Power Load Forecasting Based On Quantile ARMAX Regression Model

With the increasing consumption of electric energy in the whole society,the requirements for power systems are also increasing.In order to achieve the goal of stable power supply,the traditional dispatching method is being gradually changed and a more efficient intelligent regulation method is being adopted,which uses in-depth data mining to achieve the purpose of effective control of the power grid,thereby improving the effectiveness of regulation.On this basis,important information can be provided for power dispatching work.Among them,accurately predicting regional electricity consumption is of great significance for achieving power production,dispatching,and safe power supply.However,due to the fact that most of the forecasting methods used in the field of power load forecasting are mean value regression methods,which often only provide predictions of the average level of the prediction object,and lack the ability to predict situations with high quantiles.However,high quantile prediction can reveal the impact of autovariables on the conditional quantiles of dependent variables,so the application of mean value regression prediction in power dispatching operation risk analysis is limited to some extent.However,users’ electricity consumption will be greatly affected by certain factors,such as a significant increase in electricity consumption under high temperatures in summer,resulting in significant fluctuations in short-term power loads.At this time,mean value regression is no longer applicable.In response to this situation,this thesis first considers using the time series method to predict power load.On the one hand,considering the wide variety of electrical appliances used by various users,and the continuous increase in the proportion of electrical appliances affected by meteorological conditions such as air conditioners,the impact of meteorological factors(temperature,humidity,rainfall,etc.)on power load is increasingly prominent,so meteorological factors are considered as exogenous variables;On the other hand,considering that electricity consumption in the field of power forecasting will be affected by severe conditions,and quantile regression can well reflect the characteristics at both quantiles,and has stronger robustness.Therefore,a quantile ARMAX regression model is proposed by combining time series analysis with quantile regression.After empirical analysis,combined with 1105 days of electricity consumption data from a certain region in China,through variable screening,stationarity testing,white noise testing,and cointegration testing,it is found that there is a cointegration relationship between the electricity load series and the exogenous variable series,and quantile ARMAX regression modeling can be performed.The estimated model is obtained through model ordering and model coefficient estimation,and then the power load under high quantile conditions is predicted for the next month.The prediction results are compared and analyzed with those of the fully connected neural network model and ARIMAX model.From the prediction results,at the 0.9 quantile,the MAPE is 0.0957,and the RMSE is15236.4415.Therefore,in the case of high quantile,using the quantile ARMAX regression model for prediction has a high accuracy and stable effect.Compared with other models,the median ARMAX regression has a better prediction effect than the fully connected neural network model and ARIMAX model.The innovation of this thesis lies in:(1)Extending time series analysis to the framework of quantile regression,combining ARMAX model with quantile regression,a quantile ARMAX regression model is proposed,and its applicability is verified through empirical analysis.(2)Due to the lack of research on short-term power load forecasting at high quantiles,applying this model to this scenario solves the problem of inaccurate short-term power load forecasting at high quantiles,expands research methods and models in the field of power load forecasting,and provides reference and reference for other researchers.