Investigation of a Real-Time Dynamic Model for a PV Cooling System

Abstract

The cooling of PV models is an important process that enhances the generated electricity from these models, especially in hot areas. In this work, a new, active cooling algorithm is proposed based on active fan cooling and an artificial neural network, which is named the artificial dynamic neural network Fan cooling algorithm (DNNFC). The proposed system attaches five fans to the back of a PV model. Subsequently, only two fans work at any given time to circulate the air under the PV model in order to cool it down. Five different patterns of working fans have been experimented with in this work. To select the optimal pattern for any given time, a back propagation neural network model was trained. The algorithm is a dynamic algorithm since it re-trains the model with new recorded surface temperatures over time. In this way, the model automatically adapts to any weather and environmental conditions. The model was trained with an indoor dataset and tested with an outdoor dataset. An accuracy of more than 97% has been recorded, with a mean square error of approximately 0.02.

1. Introduction

The idea of developing a world that does not rely on fossil fuel sources has been a common objective of most industry-leading nations for years now [1]. A lot of efforts have been made to find and improve alternative energy sources that are sustainable and can meet the ever-increasing demand for energy [2]. Solar energy is one of these sources that can be harvested using many technologies, the most common of which is the photovoltaic (PV) systems. PV systems convert sunlight into electrical energy by freeing electrons from semiconductor materials, thus creating a direct current [3]. The PV industry has been rapidly growing, pushing power conversion efficiencies to 23–29% [4,5]. Furthermore, as of 2019, the accumulative capacity of installed PV systems in the world has surpassed 600 GW and is expected to reach around 1 TW by 2022 [6,7].

All of these technological advances in PVs, supported by the continuous decrease in prices of PV systems [8], have made them one of the most promising alternative energy sources out there. However, monitoring the electrical efficiency of PV systems has become increasingly more important in the past few years. PV systems are prone to a lot of factors such as solar radiation, ambient temperatures, dust, shading, and many other weather conditions [9,10]. These factors may very well affect the efficiency and the profitability of PVs. For instance, several studies have discussed the negative effects of elevated temperatures on PV modules and suggested different methods of improving their efficiencies [11,12,13]. Other studies have investigated the effect of dust on PV modules [14,15]. Nevertheless, there is no doubt that artificial intelligence models can improve the power, efficiency, and stability of such systems in potential smart grid applications [16].

The results provided by prediction models can sufficiently forecast, with certain accuracies, the power outcomes of PVs under different working conditions [17]. Therefore, researchers have been investigating ways to estimate and predict the behavior of PV modules with the use of different numerical methods. Artificial neural networks (ANN) are considered as one of the most commonly used methods for estimating the electrical efficiencies and predicting the power generation of PV modules [18]. The majority of these studies, however, merely investigated PV power outputs [19], optimum tracking behaviors [20], or fault detection [7,21].

In this work, a new and dynamic ANN fan cooling (DNNFC) system is proposed. The system consists of mini-current DC fans that are deployed on the back of a solar PV model at five different positions. At any given moment, only two fans work as the air inlet and outlet to the system. Five different patterns were created with these fans. To select the optimal pattern at any given time, a trained ANN model was deployed to select the optimal model according to the recorded surface temperature. The proposed system is a novelty since the algorithm retrains the ANN model over time with new recorded temperature data. The ANN model was trained with data from the indoor environment, and then tested using data from the outdoor environment. Based on the training results, the developed model dynamically chooses the activation of two of the five fans to cool the PV module according to the optimal air pattern. This way, the model gives the systems the ability to adapt to weather conditions and different environmental scenarios such as, dust, clouds, and rain.

The rest of this paper is organized as follows: Section 2 summarizes the related works that have been conducted in the area of enhancing the working environment of PV modules. Section 3 introduces the proposed algorithm and its working steps. Section 4 presents an overview of the experiment and the obtained results. We conclude this paper in Section 5.

