique with ANN-based panel modelling. Temperature and illuminance are monitored by MPPT Comparison the microcontroller (located at the characterization system on top of the figure), that compares the Figure 16 shows the electronic implementation of the energy harvesting system based on the real panel output voltage (sent frompanel the DC-DC system to the microcontroller through the voltage VOC technique with ANN-based modelling. Temperature and illuminance are monitored by tracking wires), and the ANN-based maximum power point voltage estimation (from temperature the microcontroller (located at the characterization system on top of the figure), that compares the and illuminance real time values) properly switch the value the Control line of thethe boost DC-DC real panel output voltage (sent to from the DC-DC system to the of microcontroller through voltage tracking wires), and theAn ANN-based maximum power point voltage estimation (from temperature converter (centre, bottom). additional line (in parallel to the Control line in Figure 16) disables the andcircuit illuminance real time values) to properlysystem switch when the value the Control linetests of the boost DCDC-DC to operate the characterization it isofrequired. Load were performed DC converter (centre, bottom). An additional line (in parallel to the Control line in Figure 16) disables using one supercap to speed up the comparison process, though a wireless sensor node can require the DC-DC circuit to operate the characterization system when it is required. Load tests were several in parallel [20] for suitable energy storage. performed using one supercap to speed up the comparison process, though a wireless sensor node The proposed technique has been compared to a classical analog VOC MPPT to evaluate its can require several in parallel [20] for suitable energy storage. performance. In the proposed model, the MPPT management routine is executed every 10 µs (t in The proposed technique has been compared to a classical analog VOC MPPT to evaluate its Figure 15), while Vmpp estimation is upgraded every 100 msroutine (T in Figure 15). every Both techniques performance. the In the proposed model, the MPPT management is executed 10 µs (t in are executed in parallel over two identical solar panels with the same orientation, seeking for the Figure 15), while the Vmpp estimation is upgraded every 100 ms (T in Figure 15). Both techniques aresame working conditions. executed in parallel over two identical solar panels with the same orientation, seeking for the same working17conditions. Figure shows the dynamic energy load process for a 1.5 F supercap using both classical (blue) Figure shows the dynamic load process a 1.5 F supercap using both classical (blue) and ANN-based17(yellow) VOC MPPTenergy techniques at fourfordifferent environmental conditions. The load and ANN-based (yellow) VOC MPPT techniques at four different environmental conditions. Thex axis process stops when the supercap is fully loaded and its voltage reaches 5.5 volts. In Figure 17, the load process stops when the supercap is fully loaded and its voltage reaches 5.5 volts. In Figure 17, represents time (1000 seconds full screen for Figure 17a–c, and 2000 seconds for Figure 17d), while the x axis represents time (1000 seconds full screen for Figures 17a–c, and 2000 seconds for Figure y axis shows the voltage evolution with time. Both techniques have been implemented in parallel, 17d), while y axis shows the voltage evolution with time. Both techniques have been implemented in placed in identical environmental conditions and orientation. Energy load processes are triggered at parallel, placed in identical environmental conditions and orientation. Energy load processes are the same time,atand sensors and identical both systems, being the control triggered the same time, andDC-DC sensorsconverters and DC-DCare converters arefor identical for both systems, being architecture thearchitecture only difference. The proposed system presents a more efficient under those the control the only difference. The proposed system presents a morebehaviour efficient behaviour conditions, considerably reducing the load time.the load time. under those conditions, considerably reducing

Figure 16. Full solar energy harvestingsystem system based based on and energy storage system Figure 16. Full solar energy harvesting onANN ANNMPPT MPPT and energy storage system (supercapacitor): solar panel (left), control system (centre, on top, includes the characterization (supercapacitor): solar panel (left), control system (centre, on top, includes the characterization system), system), DC-DC system (centre, bottom) and supercap (right). In this case, only one of the 1.5 F DC-DC system (centre, bottom) and supercap (right). In this case, only one of the 1.5 F supercaps is supercaps is connected to the boost system for characterization purposes. connected to the boost system for characterization purposes.

Sensors 2017, 17, 1794 Sensors 2017, 17, 1794

16 of 18 16 of 18

(a)

(b)

(c)

(d)

Figure Figure 17. 17. Charge Charge curves curves for for aa 1.5 1.5 FF supercap supercap for for aa VOC VOC algorithm algorithm using using aa reference reference panel panel (blue) (blue) and and the ANN-based solar panel model (yellow). Environmental conditions (irradiance and temperature, the ANN-based solar panel model (yellow). Environmental conditions (irradiance and temperature, 2 2, 55 °C; (c) 1120 W/ m2, 43 °C; (d) 500 W/ m2, 26 °C. mm average): ◦ C;(b) 2 , 55 ◦ C; (c) 1120 W/ m2 , 43 ◦ C; (d) 500 W/ m2 , 26 ◦ C. average): (a) (a) 850 850 W/m W/m,2 ,35 35°C; (b)1300 1300W/ W/

