Accurate minute-level wind power forecasting is identified as critical for the stable operation and economic dispatch of microgrids; however, it remains a challenging task constrained by the inherent intermittency and volatility of wind energy. To address this challenge, a hybrid forecasting model integrating deep learning with dynamic Kalman filtering is proposed in this paper. A dual-encoder Transformer is utilized to parallelly extract long-range temporal dependencies and nonlinear dynamic features from recurrence plots. These heterogeneous features are effectively fused via a cross-modal attention mechanism and subsequently fed into a Long Short-Term Memory (LSTM) network to capture short-term dynamic variations. Finally, real-time state estimation is performed using a Kalman filter with dynamic parameters, guided by the LSTM hidden states to mitigate noise interference. Experimental results based on actual wind farm data indicate that the proposed method achieves a Root Mean Square Error of 32 . 92 ± 0 . 43 kW, a Mean Absolute Error of 18 . 46 ± 0 . 43 kW, and a coefficient of determination of 0 . 967 ± 0 . 001 , significantly outperforming advanced benchmark models such as Temporal Fusion Transformer and Informer. These findings demonstrate that prediction accuracy and robustness are effectively enhanced by the proposed hybrid model, which exhibits superior performance in handling high-frequency fluctuations and abrupt power changes, thereby providing a reliable solution for the real-time management of microgrids. • Hybrid model couples dual-encoder Transformer with Kalman filtering. • Dual-encoder captures temporal trends and phase-space features. • Cross-modal attention effectively fuses heterogeneous data features. • Dynamic filtering enhances robustness against noise and lags.
Li et al. (Wed,) studied this question.