基于贝叶斯优化CNN-BiGRU混合神经网络的短期负荷预测

Short-term Load Forecast Based on Bayesian Optimized CNN-BiGRU Hybrid Neural Networks

高效准确的短期负荷预测在电力市场交易过程中可以提高发电设备的利用率和经济调度的有效性。为解决历史数据当中特征因素较多且特征关系不明显等问题,并充分挖掘负荷数据中时序性特征之间的联系,提出了一种基于贝叶斯优化的卷积神经网络(convolutional neural network,CNN)-双向门控循环网络(bidirectional gate recurrent unit,BiGRU)短期电力负荷预测方法。首先结合特征工程,采用皮尔逊相关系数对负荷特征参数进行初筛,再用递归特征消除(recursive feature elimination,RFE)结合回归模型对特征进行反向选择,完成特征参数筛选;搭建CNN-BiGRU网络模型,并使用贝叶斯优化对其进行超参数调优;将数据输入CNN网络,利用其提取反映特征与负荷之间复杂变化关系的高维特征向量,并将所提特征向量构造为时间序列形式,再输入到BiGRU网络中,完成短期负荷预测。以农夫山泉公司以及美的暖通公司所在地区的真实数据作为实际算例,根据实验结果显示,该模型的预测精度达到95.9%,与其他模型相比具有更好的预测效果。

Efficient and accurate short-term load forecast can improve the utilization rate of power generation equipment and the effectiveness of economic dispatch in the process of power market trading.In order to solve the problem of a large number of features in the historical data and the lack of obvious feature relationships,and to fully explore the connection of temporal features in load data,this paper proposes a convolutional neural network(CNN)based on Bayesian optimization-bidirectional gate recurrent unit(BiGRU)method for short-term power load forecasting.Firstly,a Pearson correlation coefficient is used for the initial screening of load features,and then a recursive feature elimination(RFE)combined with a regression model is used for the backward selection of features to complete the feature parameter screening.Moreover,a convolutional bidirectional gate recurrent unit(CNN-BiGRU)network model is constructed,and hyperparameter tuning is optimized using Bayesian optimization.The data are fed into the CNN network,which is used to extract a high-dimensional feature vector reflecting the complex change relationship between features and load,and the extracted feature vector is constructed into a time series form and fed into the BiGRU network to complete short-term load prediction.The real data sets of Nongfu Spring Company and Midea Company are taken as practical examples,according to the experimental results,the prediction accuracy of the model can reach 95.9%,which has better prediction effect compared with other models.