Research And Application Of Natural Gas Load Forecasting Method Based On Deep Learning

With the continuous improvement of the global industrialization level,the demand for energy in human daily production activities is increasing day by day.However,the cumulative energy consumption will cause serious pollution to the environment and economic losses to the global market.In response to this situation,countries around the world are actively investing in and developing new energy.Natural gas,as a clean energy in the transitional period,has been widely used in various fields,but with the increase of gas demand,the contradiction between supply and demand of natural gas is becoming increasingly prominent.The accurate prediction of natural gas consumption load can help relevant departments grasp the future demand for natural gas and achieve accurate energy supply,thus effectively alleviating this contradiction.Therefore,this thesis studies the problem of gas load prediction of natural gas.The main work contents include the following aspects:(1)Data processing and feature analysis.In this thesis,the missing values and abnormal data in the collected experimental data are filled and processed.Then Spearman rank correlation coefficient is used to measure and analyze the load data and its influencing factors to determine the influencing factors with high correlation with natural gas consumption,which can effectively reduce the calculation cost of the prediction model on the basis of reducing the number of features.(2)Research on natural gas load forecasting scheme.Aiming at the non-stationary and nonlinear characteristics of natural gas,this thesis combines the input sequence optimization algorithm which combines the Variational Mode Decomposition(VMD)and Sample Entropy(Samp En)algorithm with the current popular deep learning algorithm.A new gas load forecasting scheme based on the fusion of VMD-SE and CNN-Bi LSTM-Attention is proposed.Firstly,VMD algorithm is used to decompose the original natural gas load data into several eigenmodal components with limited bandwidth.Then the sample entropy algorithm is used to measure and reconstruct the complexity of the decomposed modal components,and three modal components with obvious characteristic information are obtained.Finally,on the basis of adding the attention mechanism,combined with the CNN algorithm with good multi-dimensional feature extraction ability and the Bi LSTM algorithm with excellent time series feature extraction ability,the prediction model suitable for these three modal components is constructed respectively,and the prediction results are finally obtained.(3)Establishment of short-term and ultra-short-term natural gas load forecasting models.Based on the daily and hourly gas load data provided by A natural gas station in A city,short-term and ultra-short-term gas load prediction models based on VMD-SE-CNN-Bi LSTM-Attention are constructed.Starting from the vertical and horizontal dimensions,two sets of experiments containing 16 prediction models were set up to compare the application effects of these algorithm models in two kinds of natural gas load data samples.The experimental results show that the natural gas load forecasting model proposed in this thesis has the best effect and can further reduce the forecasting error.(4)Prediction system development.Develop a natural gas load forecasting system.The system is embedded with the above optimal prediction model to accurately predict the gas load of natural gas,so as to help the natural gas station timely grasp the demand for regulating gas and ensure the supply of natural gas.