A novel data-driven multi-step wind power point-interval prediction framework integrating sliding window-based two-layer adaptive decomposition and multi-objective optimization for balancing prediction accuracy and stability

Wind power is critical for large-scale grid-connection and carbon neutrality, so accurate and stable predictions are needed to address the inherent randomness and complex coupling of wind power data. The study introduces an innovative data-driven point-interval prediction framework to overcome the limitations of current models that focus only on prediction accuracy, which leads to large uncertainties by ignoring the stability required for predictions. The proposed framework begins with an outlier processing mechanism and employs a new sliding window-based two-layer adaptive decomposition strategy that avoids information leakage while decomposing the wind power data into regular subsequences. These subsequences are then classified using Lempel-Ziv complexity analysis to minimize computational redundancy. Advanced models—including Inverted Transformer (iTransformer), TimesNet, Mamba2, and Sample Convolution Interaction Network (SCINet)—are strategically deployed to a new integrated forecasting method to capture the intricate temporal dependencies within the decomposed subsequences. A pioneering multi-objective optimization algorithm is then used, which in point prediction serves as a two-stage integration module for weighted fusion of model outputs to balance the accuracy and stability of point prediction. In interval prediction, it is used as a parameter optimization module to optimize the bandwidth parameter of the interval prediction model to balance the accuracy and stability of interval prediction. Through empirical analysis and statistical tests on two real datasets, the framework significantly reduces the mean absolute error (MAE) by 27.32 % and 58.51 %, and the error standard deviation (STD) by 27.16 % and 77.72 %, compared to iTransformer. The integration forecasting architecture establishes a robust framework for high-precision and high-stability point-interval forecasting. This framework provides grid operators with enhanced decision-support capabilities through reliable uncertainty quantification, making a substantial contribution to improving the operational stability of power systems and optimizing renewable energy utilization efficiency.