Improving the Quality of Short-Term Forecasting of Photoelectric Power Plant Generation Volumes Based on Recurrent Neural Networks

Abstract

This paper presents an in-depth study of modern approaches to forecasting electricity generation by solar photovoltaic plants (PVPs), which is an important element of the development of renewable energy in the context of decentralized energy supply and the growth of the share of "green" energy in the energy balance. Particular attention is paid to the Sensitivity-Enhanced Recurrent Neural Network (SERNN) model, which combines the advantages of recurrent neural networks with adaptive sensitivity to changes in input meteorological parameters. The proposed approach is based on taking into account key climatic factors, directly affecting the volume of solar energy generation, in particular global solar radiation, ambient temperature, cloudiness, and seasonal fluctuations. It is shown that, unlike classical models such as linear regression or traditional neural networks such as LSTM, the SERNN model demonstrates the ability to flexibly respond to rapid changes in weather conditions due to the built-in mechanism for dynamic adaptation of weight coefficients. The study examines in detail the stages of input data preprocessing that are critical for ensuring the quality of training and forecast accuracy, in particular: scaling (normalization) of indicators, encoding time features in the form of sinusoidal functions to reflect daily and seasonal cycles, as well as filtering or correcting anomalous values that may distort the model results. Based on the real observation data from several photovoltaic installations, the high efficiency of using SERNN in short-term forecasting tasks was experimentally confirmed. In particular, a significant reduction in the mean absolute error (MAE) and root mean square error (RMSE) was recorded compared to alternative approaches. The proposed model also demonstrated the stability of results during periods of high variability in weather conditions, which is especially important for energy planning and network optimization. Overall, the study confirms the potential of deep neural networks with a weather-sensitive mechanism as a promising tool for accurate forecasting of PV power generation. This opens up new opportunities for integrating renewable energy sources into the power system while minimizing technical and financial risks associated with imbalances.