Data-Driven Adaptive Bias Correction for Wind Power Curves: LSTM-Self-Attention Deep Learning Approach

Discrepancies between actual and theoretical power curves of wind turbines impact wind farm operational efficiency and power prediction accuracy. Current bias correction methods are limited in addressing the combined effects of multiple microclimate factors and long-term time series dependencies. This study introduces a novel approach utilizing a multivariate time series modeling technique that integrates LSTM and Self-Attention mechanisms. By leveraging LSTM"s time series modeling capabilities and Self-Attention"s capacity for capturing global dependencies, a bias correction model is developed to address five microclimate variables: wind speed, wind direction, temperature, weather, and wind level. The proposed cascade architecture first employs an LSTM network to extract time series features, followed by the modeling of long-range dependencies between variables using Self-Attention to dynamically correct the power curve. Experimental validation using 19704 sets of operational data from the Jiuquan Guazhou wind farm demonstrates that the proposed method reduces MAE from 0.7041 to 0.6859 and RMSE from 0.937 to 0.9136, representing improvements of 2.6% and 2.5%, respectively, compared to the LSTM benchmark model. The model"s inference time is approximately 45ms, meeting real-time application requirements. This research presents a viable technical solution for rectifying wind power curve deviations.