Research On Coal Price Forecast Method Based On Temporal Convolutional Network

With the vigorous development of information technology,the application of artificial intelligence in different fields emerges one after another,and more and more engineers are committed to bringing artificial intelligence to more application scenarios.Coal is one of the most important energy fuels in the world today,and it is the fundamental guarantee for our country’s power supply and heat supply.Its price is affected by various factors such as output,supply and demand,transportation,and the end-users’ demand.In recent years,machine learning technology for time series forecasting has made great progress,and the accuracy of some forecasting methods based on neural networks has surpassed traditional statistical forecasting methods.If the coal price can be forecasted by building a big-data model,it can provide a purely data-driven basis for fuel procurement,and further take the initiative in the fuel market,which is conducive to laying a solid material foundation for the deep peak shaving work on the power generation side in the future,resulting in economic benefits.Therefore,this paper uses the method of data mining and analysis to study the factors affecting the change of coal price,constructs the daily forecast dataset of coal price.In order to solve the timeliness problem of the forecast results,the time series multi-input and multi-output structures of the forecast model are defined.A temporal convolutional network model is constructed for experiments,and the coal prices in the future multi-period are accurately forecasted.The details are as follows:(1)The factors affecting coal prices through literature have been investigated.Then relevant data has been collected and well organized.and the feature of working days has been added to ensure the sequence continuity of the collected features,which avoids the fact that some features are not recorded during holidays.This may result in data missing in whole sequence of the dataset.Through correlation analysis,Granger causality test and other data analysis methods,the relationship between each characteristic variable collected and coal price is studied,and the lag coefficient of characteristic variables with lag influence is determined.Based on these results,the daily forecast dataset of coal price is constructed.(2)Based on the above data mining and analysis results,data validity is verified,and I determine the number of input days of multivariate time series required for forecasting,and the maximum number of days that can theoretically achieve a more accurate forecast is ensured,so that the multivariate time serie multiple-input and multiple-output structure of the forecast is also constructed.(3)A deep neural network based on the theoretical basis of temporal convolutional network has been constructed,and forecast experiments on the constructed dataset have been conducted.A Bi-LSTM network and a Bi-GRU network have been constructed for comparison.It is found that for such dataset with short accumulation time,the temporal convolutional network can exert its representation learning ability and the advantage of the residual module to effectively alleviate the over-fitting problem of recurrent neural networks.The model shows better stability while accurately forecasting future coal prices.To sum up,this paper constructs the daily forecast data set of coal price by studying the factors affecting coal price,constructs the multivariate time series multiple-input and multiple-output structure of the forecast work,and conducts experiments by constructing recurrent neural network and temporal convolution network,which provides an effective solution for middle/short-term forecast of coal price in practical application.