Load Prediction Method based on STL-LSTM during Spring Festival

Accurate short-term load forecasting plays an important role for power systems safe and stable operation as well as economic optimal dispatch. Due to the limited historical load data and the random factors, the holidays load forecasting accuracy is lower. This study proposes a hybrid load forecasting model based on Seasonal-Trend Decomposition using LOESS (STL) and Long Short-Term Memory (LSTM) neural networks, referred to as the STL-LSTM model. The proposed approach firstly decomposes the original load time series into trend, seasonal, and residual components using the STL additive model. Each component is then individually forecasted using the LSTM neural network, after which the prediction curves are aggregated to form the final load forecast. Experimental results demonstrate that the STL-LSTM model outperforms traditional approaches in terms of prediction accuracy, particularly during holidays, where load patterns are more unpredictable.