2. Background

Artificial Intelligence (AI) has become a very important forecasting tool in many applications, including solar PV generation. Several studies have explored the capabilities of AI in photovoltaic systems and have reported results in power generation, efficiency improvements, and even the stability of such systems. For instance, one study used multiple regression and ANN models to successfully predict photovoltaic power generation under different environmental parameters such as ambient temperature, irradiance, and PV surface temperature [22]. In another study, a recurrent neural network (RNN) was utilized to predict PV power outputs based on hourly patterns throughout different seasons [23]. The results showed that the RNN model outperformed other conventional statistical models when comparing mean absolute errors. There have also been studies that investigated PV power outputs based on uncertainty and prediction intervals [24,25]. However, there have been many studies that addressed the improvement of PV efficiencies. ANN models were developed to maximize the power output of solar photovoltaic systems, hence increasing their efficiencies. One study in particular showed that the improved design of an ANN model significantly enhanced the performance of a maximum power point tracking algorithm (MPPT), which increased the efficiency of the PV system [26]. Another study investigated the effect of dust particle size on the efficiency of solar panels [27]. The results showed that the developed Adaptive Neuro-Fuzzy Inference System (ANFIS) was able to successfully predict the electrical efficiency of PV modules for any given dust particle size. Incidentally, none of the previously cited studies incorporated active or passive cooling systems in their developed models. Any cooling system for photovoltaics can play an essential part in controlling the temperature of the operating PVs, hence affecting their efficiencies and power outputs. For example, cooling techniques that depend on passive elements such as heat sinks [28,29] can be deployed on the back of PV modules, as shown in [30,31]. The cost of deploying heat sinks under all PV modules, however, makes this method somewhat unviable. On the other hand, active cooling methods that depend on fluids have two main issues. First, the type of coolant impacts the cost and the viability of the system. Utilizing water to cool a vast number of PV modules in a farm is an issue. Second, active cooling requires energy consumption, which is also an issue. To deploy such methods, the electrical consumption of the tools has to be optimized to improve the efficacy of the system. To optimize active cooling systems, machine learning and AI have to be leveraged.

However, there are few studies that have developed prediction models for PV systems that are equipped with cooling techniques. For example, a previous research work investigated several nanofluid cooling technologies for a photovoltaic/thermal system [32]. An ANN model was later proposed and particularly showed the nano-phase change material (nano-PCM) cooling strategy, which accurately predicted the enhancement of the electrical efficiency of the system [33]. Similar results were also presented in [34]. Furthermore, another dynamic neural network model was developed for a PV configuration with porous fins applied on its back [35]. The results showed that the ANN model accurately predicted the performance enhancement of the module. Most of these studies, however, only focused on the efficiency and the power outcome of PV modules with cooling systems.

Very few studies have focused on developing AI models that are able forecast the surface temperature of the panel ahead of time. This type of algorithm may provide an insight into optimizing PV cooling technologies to be able to work more robustly and effectively. A previous study developed an ANFIS technique to estimate the operating temperature of a photovoltaic array [36]. The presented model was validated with experimental data from the field. The results showed that the model provided a reliable tool for predicting the temperature of the panels based on environmental variables. In another study, four AI models were trained to predict the temperature of PV modules [37]. It was concluded that all four techniques can present predictions with 93% accuracy. A more recent study investigated the best function form for predicting the temperature of the panel for several different PV capacities [38]. The investigation tested the best function form by measuring the PV temperature using the nominal operating cell temperature and the nominal operating module temperature. The results showed that the accuracy of the models improved as the months became hotter, or as the panel capacity/area became larger.

All of these studies prove that developed algorithms may be implemented in PV cooling systems as tools to improve the efficiency and the power output of the modules or to optimize the efficacy of the cooling techniques. None of these studies, however, actually applied cooling techniques onto the PV modules. This shows that there is still some ambiguity about how well an AI model can predict the temperature of a PV module with an actual cooling system applied on it. In this study, two AI models are presented for a PV panel that is cooled using forced air convection.

3. Artificial Dynamic NN Fan Cooling Algorithm (DNNFC)

The Artificial Dynamic Neural Network Fan Cooling Algorithm (DNNFC) is an active air-cooling method that depends on circulating the air under the PV model. The algorithm is a dynamic one since it has the ability to train its model dynamically over time with new recorded data. Coolant air enters from underneath the module and leaves through five different ports. Each port is equipped with a mini, low-current fan to force air in and out of the system. These fans work in both directions to push the air into the system under the PVs or to pull it out. To reduce the power consumption of the system, two fans work to circulate the air under the panels at any given moment. One fan pushes the air, and the other fan pulls it out. The fans which were not used in each pattern were sealed shut using a plastic cover specially fabricated to force the airflow movement in one direction and prevent any leaks from other ports.

