Short Term Gas Load Forecasting Based On EMD And Ensemble Deep Belief Neural Network

Natural gas has been widely applied because of the advantages of sage,reliable,clean and environmental.With uncasing expanding of gas scale in our country,the requirements on construction of gas pipeline network,maintenance,and gas storage and other aspects of peaking of people are rising.Load forecasting is an important reference for the above work,so load forecasting methods have been widely studied,improve forecast accuracy is also important.First of all,this paper analyzes the possible false modal components in the results of empirical mode decomposition method under the normal sampling rate.Although we can reduce the false component through increasing sampling points by Furier interpolation and sinc interpolation,it can never solve the problem of redundant modalities fundamentally.Therefore,an improved modal function elimination algorithm is proposed in this paper.Checking Each IMF when it is obtained to remove the false modal function and eliminate the error in the IMF,as the process of decomposition signal follows the completeness of the empirical mode decomposition,the energy principle,and the properties of the false modal component.Thereby removing false modal function and eliminating errors in the intrinsic modal function.Then,a deep belief neural network(DBN)prediction model with adaptive learning rate is introduced,which can speed up the training process by reducing the number of iterations during training DBN model.On the one hand,DBN is good at extract the inherent implicit relations between factors which seems unrelated,so that it can improve the prediction accuracy while the value of gas load is affected by many external factors.On the other hand,the traditional DBN algorithm only adjusting the values of weights and bias in the process of pre-training and tuning,but it ignores that the learning rate ? also have influence on the prediction results.This paper studies the adaptive learning rate based on DBN in order to improve the model Training speed and improve the prediction accuracy.Finally,it comes the important part of the article,that the establishment and simulation experiment of EMD based ensemble deep belief neural network prediction model are introduced in detail.AS some data of gas load were missing or misreading by instruments during daily monitoring process,the raw data need to be preprocessed before prediction so that the gas load curve sequence becomes smooth.At the same time,the load data are normalized and the influence factors are indicated by numbers to forming the input matrix of prediction model.According to the nonlinear non-stationary characteristics of gas load sequence,the optimized EMD is used to decompose the load data sequence into multiple IMF.And then,performe DBN modeling and prediction for each IMF to obtain corresponding prediction results.Finally,combine all the prediction results by a linear neural networks to formulate an ensemble output as final prediction result.The results show that the ensemble method can avoid over-fitting as well as improve the prediction accuracy.