A hybrid-driven control strategy for optimized wind farm power dispatch

Abstract

This paper proposes a hybrid-driven active power control strategy for large-scale wind farm (WF) that integrates data-driven and model-driven approaches to optimize power dispatch while reducing fatigue loads and enhancing noise resistance. The strategy employs an encoder-decoder framework in which the encoder, based on a Mixture of Experts (MoE) and Bidirectional Gated Recurrent Unit (BiGRU), captures temporal dependencies from WF time series data, and the decoder, using Graph Attention Networks (GAT), models wind turbine (WT) coupling without explicit mathematical formulations. A Deep Neural Network (DNN) adaptively fuses outputs from the data-driven and Model Predictive Control (MPC)-based strategies, delivering the best overall performance. MATLAB simulations on a WF with 32 × 5 MW WTs show that the proposed method reduces the standard deviation (SD) of shaft torque and thrust force by 21.59 % and 25.64 %, respectively, demonstrating the significant improvements of the proposed method in fatigue load reduction.