5. Conclusions 5. Conclusions The application of a characterization system for small solar panels powering wireless sensor The application of a characterization system for small solar panels powering wireless sensor networks after deployment provides an accurate local behavioural model, which allows monitoring networks after deployment provides an accurate local behavioural model, which allows monitoring the efficiency changes in the energy generation due to both environmental and aging effects. Taking the efficiency changes in the energy generation due to both environmental and aging effects. Taking this into account, this paper has proposed an in situ PV panel modelling system, complying with the this into account, this paper has proposed an in situ PV panel modelling system, complying with the severe constraints related to wireless sensor networks in size, cost and energy consumption. Based severe constraints related to wireless sensor networks in size, cost and energy consumption. Based on on the acquired data, a behaviour-based PV model is developed based on a multilayer perceptron; the acquired data, a behaviour-based PV model is developed based on a multilayer perceptron; this this ANN solution makes easier to incorporate new characterization measures into the model, ANN solution makes easier to incorporate new characterization measures into the model, adapting it adapting it to changes in the current operating conditions. to changes in the current operating conditions. Finally, a boost DC-DC converter based on the maximum power point tracking using the VOC Finally, a boost DC-DC converter based on the maximum power point tracking using the VOC algorithm, where the pilot panel monitoring the open circuit voltage is replaced by a customized algorithm, where the pilot panel monitoring the open circuit voltage is replaced by a customized behavioural model based on the proposed multilayer perceptron approach, with reduced computing behavioural model based on the proposed multilayer perceptron approach, with reduced computing requirements to allow the software model to be implemented in a low-cost microcontroller embedded requirements to allow the software model to be implemented in a low-cost microcontroller embedded in the wireless node, that manages both the data acquisition and modelling. Compared to the classical in the wireless node, that manages both the data acquisition and modelling. Compared to the classical VOC technique, experimental results show an improved performance in loading a supercap as VOC technique, experimental results show an improved performance in loading a supercap as energy energy storage device. In addition, this technique can be applied to different MPPT algorithms where storage device. In addition, this technique can be applied to different MPPT algorithms where a solar a solar panel or its behaviour is required in its operation. panel or its behaviour is required in its operation. Acknowledgments: This work was supported in part by TEC2015-65750-R (MINECO/FEDER, UE) and JIUZ2016-TEC06. Authors would like to acknowledge the use of Servicio General de Apoyo a la Investigación-SAI, Universidad de Zaragoza.

Sensors 2017, 17, 1794

17 of 18

Acknowledgments: This work was supported in part by TEC2015-65750-R (MINECO/FEDER, UE) and JIUZ-2016-TEC06. Authors would like to acknowledge the use of Servicio General de Apoyo a la Investigación-SAI, Universidad de Zaragoza. Author Contributions: Diego Antolín designed and implemented the system, developed the panel model and experimentally validated the system. Belén Calvo and Nicolás Medrano supervised all the work and analyzed the results. Pedro A. Martínez provided theoretical support to the work. Diego Antolín, Belén Calvo, Nicolás Medrano and Pedro A. Martínez wrote the paper. Conflicts of Interest: The authors declare no conflict of interest.

References 1. 2. 3. 4. 5.

6. 7.

8. 9.

10. 11.

12.

13. 14. 15. 16.

17. 18. 19.

