基于多粒度时间卷积网络的超短期风功率预测

ULTRA-SHORT-TERM WIND POWER PREDICTION BASED ON MULTI-GRANULARITY TEMPORAL CONVOLUTION NETWORK

针对传统风功率预测方法通常基于固定时间粒进行研究,但该类方法往往忽略了其他时间粒度对风功率的影响的问题,提出一种基于多粒度时间卷积网络(MGTCN)的超短期风功率预测方法,使用时间卷积网络来挖掘多粒度视角下的风力机数据特征,并设计多粒度特征融合模块来增强模型的鲁棒性,提高风功率预测精度。首先,利用随机森林算法(RF)得到与输出功率相关性较强的部分特征数据;然后,对筛选后的特征数据进行多粒度划分,通过时间卷积网络(TCN)提取各个粒度的独立特征。最后,使用挤压激励网络(SENet)对不同粒度特征进行自适应加权融合,得到最终预测值。采用中国某风场数据进行算例分析,结果表明相较于其他方法,所提方法在24步预测任务和6步预测任务上取得了最佳的预测性能,具有较高的准确性和稳定性。在24步预测任务上归一化均方根误差、归一化平均绝对值误差和决定系数指标分别为0.152、0.108和0.7214,在6步预测任务上各指标分别为0.1027,0.0683和0.8717。

To address the problem that traditional wind power prediction methods are usually based on the fixed time granularity and often ignore the influence of other time granularity.an ultra-short-term wind power prediction method based on multi-granularity time convolution network(MGTCN)is proposed.In our proposed method,the temporal convolution network(TCN)is used to mine the characteristics of wind turbine data from a multi-granularity perspective,and a multi-granularity feature fusion module is designed to enhance the robustness of the model and improve the accuracy of wind power prediction.Firstly,the random forest algorithm(RF)is used to select related feature data with strong correlation with the output power.Then,the filtered feature data is divided into multiple granularities,and the independent features of each granularity are extracted through TCN.Finally,the squeeze and excitation network(SENet)is used to adaptively weight the fusion of different granularity features to obtain the final prediction value.A wind field data in China is used for example analysis.The results show that compared with other methods,our proposed method achieved the best performance with high accuracy and stability on both the twenty-four-step prediction task and the six-step prediction task.Specifically,in terms of three common prediction performance metrics,including normalized root mean square error,normalize mean absolute error and R2,our proposed method obtained the best performance with 0.152,0.108 and 0.7214,respectively on the twenty-four-step prediction task.For the six-step prediction task,it achieved 0.1027,0.0683 and 0.8717,respectively.