基于KPCA-CNN-DBiGRU模型的短期负荷预测方法

A short-term load forecasting method based on KPCA-CNN-DBiGRU model

本文针对已有神经网络模型在短期负荷预测中输入维度过高、预测误差较大等问题,提出了一种结合核主成分分析、卷积神经网络和深度双向门控循环单元的短期负荷预测方法。先运用核主成分分析法对原始高维输入变量进行降维,再通过卷积深度双向门控循环单元网络模型进行负荷预测。以第九届全国电工数学建模竞赛试题A题中的负荷数据作为实际算例,结果表明所提方法较降维之前预测误差大大降低,与已有预测方法相比也有大幅的误差降低。

Due to the difficulty of storing large amounts of electric energy and the power load demand changing at any time,strict requirements are presented for the dynamic supply-demand balance of power generation and power load.Therefore,improving the accuracy of load forecasting is beneficial for improving the utilization rate of power generation equipment and realizing the balance of power supply and demand efficiently.Short-term load forecasting of power systems is very important for economic dispatch,power market transactions and the smooth operation of the power grid economy.With the development of intelligentsia and digitalization of power grid construction,the requirement of load forecasting accuracy is gradually increasing.Improving the accuracy of short-term load forecasting has become a practical problem to be resolved urgently.In recent years,with the continuous expansion of the power grid scale and the increase in power generation equipment,massive amounts of datasets have been provided for load data analysis and forecasting.In the face of the massive immense data,the accuracy of the traditional load prediction methods cannot meet the current requirements of the intelligent development of power grids.However,in recent years,deep learning models have developed rapidly and have been extensively used in various fields.They have also been well-applied in the field of power load forecasting.Because the factors that affect power load include meteorological,date,season and other characteristics,the input dimension of the model is too high.Only applying a neural network predictive model based on deep learning is often not conducive to improving the prediction accuracy.Hence,at present,many scholars have adopted some common traditional linear dimensionality reduction methods to reduce the original input high-dimensional feature data first,and then begin the load prediction task.However,there is often a nonlinear coupling relationship between the input high-dimensional feature data of the load prediction model.Using the traditional linear dimension reduction methods for high-dimensional input data dimension reduction is unreasonable here and cannot better retain the original data information,which is not helpful for improving the load prediction accuracy.Aiming at the problems when applying the existing neural network models based on deep learning in short-term load forecasting,such as original input data with a high dimension and a large prediction error,this paper proposes a short-term load forecasting method combining a kernel principal component analysis(KPCA),a convolutional neural network(CNN)and a deep bidirectional gated recurrent unit(DBiGRU).Taking the load data in question A of the 9th National Electrotechnical Mathematical Contest in Modeling as a practical example,the relationship between the 96-point power load data and the five-dimensional meteorological characteristic data is analyzed first.By analyzing the grey correlation degree and the scatter diagram of the load data and the five-dimensional meteorological characteristic data,it is found that there are strong correlation and nonlinear coupling relationship between the load data and the meteorological data.Second,several common nonlinear dimensionality reduction methods(KPCA,TSNE,LLE,ISOMAP)are utilized to reduce the dimensionality of high-dimensional nonlinear input data in load forecasting.Compared and analyzed the results of dimension reduction,the KPCA method is adopted to reduce the dimension of meteorological data from five dimensions to three dimensions.Then,the processed data is input into the neural network model,and the grid search method is used to fine-tune the convolutional layer,the sliding window length,batch size and other hyperparameters.Finally,the data after dimension reduction are input into the model after hyperparameter tuning for testing.The features of the load data are captured by the CNN,and then DBiGRU is applied for 96-point prediction.The results show that the prediction error of our proposed method is much lower than that of the method before dimension reduction,which verifies the rationality of dimensionality reduction for the original load data.Compared with other nonlinear dimensionality reduction methods,the prediction error of our method is also low,which verifies the superior efficiency of using KPCA for dimensionality reduction of load data.Compared with other existing prediction methods based on deep learning network models,the error is also greatly reduced,which further verifies the high efficiency and rationality of our proposed model in this paper.In addition,several dates with different seasons and different date types are randomly selected from the test set of working days and holidays.The results demonstrate that the prediction accuracy of the model in this paper is still better than other models on random days.