Han, G.; Dong, Y.; Guo, H.; Shu, L.; Wu, D. Cross-layer optimized routing in wireless sensor networks with duty cycle and energy harvesting. Wirel. Commun. Mob. Comput. 2015, 15, 1957–1981. [CrossRef] Cammarano, A.; Petrioli, C.; Spenza, D. Online Energy Harvesting Prediction in Environmentally Powered Wireless Sensor Networks. IEEE Sens. J. 2016, 16, 6793–6804. [CrossRef] Zou, T.; Lin, S.; Feng, Q.; Chen, Y. Energy-Efficient Control with Harvesting Predictions for Solar-Powered Wireless Sensor Networks. Sensors 2016, 16, 53. [CrossRef] [PubMed] Yoon, I.; Kim, H.; Noh, D.K. Adaptive Data Aggregation and Compression to Improve Energy Utilization in Solar-Powered Wireless Sensor Networks. Sensors 2017, 17, 1226. [CrossRef] [PubMed] Mandourarakis, I.; Koutroulis, E. Optimal design of a boost-type DC-DC converter for PV power-supplied wireless sensor networks. In Proceedings of the 2015 IEEE International Conference on Industrial Technology (ICIT), Seville, Spain, 17–19 March 2015; pp. 1100–1105. Penella-López, M.T.; Gasulla-Forner, M. Powering Autonomus Sensors; Springer: Dordrecht, The Netherlands, 2011. Altas, I.H.; Sharaf, A.M. A Photovoltaic Array Simulation Model for Matlab-Simulink GUI Environment. In Proceedings of the International Conference on Clean Electrical Power, ICCEP ‘07, Capri, Italy, 21–23 May 2007; pp. 341–345. Khatib, T.; Elemenreich, W. Modeling of Photovoltaic Systems Using MATLAB: Simplified Green Codes; John Wiley & Sons: Hoboken, NJ, USA, 2016. Mahmoud, S.A.; Alsari, M.M.; Reda, E.I.; Alhammadi, R.M. MATLAB modeling and simulation of photovoltaic modules. In Proceedings of the 2012 IEEE 55th International Midwest Symposium on Circuits and Systems (MWSCAS), Boise, ID, USA, 5–8 August 2012; pp. 786–789. Ropp, M.E.; Gonzalez, S. Development of a MATLAB/Simulink Model of a Single-Phase Grid-Connected Photovoltaic System. IEEE Trans. Energy Convers. 2009, 24, 195–202. [CrossRef] Can, H.; Ickilli, D.; Parlak, K.S. A New Numerical Solution Approach for the Real-Time Modeling of Photovoltaic Panels. In Proceedings of the Power and Energy Engineering Conference (APPEEC), Shanghai, China, 27–29 March 2012; pp. 1–4. Ulapane, N.N.; Dhanapala, C.H.; Wickramasinghe, S.M.; Abeyratne, S.G.; Rathnayake, N.; Binduhewa, P.J. Extraction of parameters for simulating photovoltaic panels. In Proceedings of the 2011 6th International Conference on Industrial and Information Systems (ICIIS), Kandy, Sri Lanka, 16–19 August 2011; pp. 539–544. Castañer, L.; Silvestre, S. Modelling Photovoltaic Systems Using PSpice; John Wiley & Sons: Chichester, UK, 2002. Bukshtab, M. Applied Photometry, Radiometry, and Measurements of Optical Losses; Springer: New York, NY, USA, 2012. Saberian, A.; Hizam, H.; Radzi, M.A.; Ab Kadir, M.Z.; Mirzaei, M. Modelling and Prediction of Photovoltaic Power Output Using Artificial Neural Networks. Int. J. Photoenergy 2014, 1–10, 469701. [CrossRef] Sánchez-García, J.L.; Espinosa-Juárez, E.; Tapia-Juárez, R. Photovoltaic panel characterization by using artificial neural networks and comparison with classical models. In Proceedings of the 2015 IEEE International Autumn Meeting on Power, Electronics and Computing (ROPEC), Ixtapa, Mexico, 4–6 November 2015; pp. 1–6. Karapete, E.; Boztepe, M.; Colak, M. Neural network based solar cell model. Energy Conver. Manag. 2006, 47, 1159–1178. Haykin, S. Neural Networks and Learning Machines; Pearson Education: New Jersey, NJ, USA, 2009. Principe, J; Euliano, N.; Lefebvre, W. Neural and Adaptive Systems, Fundamentals through Simulations; John Wiley & Sons: New York, NY, USA, 2000.

Sensors 2017, 17, 1794

20.

21. 22. 23. 24. 25. 26.

27.

18 of 18

Antolín, D.; Medrano, N.; Calvo, B. Reliable Lifespan Evaluation of a Remote Environment Monitoring System Based on Wireless Sensor Networks and Global System for Mobile Communications. J. Sens. 2016, 2016, 1–12. [CrossRef] Mohan, N.; Underland, T.M.; Robbins, W.P. Power Electronics: Converters, Applications, and Design; John Wiley & Sons: New York, NY, USA, 2003. Hart, D.W. Power Electronics; McGraw-Hill: New York, NY, USA, 2011. Texas Instruments. Application Note-1994: Modeling and Design of Current Mode Control Boost Converters. Texas Instruments Application Report SNVA408B; Texas Instruments: Dallas, TX, USA, 2013. Balato, M.; Costanzo, L.; Vitelli, M. Maximum Power Point Tracking Techniques; Wiley Online Encyclopedia of Electrical and Electronics Engineering; John Wiley & Sons: New Jersey, NJ, USA, 2016; pp. 1–26. Esram, T.; Chapman, P.L. Comparison of photovoltaic array maximum power point tracking techniques. IEEE Trans. Energy Conver. 2007, 22, 439–449. [CrossRef] Gupta, A.K.; Saxena, R. Review on widely-used MPPT techniques for PV applications. In Proceedings of the 2016 International Conference on Innovation and Challenges in Cyber Security (ICICCS-INBUSH), Noida, India, 3–5 February 2016; pp. 270–273. Pressman, A.; Billings, K.; Morey, T. Switching Power Supply Design; McGraw-Hill: New York, NY, USA, 2009. © 2017 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).