Enhancing wind power forecasting accuracy through LSTM with adaptive wind speed calibration (C-LSTM)

Wind power constitutes a pivotal component in the quest for carbon neutrality, serving as a principal renewable energy source. Enhancing the accuracy of wind power forecasting facilitates more efficient exploitation of this resource, with deep-learning models, notably Long Short-Term Memory (LSTM), proving effective in advancing forecasting capabilities within this domain. Nevertheless, the accuracy of wind power forecasting is undermined by the inaccurate forecasted wind speed, which diminish the reliability of such predictions. To address this challenge, we propose the model "LSTM with Adaptive Wind Speed Calibration (C-LSTM)", which integrates a mechanism into LSTM that autonomously calibrates forecasted wind speed during the training and inference phase. Specifically, considering the inherent continuity of wind speed, C-LSTM fuses historical wind speed with forecasted wind speed using adaptive weighting parameters. This integration is harmonized with the concurrent updating of the other parameters of C-LSTM, thereby ensuring a dynamic adaptation process that bolsters the model's capacity to coordinate discrepancies between forecasted and actual wind speeds. Experiments conducted across 25 distinct wind turbines have demonstrated that C-LSTM significantly outperforms LSTM in both Mean Squared Error (MSE) and accuracy metrics when the latter directly incorporates forecasted or historical wind speeds. This disparity underscores the efficacy of the adaptive wind speed calibration technique employed within the C-LSTM framework.