基于ADP-DANN的小样本风电集群短期功率预测方法OA

China's wind-power sector has entered a cluster-based development stage.However,the few-shot problems caused by cross-regional spatio-temporal feature mismatches and limited operational data from newly built farms seriously restrict the accuracy and generalization ability of forecasting models.This paper proposes an adaptive dual-predictor domain adversarial neural network(ADP-DANN)that effectively enables few-shot forecasting problems through multi-level cross-domain knowledge transfer.Firstly,a dynamic adaptation layer is designed to flexibly handle heterogeneous inputs with varying numbers of wind farms between the source and target domains,ensuring the integrity of input information.Secondly,by integrating sequence memory and shared feature extractor with context mining,combined with domain adversarial training,efficient extraction and alignment of domain invariant spatio-temporal features have been achieved.Finally,a dual-predictor structure is established to adapt independent forecasting functions to the target domain based on shared features,achieving fine-grained transfer from feature space to task space.The comprehensive experiment based on measured data from multiple stations shows that the proposed ADP-DANN outperforms the baseline model in various forecasting time-scales.The ablation experiment further validates the indispensability and synergistic effects of core components such as the dynamic adaptation layer,dual-predictor,and domain adversarial mechanism.In addition,with the help of t-distribution stochastic neighbor embedding feature visualization and SHAP interpretability analysis,the effective domain alignment ability of the model in the feature space and the physical rationality of its forecasting decisions have been confirmed from both data and mechanism perspectives.This provides a high-performance and clear-mechanism transferable modeling framework for few-shot power forecasting of wind power clusters,which is of great value for intelligent power grid scheduling and efficient accommodation of renewable energy.