Researchers from Valencia Polytechnic University in Spain have developed a new method for predicting the power generation of PV systems.
The novelty lies in developing a hyper parameter optimization model for feedforward-artificial neural networks (FF-Ann’s) with the help of genetic algorithms (gas) techniques. Hyper parameters are external configurations set before the learning process begins that control the learning process.
“Although integrating machine learning and intelligent algorithms offers considerable potential, there are still challenges in optimizing modeling techniques for complex non -linear systems,” the group said. “The study therefore focuses on tackling the challenges in accurately predicting solar -PV current generation, a complicated task due to the inherent variability of renewable energy sources and the complexity of non -linear energy systems.”
FF-Anns are artificial neural networks (Ann’s) that process information in one direction. They do this from the input layer, via the hidden layer to the output layer, while conventional Ann’s information flows in different ways, using feedback klussen and other memory mechanisms. FF-Ann includes various hyperparameters, namely the number of neurons, transfer functions, input weight, low weight and prejudices.
Gas, on the other hand, is used before the FF-Ann is performed to optimize the configurations. They are inspired by the natural selection processes, starting with hyper parameters that are defined as ‘parents’. These then create offspring, which is a solution for the optimization problem, and continue to make later offspring until an optimal solution is achieved.
“The population of the GA is the various parameter settings that the Ann can accept,” 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 -average square error (RMSE). This is a generally accepted and used practice in literature; This choice is based on its ability to offer a clear and objective size for model performance, which makes comparison and evaluation in different scenarios possible. “
The proposed approach was used in a database of a real installation on the roof in Valencia. The installation consisted of 12 monocrystalline 350 W panels accompanied by weather information from a nearby station. Recording took place from 1 May 2021 to 30 April 2022. Seventy percent of the data points were used for training, 15 percent for testing and 15 percent for validation.
“This analysis included the comparison of the results obtained by training the network using aggregated data sets at annual, seasonal and monthly levels,” the academics added. “The annual training can make it possible to record trends and patterns in the long term, which offers an overview of system behavior over time. Seasonal training can make analyzing seasonal variations possible, taking into account changes in weather conditions and environmental conditions. Monthly training can make it possible to investigate variations in the short term in the short term, making it possible to make more specific and detailed patterns in system behavior. “
They then tested the new model on the multiple linear regression (MLR) model, a commonly used statistical method for analyzing the relationship between multiple predictive variables and a response variable; And the non -Linear AutoRegressive (NAR) Model, which is based on the ability of neural networks to model non -linear relationships in time series.
“The prediction capacity of the Ann Optimized by GA is close to actual measurements, with minimum RMSs of 13.4 W for the forecast 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 Augustus, “the results showed. “To evaluate which of the five methods have had better performance, the average RMSs 24 W, 59 W, 72 W, 53 W, 53 W, 69 W and 219 W are for the annual, seasonal, monthly GA-Fann Methodologies, MLRR , Nar and Base Ann respectively. “
After this, the new FF-Ann method was compared in benchmark tests with ultramodern PV energy gaming methods, namely QT-Marf, RNN-LSTM, Iambn, CNN-LST, CNN-Gru, Elm, Ann and SVR. They were tested in eight cases, with installations ranging from 1500 W to 2,700 W.
The group discovered that the new model showed superior performance in terms of RMSE and determining coefficient (R).
“For the day 01/09/2023, for example, the GA-FANN reaches an RMSE of 20W and an R of 0.99851, while the best benchmark method, QT-Marf in case 1 (1,600 W), has an RMSE of 43W and a R from 0.99599. In addition, the proposed model achieves an RMSE of 16W and an R of 0.99976 on days such as 15/08/2022 compared to the best benchmark performance in case 4 (1500 W) with RNN-LSTM, with an RMSE of 20W and an R from 0.99715, “the Academics concluded.
Their findings were presented in “Optimizing photovoltaic power plants with dynamic refinement of neural network structure“Published in Scientific reports. The code that is linked to the approach is accessible via the Harvard Datse Repository.
This content is protected by copyright and may not be reused. If you want to work with us and reuse part of our content, please contact: editors@pv-magazine.com.