These five fans have five different configuration patterns in the system, as shown in Figure 1. These patterns work as follows:

1-

Pattern 1: The first configuration pattern lets the cooling air enter from the upper right corner and leave from the upper left corner.

2-

Pattern 2: This pattern is configured to allow air to go inside the system from the upper left corner and leave from the lower left corner.

3-

Pattern 3: The third configuration pattern lets the cooling air go inside from the lower right corner and leave from the central port.

4-

Pattern 4: Configured to allow air to enter from the center and leave from the upper right corner.

5-

Pattern 5: This pattern lets air enter the system from the lower left corner and leave from the upper right corner.

Twenty different cooling path combinations could be created in this setup using one inlet and one outlet. However, these patterns have been selected for two main reasons. First, these patterns cover all the configurations that can be created horizontally, vertically, or diagonally. The cooling patterns were chosen based on the lines of symmetry of the module, while utilizing all the inlet/outlet ports of the DC fans. Therefore, each pattern is unique as compared to the others. The difference between the selected patterns and other patterns is that the left/right or up/down directions only “mirrored cooling paths”. Second, these patterns cover all the configuration distances between the holes. The first pattern covers the width distance. The second pattern covers the length distance, and the final pattern covers the diagonal distance. Finally, the third and the fourth patterns cover the special configured distances in the system. Each one of these configurations circulates the air under the PV model to cool it down. However, each one of these five configurations has a different impact on the PV temperature according to the temperature distribution on the module and the value of the measured temperature.

To select between these five configurations, a real-time “online training” artificial neural network model was constructed and trained. This model is first fed with the working configuration and the average temperature of all temperature sensors attached to the PV surface. Subsequently, the model selects a new pattern, depending on the last pattern and the average temperature measured. Figure 2 shows how the NNFC algorithm works. It can be observed from the figure that the algorithm consists of three parts: the temperature averaging process, the trained neural model, and the controller that controls the fans according to the output of the neural model. The following subsection gives an overview of the ANN model that was trained.

3.1. Electrical and Thermal Performance

The electrical efficiency of the PV module depends on the maximum power output, the area of the module, and the solar irradiance. Under standard test conditions, the solar irradiance is set to 1000 W/m². The following equation governs the electrical efficiency of the panel:

$${\eta }_{el}=\frac{P_{mp}}{A_{PV}\times G}\times 100\%$$

(1)

The maximum power output can be determined from the readings of the maximum current, I_mp, and the maximum voltage, V_mp. These values were later used to present the data in Figures 5 and 8.

The thermal efficiency of the system is determined with Equation (2). The thermal performance depends on the mass flow rate and specific heat of the coolant, the difference in the coolant’s temperature, and the solar irradiation as well.

$${\eta }_{th}=\frac{\dot{\mathrm{m}}{C}_p\left(T_{out}-T_{in}\right)}{A_{PV}\times G}\times 100\%$$

(2)

where the volumetric flow rate of the low-current DC fan is around 1.25 m³/min at rated speed.

3.2. Real-Time Dynamic Neural Network Model

A feed-forward, back-propagation neural network model was selected for the NNFC system. The ANN had one hidden layer, whereas the hidden layer consisted of 12 neurons. This number was selected because the neurons had to be twice as many as the input features. The input layer consisted of six features and a bias term. These features were the average surface temperature and five binary features that showed the state of each one of the cooling fans (i.e., 1 for “on” and 0 for “off”). The output layer of the model contained one output, which was the predicted surface temperature under the entered configuration. Each neuron in the hidden layer utilized the Sigmoid function, as shown in Equation (3). Finally, the output of the system is the Sigmoid function of the hidden neurons’ output, calculated in Equation (3), as shown in Equation (4).

Table 1. Variable Definitions.

