Analysis of the impact of data windowing on LSTM-based PV power generation forecasting

Long Short-Term Memory (LSTM) networks have become a commonly used tool in time series forecasting due to their ability to identify complex patterns and relationships within data. The LSTM model's forecasting is based on the input window width, the label corresponding to the next data points in the sequence, and the offset time between the input window and the label. These parameters can influence forecasting accuracy. This paper examines the effect of data windowing on LSTM-based PV power generation forecasting. Different scenarios involving input window widths of 5 minutes, 15 minutes, and 30 minutes were explored. For each input window width, the offset time was adjusted to durations of 5 minutes, 15 minutes, and 30 minutes, respectively. The LSTM model’s performance was evaluated using root mean squared error (RMSE). The measured PV power generation datasets used in this study were collected from a 780 kWP PV power plant located in Thailand. The PV power data were recorded every minute from June 2022 to August 2022, totaling 59,076 data points. At each offset time, the 30-minute input window width provided the lowest RMSE, indicating that longer input window widths lead to better predictive performance. For all input window widths, the RMSE increases as the offset time increases, suggesting that predictions become less accurate as the prediction time moves further away from the input window. Therefore, understanding the balance between window width and offset time is crucial for optimizing LSTM model performance. For practical applications, one should consider the trade-off between computational requirements and desired accuracy.