A Novel Multi-Step Short-Term Power Forecasting Model for Electric Power Systems Based on Deep-Learning

Load forecasting and renewable energy forecasting are fundamental for the optimization and scheduling of new power systems, which play a crucial role in ensuring the safe, stable, and economical operation of the system. This paper proposes a short-term forecasting model based on Transformer-BiLSTM-Patch to meet the demand for short-term forecasting of load and renewable energy. The model designs Patch Embedding to improve attention calculation from traditional point-to-point interactions to segment-to-segment interactions, which is capable of capturing local trends. The model eliminates the mask mechanism during the decoding stage to fully utilize information from the entire time series, that is particularly effective in addressing scenarios specifically where auxiliary variables for future time steps are known (such as weather forecast data). Furthermore, BiLSTM is integrated into the Decoder to further enhance the model's capability in mining temporal features. Finally, comparative tests conducted on multiple datasets demonstrate that the proposed model significantly improves the performance of Transformer models in time series forecasting.