Variable	Meaning
$F$	Number of Features of the model. This number is equal to 6 in this model
$w_i$	The variable weight that needs to be optimized
$b$	The bias term, also needs to be optimized in the training step of the model
$y_i$	The output of neuron “i” in the hidden layer
$N$	The number of neurons in the hidden layer

shows the meaning of all the variables utilized in these equations.

$$y_i=\frac{1}{1+e^{-\left(b+{\sum }_{j=0}^F{w}_j{f}_j\right)}}$$

(3)

$$Output=\frac{1}{1+e^{-\left(b+{\sum }_{i=0}^N{w}_i{y}_i\right)}}$$

(4)

After constructing the model, it had to be trained to optimize the weights and bias terms for each node in the hidden and the output layers. A dataset was constructed with the five configuration pattern values and the average surface temperature of the PV module before running the configuration and after running it for a threshold time value, “t”. A time interval of one minute was selected as the threshold timing. To obtain the training dataset, the PV module was placed in a photovoltaic performance simulator device (PVPS) inside a laboratory. The experiments from which the training data were collected were conducted indoors to guarantee ideal testing conditions and equal temperature distribution across the PV module. The PVPS device applied a radiative flux onto the module at 1000 W/m², which caused the module to heat up. When the surface temperature of the module reached a maximum steady value, one of the fans’ configurations (shown in Figure 1) was activated. The surface temperature of the module was recorded every one minute, until the temperature reached the approximate room temperature. A more detailed explanation of the procedure is discussed in Section 4.1.2.

A total of 305 data tuples were recorded (i.e., one hour for each fan configuration). Each tuple consisted of the configuration value, the surface temperature, the temperature of the inlet air, and the temperature of the outlet air. Thus, the training data that were used consisted of a 305 × 7 matrix. The first six columns were used as input data for the model, and the last column was used as the target value. It is worth mentioning here that the hot outlet air can be utilized for heating purposes, which can be the subject of a future study.

To utilize this model in an efficient way, the calculated surface temperature was registered in the model at five different times, and each time with a different cooling configuration value. Subsequently, the configuration with the lowest predicted surface temperature was sent to the fan controller. Hence, there was no need to modify the model to apply the predicted configuration. Algorithm 1 shows the pseudo code of NNFC system.

Algorithm 1. Pseudo code of the NNFC.

Temp (i): The temperature recorded from sensor i

Pattern: The configuration patterns

Pattern (t − 1): The running pattern

NNFC: The neural network model that was trained with the experimental dataset

T: The threshold time used to select a new pattern to run

Sensor: Array of all values read by the temperature sensors

While T

For i in sensors

temp = temp + i

End for

temp = temp/len (Sensor)

min = temp

For j in pattern

temp2 = NNFC (j)

if (temp2 < min) then

min = temp2

P = j

End if

End for

Run (P)

dataset.add (pattern (t − 1), temp)

NNFC = Re-train (NNFC, dataset)

End while

The algorithm is a dynamic one that re-trains its model with new data. Every time the threshold value “t” passed, the datalogger recorded new temperature data from the sensors. The collected data were averaged; hence, the running pattern with the average temperature data was saved in the dataset. Finally, the algorithm was re-trained again with the set of the old data combined with the newly added data. In this way, the algorithm adapts to different environment or weather situations since it can tune its variables with new data on the fly.

4. Experimental Setup

In this study, two sets of experiments were conducted: an indoor experiment and an outdoor experiment. One 100 W, 36-cell monocrystalline photovoltaic panel was used in both sets.

Table 2. General characteristics of the PV panel.

Technical Data of the PV Panel
Model	CNSDPV100(36)M5
Maximum power output	100 W
Open circuit voltage (Voc)	22.7 V
Short circuit current (Isc)	5.96 A
Cell efficiency under STD (%)	17.94%
Temperature coefficient of Vm	−0.340%/K
Temperature coefficient of Im	0.049%/K
Temperature coefficient of Pmax	−0.430%/K
Dimensions	1070 mm× 670 mm× 35 mm

presents the general characteristics of the PV module. The PV panel was mounted on a steel frame, which was tilted at a 25-degree angle from the horizon. A forced convection cooling system was applied on the back of the PV module to lower the operating temperature of the module and hence improve its efficiency. The cooling system consisted of a stainless-steel chassis with five low-energy reversible DC fans attached to it. Those fans were distributed in a way that forced air to move in specific cooling paths on the backside of the PV panel.

To monitor the temperatures in this study, eight K-type thermocouples were used with a data harvesting device (Hioki LR8431). Three of them were attached to the back of the PV module at different locations to measure its average surface temperature. Those sensors were insulated using a special insulating adhesive tape so that they would only measure the surface temperature on the back of the module. The other five sensors were installed at the fan ports to measure the inlet/outlet air temperatures. All the fans were connected to a microcontroller that controlled the working state and direction of the fans using PWM. In this section, the cooling paths, the procedure of the experiments, and the uncertainty analysis of the tests are described in detail.

