| Work Detail |
Scientists in Spain have used genetic algorithms to optimize a feedforward artificial neural network for the prediction of energy generation of PV systems. Genetic algorithms use “parents” and “offspring” solutions to achieve better results in subsequent generations.” Researchers from Spain’s Valencia Polytechnic University have developed a novel method for forecasting the power generation of PV systems. Its novelty lies in developing a hyperparameter optimization model for feedforward artificial neural networks (FF-ANNs) using genetic algorithms (GAs) techniques. Hyperparameters are external configurations set before the learning process begins that govern the learning process. “While integrating machine learning and intelligent algorithms offers significant potential, there are still challenges in optimizing modeling techniques for complex nonlinear systems,” the group said. “The study, therefore, focuses on addressing the challenges in accurately predicting solar PV power generation, a complicated task due to the inherent variability of renewable energy sources and the complexity of nonlinear power systems.” FF-ANNs are artificial neural networks (ANNs) that process information in one direction. They do so from the input layer, through the hidden layer, to the output layer, while conventional ANNs allow information to flow in different ways, using feedback loops and other memory mechanisms. FF-ANN includes several hyperparameters, namely the number of neurons, transfer functions, input weight, layer weight, and biases. GAs, on the other hand, are used before the FF-ANN runs to optimize its configurations. They are inspired by the natural selection processes, starting with hyperparameters that are defined as “parents.” These then create offspring, which is a solution to the optimization problem, and continue to create subsequent offspring until an optimal solution is reached. “The population of the GA is the different parameterizations that the ANN may assume,” the group explained. “In each generation, the quality of each individual in the population is evaluated using a fitness function, which, in the proposed case, is the root mean square error (RMSE). This is a widely accepted and used practice in the literature; this choice is based on its ability to provide a clear and objective measure of model performance, enabling comparison and evaluation in different scenarios.” The proposed approach was used on a database from a real rooftop installation in Valencia. The installation consisted of 12 monocrystalline 350 W panels accompanied by weather information from a nearby station. Recording occurred from May 1st, 2021, to April 30th, 2022. Seventy percent of the data points were used for training, 15 percent for testing, and 15 percent for validation. “This analysis has included the comparison of the results obtained by training the network using aggregated data sets at the annual, seasonal, and monthly levels,” the academics added. “The annual training may allow for capturing long-term trends and patterns, providing an overview of system behavior over time. Seasonal training may allow for analyzing seasonal variations, considering changes in weather and environmental conditions. Monthly training may allow for individually examining short-term variations, which enables capturing more specific and detailed patterns in system behavior.” They have then tested the novel model against the multiple linear regression (MLR) model, a commonly used statistical method for analyzing the relationship between multiple predictor variables and a response variable; and the nonlinear autoregressive (NAR) model, which is based on the ability of neural networks to model nonlinear relationships in time series. “The prediction capacity of the ANN optimized by GA is close to actual measurements, with minimum RMSEs of 13.4 W for the prediction with monthly data for March, 31.8 W for the forecast with seasonal data for February, and 15.6 W for the prediction with annual data for August,” the results showed. “To evaluate which of the five methodologies has had a better performance, the average RMSEs obtained are 24 W, 59 W, 72 W, 53 W, 69 W, and 219 W for the annual, seasonal, monthly GA-FANN methodologies, MLR, NAR, and base ANN respectively.” Following that, the novel FF-ANN method was compared in benchmark tests to state-of-the-art PV energy prediction methodologies, namely QT-MARF, RNN-LSTM, IAMFN, CNN-LSTM, CNN-GRU, ELM, ANN and SVR. They were tested on eight cases, with installations ranging from 1,500 W to 2,700 W. The group found that the new model showed superior performance in terms of RMSE and coefficient of determination (R). “For example, for the day 01/09/2023, the GA-FFANN achieves an RMSE of 20W and an R of 0.99851, while the best benchmark method, QT-MARF in case 1 (1,600W), has an RMSE of 43W and an R of 0.99599. Furthermore, on days such as 08/15/2022, the proposed model achieves an RMSE of 16W and an R of 0.99976, compared to the best benchmark performance in case 4 (1,500W) with RNN-LSTM, which has an RMSE of 20W and an R of 0.99715,” the academics concluded. Their findings were presented in “Optimizing photovoltaic power plant forecasting with dynamic neural network structure refinement,” published in Scientific Reports. The code associated with the approach can be accessed via the Harvard Dataverse repository. |
