Research on a wind power interval prediction method coupling process characteristics and deep transfer learning

• Identify fluctuation types via sliding window and SOM-based classification. • Construct fluctuation-aware prediction intervals using cloud error modeling. • Propose a two-stage transfer strategy for adapting interval models to targets. Wind power prediction in scenarios with limited data, exemplified by newly-built wind farms, faces challenges of insufficient historical data and difficulty in feature transfer. To address this, this paper proposes a wind power interval prediction method that couples process characteristics and deep transfer learning. Firstly, to overcome the issue that traditional fixed time window partitioning methods struggle to effectively capture fluctuation processes, a wind power fluctuation type identification method based on fluctuation process clustering is proposed. This method utilizes a sliding window algorithm and SOM clustering to identify wind power fluctuation types. Secondly, to resolve the problem that prediction accuracy is constrained by the distinct error characteristics of different fluctuation types, the uncertainty of prediction errors for each fluctuation process is characterized. The error cloud model is employed to derive the error intervals, which are then superimposed onto the point predictions from the Informer deterministic prediction model to obtain the interval prediction results. To tackle the issues of insufficient historical data and difficult transfer of process characteristics for newly-built wind farms, a “dual-stage transfer” strategy is designed: first, the well-trained Informer prediction model parameters are transferred to the target domain using a “freeze-finetune” strategy; subsequently, a hierarchical transfer mechanism is designed to transfer the constructed error cloud model. The fusion of these two components achieves interval prediction for the target domain. Case simulations validate that the proposed model and method significantly outperform other models involved in this study.