深度集成风速驱动的风电功率预测

A Deep Ensemble Wind-Speed-Driven Wind Power Forecasting

摘要: 风电功率预测是风电场优化发电计划、减少弃风并提高经济效益的有效手段，而精准风速预测是功率预测的关键。本文提出深度集成风速驱动的物理约束共形功率模型以达到对风电功率的精准预测。该模型首先采用长短期记忆神经网络、时间卷积网络和Transformer的深度集成并由XGBoost堆叠估计当前风速，保证多日滚动时风速输入的时序连续；进而在风速与功率映射上构造单调物理功率曲线，之后叠加XGBoost残差学习，结合分位数回归和分割共形方法给出具备置信度控制的功率预测区间。真实风电场预测及对比实验结果显示，该模型有效提升了风电功率预测的精度，区间覆盖率更稳定。且模型能利用该优势将预测区间直接转化为爬坡骤降预警信号，为风电场运维决策提供物理一致性强且工程可用的出力预测方案。

Abstract: Wind power forecasting is an effective means for wind farms to optimize power generation plans, reduce curtailment, and improve economic efficiency, with accurate wind speed prediction being the key to precise power forecasting. This paper devotes to a forecasting framework called Deep Ensemble Wind-Speed-Driven Physically-Conformal Power Forecasting Model, to obtain more accurate prediction of wind power. In the first stage, an ensemble of LSTM, TCN and Transformer models stacked by XGBoost is proposed to estimate the current wind speed. This design keeps wind-speed inputs continuous in time for multi-day rolling forecasts. Thereafter, a monotonic physical power curve relating wind speed to power is employed, with the addition of XGBoost-based residual learning. Then, the power prediction intervals of predefined confidence can be derived based on quantile regression and split conformal method. Experiments using actual wind farm data and comparative analyses have shown that this model effectively improves the accuracy of wind power prediction and ensures a more stable coverage probability. Moreover, the model can leverage this performance advantage to directly convert the prediction intervals into warning signals for rapid power ascent and descent. Therefore, it provides a physically consistent and engineering-usable output power prediction solution for the operation and maintenance decisions of the wind farm.