Enhancing photovoltaic power forecasting using deep learning techniques by considering the realized power production period: A case study on 160 kWp rooftop PV system in Thailand

Abstract

This paper proposes an improved PV power forecasting model by applying a modified method to three forecasting techniques: PSO-ANN, GRU, and LSTM. The key enhancement focuses on improving prediction accuracy during nighttime periods when PV systems do not generate power, an area where traditional models often yield high error rates. Reducing these inaccuracies is vital to ensure reliable energy planning and dispatch. The case study uses one-year, minute-by-minute time series generation data from a 160 kWp PV plant in Thailand. Before the modification, LSTM outperformed other methods with an MAPE of 3.91% and an RMSE of 233.09 kW. After applying the modification, adjusting nighttime power outputs to zero, the modified LSTM model achieved improved performance, with an MAPE of 2.97% and an RMSE of 232.64 kW, outperforming the modified PSO-ANN and GRU models. The simulation results confirm that this simple yet effective adjustment significantly enhances prediction accuracy by addressing a key limitation of conventional models: inaccurate power estimates during non-generating periods. Accurate PV power forecasting is essential, particularly in the early-stage investment and operational planning of PV systems.