Enhancing power grid stability with a hybrid framework for wind power forecasting: Integrating Kalman Filtering, Deep Residual Learning, and Bidirectional LSTM

Accurate short-term wind power forecasting is essential for grid stability and the seamless integration of renewables. However, existing methods often struggle to capture complex temporal dependencies and adapt to real-time data, leading to suboptimal predictions. To address these challenges, this paper employs a dual-stage Kalman filter to decompose time series data into trend and residual components, effectively isolating key patterns while minimizing noise. The trend component is predicted using an adaptive multi-layer perceptron (MLP) model that evolves with real-time data, while the residual component is refined using a deep residual learning-based bidirectional long short-term memory (DRL-Bi-LSTM) model, which excels in capturing intricate temporal dependencies. To further improve prediction accuracy, a recurrent neural network dynamically optimizes the combination of MLP and DRL-Bi-LSTM outputs by adjusting their relative weights. The proposed model was evaluated on a real-world dataset from a Spanish wind farm consisting of 168 hourly observations, achieving a MAE of 0.48 and a RMSE of 0.59 on the test set. These results significantly outperform classical models, including ARIMA (MAE = 2.99), SVM (MAE = 2.13), and LSTM (MAE = 1.78). Compared to hybrid models like KF-ARIMA, MLP-LSTM, and Transformer-based architectures, the proposed model achieved 21–47 % lower MAE and 31–54 % lower RMSE. Additionally, it surpassed recent state-of-the-art models such as SSA-LSTM, CEEMDAN-GRU, and GA-ESN, with average improvements of 31.79 % in MAE and 42.29 % in RMSE across tested benchmarks. Its ability to dynamically adjust to real-time fluctuations, and capture multi-scale temporal patterns, makes it a practical and scalable solution for smart grid applications.