Power Load Forecasting Based On Similar Situations And Attention Mechanism Improved Convolutional Neural Network

Accurate load forecasting is the basis of ensuring the safe,economic and stable operation of power system,which is of great significance to power system planning,dispatching and safety management.However,with the renewable energy power into the network,people have higher and higher requirements for the accuracy and stability of prediction,and the traditional prediction methods are gradually difficult to meet the needs.Therefore,this paper proposes an improved convolutional neural network prediction method based on similar situation and attention mechanism.Firstly,the problem of feature extraction of power load forecasting is studied,and the method of candidate feature determination based on load characteristic analysis and "situation" feature extraction is proposed.Through the variable granularity analysis of historical data,it is found that the power load has multiple quasi periods such as day,week and year.There is not only a strong correlation between the loads at the adjacent time,but also a strong correlation between the loads at the "long distance" time in the quasiperiodic adjacent phase.Then,the concept of "forecasting situation" is proposed by comprehensively considering the "historical inside information" and"time and space outside environment" of power load,and the feature extraction methods of"historical inside informtion" based on correlation analysis and "time and space outside environment" based on influencing factor analysis are designed.Finally,the candidate feature set of load forecasting is obtained by combining the "historical inside information"feature and "time and space outside environment,feature.Secondly,the problem of feature reduction and power load forecasting is studied,and an improved convolutional neural network forecasting method based on attention mechanism(AMICNN)is proposed.In this method,attention mechanism is introduced in convolution layer,and the weight distribution between channels is carried out according to attention.The convolution structure is improved by multi-channel weighted synthesis,and the effect of feature reduction is optimized.Furthermore,the idea of feature reduction based on multiple pooling weighted synthesis is proposed,and a new pruning mean pooling method is designed.A new attention layer is introduced,and the weight distribution is based on the attention value,and the fusion of different reduction results is realized by weighted synthesis.Finally,the improved CNN prediction method based on attention mechanism is given,and its effectiveness is verified by a case study.Thirdly,in order to further improve the prediction performance of AMICNN,the classification and organization of training data is studied,and an improved multi view variable weight collaborative fuzzy clustering method is proposed.In this method.the view weight based on information entropy is introduced,and the weight adjustment term based on view state is designed from the perspective of maximizing the separation between classes,that is,state adjustment.Then,the variable weight is constructed by combining entropy weight and state adjustment weight,and the penalty term of variable weight synthesis based on the clustering effectiveness of each view is introduced into the objective function.The fuzzy clustering method based on multi view variable weight collaboration is established.The effectiveness of the method is verified by the simulation on open data sets and power load data sets.Fourthly,the prediction performance improvement of AMICNN is studied,and the prediction method based on similar situation is proposed.This method introduces the concept of "similar situation",and measures the situation similarity by means of the distance between the predicted situation state and its historical situation state.Then,the multi view collaborative fuzzy clustering method based on variable weight comprehensive improvement is used to classify situations and select similar situations,and the prediction method based on similar situation improved AMICNN is given.Finally,simulation experiments are carried out to verify its effectiveness and advanced nature.The results show that compared with CNN and BPNN,the MAPE of SSAMICNN is reduced by 0.84%and 2.15%respectively,and the prediction error variance is reduced by 0.95%and 1.15%respectively.