Optimizing Photovoltaic Power Plant Forecasting with Dynamic Neural Network Structure Refinement

Abstract

Abstract Energy systems face a challenge characterized by the inherent uncertainty associated with accurate renewable power generation forecasts. Despite the availability of weather prediction methods, achieving precise predictions for photovoltaic (PV) power generation remains a critical necessity. In response to this challenge, this study presents a novel approach that leverages genetic algorithms to optimize PV power plant forecasting. The proposed algorithm dynamically refines the neural network's structure during training, minimizing the mean square error by adjusting parameters such as the number of neurons, transfer functions, weights, and biases. An evaluation of twelve representative days, each representing one month, is conducted using annual, monthly, and seasonal data. Evaluation metrics are employed to assess forecast accuracy, including root mean square error, R-value, and relative percentage error. The research uses MATLAB for modeling, training, and testing, with a 4.2 kW photovoltaic plant utilized for data validation. Results showcase the effectiveness of genetic algorithms, with mean squared errors as low as 20 on cloudy days and 175 on sunny days. Moreover, the genetic algorithm-based artificial neural network optimization methodology achieves forecasting vs. target regressions ranging from 0.95824 to 0.99980, underscoring its efficiency in providing reliable PV power generation predictions.