4.1. Procedure

Two sets of experiments were conducted in this investigation. The first set of experiments was administered outdoors, while the second set was indoors. However, similar procedures were carried out in both tests.

4.1.1. Outdoor Experiments

During the outdoor experiments, the PV module was placed outside, facing south, from morning onwards, every day for five consecutive days. At the beginning of each day of the experiment, the module was set with all cooling fans turned off. The average surface temperature of the module was recorded as its surface temperature increased. When the average surface temperature reached a steady state at around solar noon, one of the cooling patterns, shown in Figure 1, was randomly activated. The cooling process lasted for 3 h, during which the temperatures of the inlets, outlets, and the back surface of the panel were recorded. Figure 3 shows the experimental setup outdoors.

All experiments were conducted under similar weather conditions. The ambient temperature ranged from 30 °C to 32 °C, and the global irradiance was recorded at around 850 W/m² from a nearby weather station during the cooling process on all five days. It is worth mentioning here that the type of low-current fans used in this experiment may not be durable enough for harsh weather conditions, and it is advised to use a more durable type. These experiments were conducted on Al Zaytoonah University’s campus in August 2019. The ambient temperatures and solar irradiance readings agreed with the data presented in [39] for this region in that time of the year.

4.1.2. Indoor Experiments

The PV module with the fan-cooling system was tested indoors using the Photovoltaic Performance Simulator (PVPS) apparatus [3,10]. The PVPS worked by simulating sunlight using a light source and controlling the amount of radiant flux directed at the photovoltaic module. A procedure similar to the outdoor experiments was followed in the indoor tests. At the beginning of each experiment, the PV module was installed horizontally inside the PVPS, about 40 cm away from the light source, as presented in Figure 4. The cooling fans were shut off while the PVPS was applying 1000 W/m² radiant flux perpendicular to the surface of the panel inside a temperature-controlled room at 22 °C.

When the average surface temperature of the module reached a maximum steady value, one cooling pattern was randomly activated. The cooling time lasted for 3 h, during which the average surface temperature, inlet, and outlet temperatures were recorded. The tests were conducted at the Alternative Energy Department (AED) on the same campus.

4.2. Uncertainty Analysis

In every experimental work, inaccuracies in the measurement process produce systematic errors, which create differences between the measured values and the true values in the tests. A precise way of quantifying these differences is through the performance of an uncertainty analysis, as presented by Kline and McClintock [40]. An uncertainty value can be determined for each measuring device using the datasheets of the probes or by repeated trials. However, to determine the uncertainty of all measurement properties in the entire experimental process, the following approach is used [41,42]:

$$R=R\left(x_1+x_2+x_3+\dots +x_n\right)$$

(5)

where x is the independent variable of the property, R.

Consequently, the function, ω_R, is the uncertainty result of the measurement, where the uncertainties of the independent variables are presented in ω_1, ω₂, …, ω_n [40]:

$${\omega }_R={[{(\frac{\partial R}{\partial x_1} {\omega }_{x1})}^2+{(\frac{\partial R}{\partial x_2} {\omega }_{x2})}^2+{(\frac{\partial R}{\partial x_3} {\omega }_{x3})}^2+\dots +{(\frac{\partial R}{\partial x_n} {\omega }_{xn})}^2]}^{1/2}$$

(6)

In the outdoor experiments, several probes were used: K-type thermocouples (${\omega }_t$ = ±0.4%), a multi-meter to measure the electrical power (${\omega }_p$ = ±0.2%), and the pyranometer in the weather station (${\omega }_{py}$ = ±3.5% global reading). The same equipment was used in the indoor experiments, except for the outdoor pyranometer, which was replaced by the radiant flux sensor (${\omega }_{fl}$ = ±5%) of the PVPS. Furthermore, a data harvesting device was used in both experimental sets; however, for technical reasons, its uncertainty value was discarded. All of the uncertainty values of the other probes were taken from their datasheets. Accordingly, the collective uncertainty for the outdoor and indoor experiments are calculated as follows:

$${\omega }_R{}_o=\sqrt{{\omega }_t^2+{\omega }_p^2+{\omega }_{py}^2}=3.52\%$$

(7)

$${\omega }_R{}_i=\sqrt{{\omega }_t^2+{\omega }_p^2+{\omega }_{fl}^2}=5.02\%$$

(8)

4.3. DNNFC Training Process

MATLAB was utilized to train the model with the recorded data. Six inputs and one output were used. The surface temperature value was used as the output or target data and as one of the features (t − 1).

