军事知识服务平台
基于ST-GAT与两阶段误差修正的区域风电场群超短期功率预测
Ultra-short-term Power Prediction for Regional Wind Farm Clusters Based on ST-GAT and Two-stage Error Correction
摘要 针对区域风电功率预测面临的全场建模维度高,以及线性升尺度方法受"恒定比例假设"局限而产生非线性系统偏差等问题,提出一种基于时空图注意力网络(spatio-temporal graph attention network,ST-GAT)的"降维升尺度+物理残差修正"两阶段超短期预测方法.首先,利用K-means聚类选取代表站,并结合线性升尺度策略高效获取区域功率的基础趋势;随后,鉴于尾流效应证实了场内上下游机组间存在受气象驱动的动态空间关联,构建风向引导的ST-GAT误差修正模型.该模型引入风速、风向作为物理驱动因子,利用图注意力机制挖掘空间拓扑的演变规律,从而对升尺度偏差进行精准补偿.基于某大型风电场实测数据的算例表明,所提方法在保证计算效率的同时显著提升了预测精度,决定系数R2达0.710 9,相较于基准升尺度方法提升约11.4%,且归一化均方根误差降至14.98%.所提方法可有效提高复杂风况下区域风电功率预测的准确性.
Abstract:
Aiming at the challenges of high dimensionality in full wind farm modeling for traditional regional wind power prediction,as well as the nonlinear systematic deviations induced by the constraint of constant proportionality assumption in conventional linear upscaling approaches,this paper proposes a two-stage ultra-short-term prediction method combining dimensionality-reduction upscaling and physical residual correction based on the spatio-temporal graph attention network(ST-GAT).Firstly,the K-means clustering algorithm is employed to screen representative wind turbines,and a linear upscaling strategy is adopted to efficiently acquire the basic variation trend of regional wind power.Furthermore,given that the wake effect confirms meteorology-driven dynamic spatial correlations between upstream and downstream turbines inside wind farms,this study constructs a wind direction-oriented ST-GAT error correction model.By introducing wind speed and wind direction as physical driving factors,the established model adopts the graph attention mechanism to mine the evolutionary characteristics of spatial topological relationships,so as to realize precise compensation for upscaling deviations.Numerical experiments based on field-measured data from a large-scale wind farm verify that the proposed method can significantly boost prediction accuracy while maintaining high computational efficiency.The achieved coefficient of determination is 0.710 9,and the normalized root mean square error is reduced to 14.98%.Compared with the benchmark upscaling method,the value of coefficient of determination is improved by about 11.4%,which effectively enhances the prediction accuracy of regional wind power under complex wind conditions.
刊 名 广东电力 ISTICPKU
年,卷(期): 2026, 39(10)
分类号: TM614
机标分类号: TM614 TM715 TK81
在线出版日期: 2026年10月14日
基金项目: :广东电网有限责任公司电力调度控制中心科技项目,