Research On Power Load Forecasting Method Based On Big Data Dimensionality Reduction And Optimized Neural Network

The large-scale grid connection of distributed energy has brought challenges to the power system.The stable operation of the power system has strong constraints on load balance.Accurate power load forecasting is of great significance in ensuring the stability and economic operation of power system.It affects many decisions of the power system,such as economic dispatch,automatic power generation control,safety assessment,maintenance scheduling and energy commercialization,etc.Based on the in-depth mining of load data characteristics,the paper proposes a power load forecasting method based on big data dimensionality reduction and optimized neural network.Firstly,in order to avoid the occurrence of"dimension disaster",elastic network（EN）algorithm is introduced to select and reduce the dimension of high-dimensional meteorological big data.By adding L₁ norms and L₂ norms to the penalty items,the elastic net combines the advantages of the least absolute shrinkage and selection operator（LASSO）and ridge regression.It overcome the problem that the dimensionality reduction effect is affected by the collinearity and group effect in the data during the dimensionality reduction process of LASSO,and can extract important variables from high-dimensional meteorological big data and eliminate bad data,thus completing the variable selection of high-dimensional meteorological big data.Through comparative experiments,it is verified that the elastic net works better when dealing with data with strong collinearity and obvious group effect.Secondly,in order to further reduce the complexity of the prediction model,a multi-meteorological factor fusion method based on Elastic network sparse kernel principal component analysis（EN-SKPCA）is proposed.Kernel principal component analysis（KPCA）is used to reduce the dimensionality of the meteorological factors selected by the elastic net,and then the reconstruction principal component is expressed as a regression optimization problem,and the elastic net penalty is added in the regression process,and finally calculate the sparse principal component through iteration.EN-SKPCA can fuse multi-dimensional meteorological factors into a few sparse principal components,and solves the problem that reconstructed principal components of KPCA are difficult to explain.Through the experiment and analysis of the iris flower data set,it is verified that the dimensionality reduction effect of EN-SKPCA is better.Then,from the aspects of timeliness and accuracy,two load forecasting models are proposed,namely,the Flower Pollination Algorithm（FPA）to optimize the BP neural network and the FPA to optimize the Long short-term memory（LSTM）neural networks.First,for the problem of slow convergence speed of BP neural network and easy trapping in local minimums,FPA is proposed to optimize the weights and thresholds of BP neural network.Secondly,in view of the randomness caused by the empirical setting the hyperparameters of LSTM neural network,FPA is proposed to optimize the number of hidden layers,the number of neurons in the hidden layer and the learning rate in the hyperparameters of LSTM neural network.Through comparison experiments and analysis with particle swarm algorithm,it is proved that the flower pollination algorithm is faster and has better optimization effect,which provides a solid foundation for its application in optimizing neural networks.Finally,through multiple sets of comparative experiments and analysis,it is proved that the method in this paper effectively improves the accuracy and stability of short-term power load forecasting,and provides a new idea for regional power load comprehensive factor forecasting under the big data environment.