A context-driven dynamic modeling method for ultra-short-term wind power forecasting

Abstract

Ultra-short-term wind power forecasting is a critical technology for ensuring secure and stable operation of power systems and promoting new energy integration. Current research usually employ offline models, some scholars study on online modeling strategies to address the problem of concept drift, but have difficulty in determining model update timing and catastrophic forgetting. Therefore, a context-driven dynamic modeling method for ultra-short-term wind power forecasting is proposed in this paper, including three components: initial model construction, online concept drift detection, and online model fine-tuning. First, a sequence-to-sequence model is adopted to construct the initial forecasting model based on all historical power data. Then, the divergence degree of contextual relevance among samples similar to model’s inputs is calculated for online concept drift detection. Finally, numerical weather prediction (NWP) are introduced to obtain a sample set with both similar input power and NWP wind speed if concept drift is detected, thereby enabling online model fine-tuning. Operational data of two wind farms in China is used to verify the effectiveness and robustness of the proposed method. Results show that, compared with offline and three traditional online methods, the proposed method improves forecasting accuracy by 15.60% to 17.92% and 11.92% to 15.30% under five basic models, respectively, when root mean squared error is used as the evaluation index.