Short-Term Wind Power Forecasting Based on a BiLSTM-Transformer Hybrid Architecture

Mingjia LV

Abstract: High-accuracy short-term wind power forecasting is essential for stable power grid operation. This paper proposes a BiLSTM-Transformer hybrid framework for time series prediction to address wind power volatility. The method leverages deep learning by integrating refined feature engineering (anomaly detection and PCC) with a dual-channel architecture. It captures local temporal dynamics via BiLSTM and long-range dependencies via Transformer, followed by an adaptive feature fusion mechanism. Experiments on a real-world SCADA dataset yield an MAE of 181.24 kW, RMSE of 202.58 kW, MAPE of 6.16%, and a coefficient of determination of of 0.916. Compared to standalone LSTM and Transformer models, the framework reduces RMSE by 24.5% and 11.4%, respectively, demonstrating superior performance in short-term forecasting.

Keywords: Wind Power Forecasting, BiLSTM-Transformer, Deep learning, Short-term forecasting, Time Series Prediction, Feature Fusion