Forecasting Study of Photovoltaic Power Generation Based on Long Short-Term Memory Networks

This paper proposes a hybrid neural network architecture (LSTM-Att) that integrates stacked Long Short-Term Memory networks and an attention mechanism to achieve high-precision forecasting of the ultra-short-term power generation capacity of photovoltaic power stations for the next 24 hours. The dataset of the 11th China National College Students Competition on Energy Economics was selected to construct a machine learning model with meteorological parameters, historical power sequences, and time features as input variables, and predicted power generation as the output variable. The experimental results show that the proposed model has certain advantages in photovoltaic power generation forecasting, and the research findings can provide a new technical path for building adaptive photovoltaic forecasting systems.