Gas Load Forecasting Based On EEAE-LSTM Multi-model Fusion Algorithm

With the continuous development of China's natural gas industry,gas load forecasting has become an indispensable task.Reasonable and accurate prediction has important reference significance and research value for gas pipeline network construction,dispatch planning and operation and maintenance management.Therefore,effectively improving the accuracy of gas load prediction is an engineering problem that needs to be solved urgently.Firstly,in order to understand the general situation of gas load forecasting,this paper introduces the research background,significance and related knowledge of gas load forecasting.The reader can have a certain understanding of gas load forecasting,forming a broad and large gas load forecasting concept.Secondly,the gas data is characterized and preprocessed.Since the change law and development trend of the data are the basis for accurate prediction,this paper focuses on the intrinsic characteristics of the load sequence and the external influence factors of the load.The statistical method of autocorrelation plot and unit root test method were used to analyze the unsteadiness of the load,and the annual periodicity,seasonal periodicity and weekly periodicity were analyzed in combination with the gas load trend graph.Then the main influencing factors of the gas load data are calculated according to the correlation coefficient,which provides the basis for the determination of the model input data.In the gas data,some bad data will inevitably occur due to various reasons.In order to improve the accuracy of the prediction,this paper uses an improved horizontal data comparison method to identify and correct the bad data to ensure the overall trend of the load curve and smoothness.Then this paper introduces the LSTM neural network and its optimization method,and establishes a gas load prediction model.The LSTM network has a unique memory structure,it can store the long-term time correlation of data sequences,and solve the problem of information association before and after time series prediction.However,since the sampled time series data is affected by many complicated factors,it contains many noises and has high-dimensional features.However,the LSTM network can only mine time data information in the same dimension in different time periods,and cannot perform multi-dimensional characteristics on the data.extract.In order to solve the problem of data dimensionality reduction,this paper uses the automatic encoder to extract theinfluence factors to obtain its high-order feature representation,mine its deep feature factors and use correlation analysis to verify,establish the AE-LSTM model and The model results were compared and analyzed.However,AE-LSTM does not solve the problem that information mutual interference affects the prediction result of LSTM model.Therefore,the signal processing method EEMD is used to obtain the modal characteristic information of load at different scales,and the mutual influence of noise and other different information is eliminated.In order to reduce the computational redundancy,this paper uses the sample entropy algorithm to reconstruct the components with significantly different trend information according to the similarity of the decomposition result complexity.The EEMD-LSTM model was established and the results were analyzed and evaluated.AE-LSTM and EEMD-LSTM respectively solve two problems in the single LSTM model,namely data dimensionality reduction and reduction of different information interference.The results of two simulation experiments prove the effectiveness of their respective algorithms and improve the prediction effect of the LSTM model.Finally,in order to further improve the accuracy of gas load prediction,and comprehensively solve the problem that LSTM network can't extract multi-dimensional features of time series data,this paper combines the above two optimization algorithms with LSTM network to design and fuse the system to form a new gas load forecasting model based on EEAE-LSTM multi-model fusion algorithm.At the same time,the model completes the feature extraction of the influencing factors and the load sequence,and comprehensively considers the influence of the two kinds of feature information on the load forecasting.The LSTM network can not make up for the defects of feature factor extraction of high-dimensional influencing factors and reduce the interference degree of load information on the final prediction results under different feature scales.In the model prediction process,the feature factors generated by the automatic encoder and the EEMD sample entropy decomposition reconstructed components constitute a component training matrix,and the prediction model is established as the input data of the LSTM network.The EEAE-LSTM model fully considers the characteristics of the influencing factors of gas data,the intrinsic characteristics of the load data itself,and the characteristics of the time scale of the load data.The use of LSTM neural network as the basic underlying predictive network makes full use of its strong learning ability of timing correlation.By establishing the EEAE-LSTM gas load forecasting modeland then comparing it with other models,it is proved that the model has good prediction performance.