Table 3. ANN model variables.

Parameter	Value
Number of hidden layers	1
Number of neurons in the hidden layer	12
Number of features	6
Number of output values	1
Cost Function	Mean square error Function
Number of Epochs	10,000

shows a summary of the model configuration parameters. The model was trained with different time threshold values “after one, two, five, and ten minutes”. In each process, the model was recreated, the mean square error (MSE) was calculated, and its accuracy was computed. The biggest threshold value used was “10” since the data that were harvested only had a one-hour time period for each pattern.

5. Results

Section 5 is divided into two subsections: The first comprises the experimental results, which represent the electrical/thermal efficiencies, surface temperatures, and output power results from the indoor and outdoor experimental sets. The second subsection represents the machine learning result, which illustrates the algorithms of the implemented models and their accuracies.

5.1. Experimental Results

The indoor and the outdoor experiments showed similar outcomes regarding the best and worst forced convective cooling patterns, in terms of electrical efficiency, thermal efficiency, surface temperatures, and power outputs. For instance, the highest electrical efficiency obtained from the module was through the implementation of cooling pattern 5 in both experimental sets, whereas the lowest efficiency was obtained from pattern 1. Figure 5 illustrates the electrical efficiency of the module for all five patterns during the indoor and outdoor experiments. Similarly, pattern 5 yielded the highest thermal efficiency in both experiments, and pattern 1 produced the lowest, as shown in Figure 6.

Accordingly, cooling pattern 5 dropped the average surface temperature of the module from the initial temperatures by 27.5% and 32.8% during the outdoor and indoor experiments, respectively. However, cooling pattern 1, for example, decreased the average surface temperature in the outdoor experiments by 12.5%. Figure 7 shows the average surface temperature of the module throughout all five cooling patterns.

Figure 8 shows the power performance of the PV module in both experiments. As illustrated, pattern 5 produced the best maximum power output in the outdoor and indoor experiments at 90 W and 80 W, respectively. However, cooling patterns 1 and 2 produced the lowest power output in the experiments.

In the experiments, cooling pattern 5 showed relatively better electrical and thermal efficiency than other cooling patterns. It also registered the lowest average surface temperature from the back of the PV panel, as illustrated in Figure 7. It is noticeable that pattern 5 covered the longest distance underneath the panel, which conceivably created the largest cooling area and therefore the largest amount of heat. Changing the length and depth of a cooling path can affect the behavior of the PV [43,44,45,46]. On the other hand, cooling pattern 1 produced the lowest efficiencies and the highest surface temperatures. This cooling pattern covered a relatively short cooling distance (i.e., relatively small convective area) and was located on the edge of the panel, which may explain the relatively poor results.

5.2. DNNFC Results

After training the ANN model on the indoor dataset, the model was simulated with the data collected from the outdoor experiment. As mentioned, five different t threshold values were leveraged. Figure 9 shows the accuracy of the model under different threshold values. It can be observed that with small t values, the accuracy increases; however, the process and computation of the controller becomes heavy. To reduce computation time, a threshold value of 5 min was utilized.

Figure 10 shows the accuracy of the training process of the model with t = 5. We can observe that the overall accuracy exceeded 98%. Moreover, an MSE value of 0.0259 was recorded for this model.

6. Conclusions

The technology of photovoltaics for electricity generation has proliferated in the past few years. One of the main issues that impacts the generated power is the PV surface temperature. This has motivated researchers to develop and produce different methods and techniques that would cool PV models. In this work, an active cooling method was proposed. The proposed method consists of different of fans attached to the back of the PV model. Different patterns for fan operations were proposed in this work. A neural network model was trained to select one of the patterns that had to be used at any selected time, based on the surface temperature and in order to optimize the cooling process. The proposed model is a dynamic training model that attempts to retrain itself with new harvested data while running. In this process, the model has the ability to adapt to different environmental and weather conditions. The model was trained with indoor data and conducted a simulation with outdoor data. An accuracy higher than 97% (with an MSE of 0.0259) was achieved from the model. In a future work, we will attempt to use the proposed algorithm on different types of fan-cooled PV modules.