Hierarchical Power Load Forecasting Based On Quantile Regression Neural Network

Load forecasting of power systems has always been a very important basis for grid regulation and planning,and scholars have never stopped their research on that.The thesis is about hierarchical forecasts of power load demand in different regions.This study is closely related to our smart grid regulation and power resource allocation,etc.The accurate prediction results can provide us with better grid services,and also save resources and rationally allocate resources.This paper mainly pursues the accuracy and comprehensive performance of the model.It is not only limited to the point prediction of power load,but also combines the model with quantiles to generate interval prediction and probability density prediction,thus it will give more complete forecast information.At the same time,by taking into account the hierarchical structure of the region,this paper proposes a more practical and effective hierarchical prediction model.The study in this thesis is based on the GEFcom2017 data set,including the hourly power load requirements of 10 regions.Firstly,the data set is pre-processed,and the XGBoost algorithm is used to filter the variables that affect the power load demand.As a result,8 variables are screened for the following modeling.In this paper,the QRNN model is used to forecast the power load demand.The QRNN model performs well in point forecast,interval forecast and probability density forecast.In point prediction,the average MAPE value is 1.02% for different regions,and the prediction effect is stable at a high level.In interval prediction,a new comprehensive index,CPIA,is proposed to evaluate the effect of interval prediction.It shows that the QRNN model can give a better overall evaluation of interval prediction,and the QRNN model also performs well in different regions.Finally,in the process of probability density prediction,we use the results of QRNN model to make the probability density prediction graph for representative areas,which shows that QRNN model has good prediction effect at different time points.The APL indicators in different areas are also calculated,and the results show that the overall prediction effect is good.It is concluded that the QRNN model is a very robust neural network model,which can not only deal with the data of complex nonlinear relations,but also make interval prediction and probability density prediction.And it can also generate more complete and effective prediction results.Next the ANN model,RBF model and QR model are compared with the QRNN model on the case data.The results show that QRNN model is better than other models in different areas,considering that the models with quantiles can be used for interval prediction and probability density prediction,two new hierarchical time series prediction models,QRNN-hts and QR-hts,are proposed.The QRNN-hts and QR-hts models are based on the independent benchmark prediction models,QRNN and QR models,which are combined with seven hierarchical prediction methods respectively.The new generated models are not only better than the traditional hierarchical time series prediction models: ARIMA-hts and ETS-hts,but also get better performance than their benchmark models.The optimal hierarchical prediction method of the new model is mainly embodied in the combination adjustment method.In addition,the traditional hierarchical time series prediction model only considers the time series data of the target prediction variables,but when the target variables have many influence factors,it can't take into account other dependent variables to improve accuracy.The QRNN-hts model proposed in this paper just solves this problem,and obtains better prediction results than the independent model QRNN.Finally,the QRNN-hts model is used to do the probability density prediction on power load demand,and good prediction result is obtained.