A Collaborative Optimization Framework for Wind Farm Clustering and Power Forecasting Using Improved PSO
TL;DR
Validation using empirical data from three wind farms of China Longyuan Group demonstrates that the proposed Clustering-Forecasting collaborative optimization framework maintains stable clustering and forecasting performance under different wind farms, providing a new theoretical basis and technical support for the efficient large-scale wind power clusters.
Abstract
: The strong stochasticity of wind power limits the prediction accuracy of traditional models, making it difficult to meet the demands of refined dispatching. Fine prediction and efficient optimization control of wind farms can be achieved through clustering of wind power clusters. Existing methods rarely consider clustering and prediction jointly, which will lead to loose cluster structures and exacerbated prediction errors. To address these challenges, this paper proposes a Clustering-Forecasting collaborative optimization framework based on the deep coupling of Reinforcement Learning and an improved PSO algorithm. This method embeds the Q-Learning mechanism into the PSO iterative process, enabling full-cycle adaptive dynamic optimization of inertia weights and learning factors through real-time agent-environment interaction. On this basis, the improved algorithm performs multi-feature dimensionality reduction clustering on wind turbines, and independent Long Short-Term Memory (LSTM) and Transformer power prediction models are constructed for each subgroup, achieving a closed-loop optimization from feature decoupling to time-series modeling. Validation using empirical data from three wind farms of China Longyuan Group demonstrates that the proposed framework maintains stable clustering and forecasting performance under different wind farms. On the original wind farm dataset, the proposed method achieves a silhouette coefficient of 0.7742 at the optimal cluster number of k = 9. Furthermore, in the 100-step prediction task, the proposed method reduces RMSE by 60.8% compared with the single LSTM baseline, and by 26.3% compared with the single Transformer. Additional validation results on two newly introduced wind farms further verify the cross-farm robustness and generalization ability of the proposed framework under different wind farm scenarios, providing a new theoretical basis and technical support for the efficient large-scale wind power clusters.