Figure 1. Diagram of operation of the intelligent photovoltaic system For a better understanding of the operation of the proposal, a scheme was made that represents how the prototype works. Figure 1 shows how artificial intelligence was integrated with the electronic system. This was done through an Arduino board and in real time the optimal angle of the sun was obtained and with a series of instructions the solar panel will be moved to better capture solar energy. The entire procedure for the implementation of the intelligent photovoltaic system for the optimization of electrical energy is described in the subsections. Figure 2. Solar tracker circuit design A neural network algorithm in the Python language was applied to the intelligent photovoltaic system This system predicts the optimal direction or angle of sunlig ht and sends an analog signal through the seria port to the microcontroller. The microcontroller processes this signal and sends a signal to the electric moto: to control its speed and direction, allowing movement of the panel to obtain maximum energy capture The microcontroller contains a control algorithm to optimize t he tracking of sunlight and minimize the error ir the position of the panel as shown in Figure 2. This algorithm takes into account factors such as the latitud and longitude of the installation site, the time of day and season of the year to determine the direction of the panel, later in the conceptual design components were included, such as a power source to feed the system a protection system against overloads and short circuits. As shown in Figure 3. Figure 4. Intelligent photovoltaic system prototype We proceeded to load the necessary program on the Arduino board so that it can receive the angles from the neural network and control the DC motors for solar tracking. The DC motors were assembled on the frame and fixed in place by connecting the motor levers to the solar panel frame and making sure they were secure. This construction process was carried out with the objective of creating a solar tracking system based yn artificial intelligence, which can optimize the generation of electrical energy. Importantly, the use of urtificial intelligence techniques in this system allows for more precise tracking of the sun’s position, which in ‘urn can significantly increase the efficiency of the solar panel. This type of solar tracking system can be an -ffective solution to improve the efficiency and reduce the cost of solar power generation in small and medium scale applications, see Figure 4. For the application of the neural network model, the database had to be created for which it was defined based on indicators of the photovoltaic module, meteorological data, and time series data. In this way, the data collection of different indicators was carried out in order to obtain the optimal angle of the position of the sun. These indicators, detailed in Table 2, include temperature, humidity, wind speed, direct and diffuse radiation radius, direction speed, and precipitation. Data collection was done using a variety of sources, including weather stations and specialized measuring devices. To ensure data quality, specific conditions were applied, such as the address of the panel and the latitude, longitude, and altitude of where the study would take place. It should be noted that the closest meteorological station to the study point corresponds to the Ceres station, located in the department of Lima, province of Lima and district of Ate. Its metadata indicates a latitude of 12°1'43.2"S, longitude of 76°55'37.6"W and an altitude of 339 masl. Once the data was obtained, it was cleaned and pre-processed, which implied the elimination of empty columns and rows, the conversion of dates and times into a single datetime type column, the elimination of duplicates, the extraction of attributes and studies of correlation as observed in Figure 5, and obtaining attributes of the time series. Subsequently, the NeuralProphet model was used to create a neural network that could be fitted to the training data. The model included a series of layers of neurons, each of which was connected to the next layer, and appropriate activation functions were used for each layer. Despite the fact that very low correlations were obtained between the variables and the optimal angle, the time series was used as input for the neural network. Figure 6 shows the correlogram of our datasets to determine the importance of each of the network attributes contained in the dataset. Now the correlation must be calculated for all the variables of the datasets, let’s see what the correlation matrix is like. This matrix shows the correlation between two or more variables in the data set through tabulations or statistics. Figure 8. Correlation heat map between variables and the optimal angle Figure 7. DataSet of meteorological indicators Figure 6. Correlation of data variables As shown in Figure 8, after the treatment of the data, the heat map of correlation between all the ariables and the optimal angle was obtained. The result was very low correlations between variables, so the ise of all attributes is ruled out. On the other hand, it is observed in Table 3 how the datasets will be treated as time series, as input to the neural network. was studied. In addition, time series attributes were obtained, and the data set was divided into training and test sets. The model used was NeuralProphet to create a neural network model that would fit the training data. The model included several layers of neurons, each of which was connected to the next layer, and appropriate activation functions were used for each layer. When performing the correlation between the variables and the optimal angle, it was observed that the relationship level was low. Therefore, it was decided to discard the use of all the attributes, and it was decided to treat the data set as a time series. Figure 7 shows how the time series will be treated as input to the neural network. Figure 9. Training the neural network Then, the model was fitted to the training data using the “fit” function, which took as input the data in its original format, “df raw_daily”, and specified the frequency of the data, which in this case was “H” for indicate that the data were in hourly intervals. During training, the model adjusted the weights of the neural connections to minimize the specified loss function (in this case, the Huber function), and the process was repeated for 40 epochs. The training process also automatically imputed missing values and normalized the data as specified in the model parameters. At the end of the training, several model performance metrics were needed, such as the mean absolute error and the loss of the model in the training set. These metrics were used to assess the quality of model performance and to compare different model configurations in case parameter changes have been made. For this, in (1) was used; The main function of the system was to predict the optimal angle to send to the microcontroller using the trained model and send this information through the serial port. The prediction of the optimal angle was made using the “pred_angulo_ptimo” function, which uses the model to make a prediction for the future and return the value of the optimal angle. The “‘send_angle_optimum” function was in charge of sending the optimal angle through the serial port. As shown in Figure 11. In Figure 10 the implementation of the intelligent model was carried out in an automated system to send an optimal angle to a microcontroller. This was done using the Python programming language and the NeuralProphet library for modeling the time series. A time series dataset was used to train the model and the serial port was configured for communication with the microcontroller. Table 4. Results of the prediction of the neural network model Figure 11. Performance curve of the algorithm Figure 12. Comparison chart of real data and predictive data BIOGRAPHIES OF AUTHORS
