An innovative memory-enhanced Elman neural network-based selective ensemble system for short-term wind speed prediction

The inherent volatility and uncertainty of wind speed can exert high pressure on power grid operations, making an accurate wind speed prediction model essential. Most existing decomposition-ensemble forecasting studies still have certain limitations, where traditional ensemble strategies struggle to effectively coordinate the information differences among subsequences generated by multiple predictive models. Additionally, most ensemble model selection strategies focus solely on fitting ability, neglecting the diversity of predictive models. Therefore, this paper proposes an innovative memory-enhanced Elman neural network-based selective ensemble system for short-term wind speed prediction. Firstly, this paper introduces for the first time the use of a memory-enhanced Elman neural network as an ensemble strategy. This approach effectively memorizes prediction information for each subsequence and discerns informational differences among them, efficiently coordinating and integrating the sub-prediction results. Secondly, a two-stage predictive model selection optimization mechanism is then established and incorporated into the forecasting system, dynamically selecting the optimal sub-prediction model. Finally, this system proposes a self-optimizing preprocessing technique that adaptively selects the appropriate key parameters for denoising and signal decomposition across different datasets. The experimental results from three wind speed datasets collected from different geographical locations and time domains demonstrate that the proposed forecasting system has high predictive accuracy and good stability. Compared with other advanced models, the performance of the proposed forecasting model can be improved by up to 70 %.