Multi-step prediction of wind power ramping based on dynamic historical window adjustment

Wind power ramp events, with their sudden nature, can significantly impact grid stability. Traditional forecasting methods often employ fixed-size time windows, which fail to adapt to the intermittent fluctuations characteristic of power ramp scenarios. To address this limitation, this paper proposes an A-TSMixer model incorporating dynamic historical windows for wind power ramp forecasting. Building upon TSMixer's inherent strength in long-sequence prediction, the model introduces an adaptive window mechanism that dynamically adjusts the reference sequence length based on power fluctuation intensity. Additionally, to mitigate the adverse effects of high-frequency noise in raw data, the model employs RTS smoothing algorithm with bidirectional processing for data preprocessing. Experimental results demonstrate that the proposed A-TSMixer model achieves significantly improved forecasting accuracy compared to conventional approaches in wind power ramp prediction.