Gas Load Forecasting Based On Fuzzy Coding Genetic Algorithm And Improved LSTM-BPNN Residual Error Correction Model

Natural gas,as a cleaner energy source than coal and oil,has attracted the attention of governments all over the world.In recent years,as the demand for natural gas by governments in various countries has increased,it has also brought huge dispatch pressure to energy managers.When natural gas dispatch is uneven,industries with too little dispatch will lead to huge economic losses.In addition,the gas contract is an advance purchase contract.When the purchase volume is not enough to meet the social demand,the purchaser's breach of the contract will also cause huge economic losses.Reasonable prediction of natural gas load can reduce the above economic losses.In the field of gas load forecasting,many researchers at home and abroad have conducted in-depth research.Most researchers use different single models to predict gas load.These prediction models include statistical models,machine learning models,and deep learning models.In recent years,LSTM and related hybrid models have been widely used.The mainstream solution is to first decompose the gas sequence with a special algorithm,then optimize the model structure,and finally use LSTM to predict the gas load.Among them,there are also a few residual hybrid models based on LSTM.the study.This article will deeply study the gas prediction model based on the residual mixed algorithm.The current mainstream scheme of the residual mixed model includes a single algorithm to optimize the residual model,a special algorithm to decompose the residual sequence and then use the model to predict the residual,and the model directly predicts the residual.These models have been used in wind energy,electricity,gas and other fields,thus proving their effectiveness.This paper proposes an improved residual error model for gas load forecasting.In view of the complex characteristics of the periodicity and randomness of the gas load data itself,the general LSTM hybrid model only optimizes the prediction frontend,ignoring the optimization of the LSTM prediction back-end residuals.The general residual hybrid model and based on LSTM The residual mixed model only optimizes a single model in the mixed model,ignoring the simultaneous optimization of the primary and secondary models,and only the residual sequence is used as the input factor of the residual model,ignoring the analysis of other influencing factors,so this article focuses on LSTM The hybrid model has been improved as follows: first,BPNN is added to predict the residual error of the LSTM backend;second,fuzzy coding genetic algorithm(FCGA)and Adam algorithm are used to optimize the LSTM-BPNN residual correction model at the same time;third,two new ones are proposed The gas residual influence factor,and used for gas residual prediction,which is based on fuzzy coding genetic algorithm and improved LSTM-BPNN residual correction model-LSTM-FCGABPNN(Adam),the basic process is to use Adam to optimize the main The model LSTM predicts the initial value,and then uses the fuzzy coding genetic algorithm to optimize the initial weight and threshold of the BPNN,and then uses the new residual influence factor to input into the Adam optimized BPNN to predict the residual value,and finally uses the residual value corrected preliminary value as The final predicted value.In order to verify the effectiveness of this model in gas load forecasting,this paper uses gas load data from a certain area of Shanghai from 2005 to 2014 to carry out forecasting experiments,and compares the experimental results with the classic models LSTM,LSTM-BPNN,LSTM-BPNN(Adam),LSTM-GA-BPNN(Adam)for comparative analysis,the experimental results show that the MAPE of this model is the lowest with a value of 0.059.This proves that the proposed model has higher prediction accuracy and is of significance for improvement.In the next step,we will continue to study the gas residual mixture model and the influencing factors of gas residual,as well as the calculation efficiency problems caused by the mixture model and the problem of automatically selecting the gas load influencing factors from the original data.