FDNet: Frequency filter enhanced dual LSTM network for wind power forecasting

The inherent volatility and intermittency of wind power present significant forecasting challenges, undermining the efficient integration of wind energy into the power grid. Existing methodologies, notably long short-term memory (LSTM) networks, encounter significant limitations due to their inefficiencies in processing long sequences, difficulties in capturing multi-scale temporal dynamics, and heightened sensitivity to noisy data, which can severely hamper model performance. To address these challenges, This paper proposes the frequency filter enhanced dual LSTM network (FDNet), a novel approach that directly addresses the constraints of the LSTM and improves the accuracy and stability of wind power forecasting. Specifically, FDNet employs the patching operation to divide the original time series into several sub-sequences, potentially boosting the computational efficiency. Furthermore, a specific frequency filter is designed and incorporated into FDNet, effectively reducing the influence of noise. Finally, a dual LSTM structure is employed, which enables FDNet to adeptly discover both short-term local temporal patterns and long-term global temporal patterns inherent in wind power data. Extensive experiments across three datasets demonstrate that FDNet significantly outperforms existing methods, achieving up to 11.0% reduction in mean absolute error and 8.1% in root mean squared error on the HL dataset, underscoring its effectiveness in wind power forecasting.