Short-term Load Forecasting Based On Improved Whale Algorithm And Cyclic Neural Network

With the wide application of electric energy,the economic development of countries is increasingly dependent on electricity,and the demand for electricity's quality is also increasing.Due to the particularity of electrical energy,it cannot be stored in large quantities,and needs to be used immediately.To establish a balance between the production of electrical energy,the transmission of electrical energy and the use of electrical energy,it is necessary to accurately estimate the amount of load in the power system.Therefore,this thesis studies a short-term load forecasting method based on similar day and cyclic neural network and a short-term load forecasting model based on improved whale algorithm to optimize the cyclic neural network.After analyzing the factors affecting the power load,the similar day interval is selected,the meteorological feature vector is constructed,and the similarity day is selected by the grey correlation analysis method.A similar day selection method based on the grey correlation degree is proposed.The cyclic neural network model of short-term electric load forecasting based on similar days is studied.The related parameters of cyclic neural network and BP neural network are analyzed.The dynamic learning rate and small batch training method are used to train the circulating neural network.The effect of the number of hidden layers in the neural network on the network fitting effect.Using the deep learning framework PyTorch,programming language Python to build a cyclic neural network model,and BP neural network comparison experiments,verify the superiority of the similar day-based cyclic neural network model,which can accurately predict short-term power load.Cyclic neural networks tend to fall into local optimum,resulting in poor prediction accuracy of some load points.The excellent global optimization ability of the whale optimization algorithm can prevent the cyclic neural network from falling into the local optimal state,but the whale algorithm does not perform well in the high-dimensional optimization problem.Combining the differential evolution algorithm with high-dimensional global optimization ability with the whale algorithm,each iteration uses the mutation strategy in the differential evolution algorithm to generate mutations for each whale individual other than the optimal whale,surrounded by the shrinking in the whale algorithm.The mechanism or spiral mechanism updates the whale position until the fitness function converges.An improved whale algorithm is used to optimize the weight of the recurrent neural network,and then the gradient descent algorithm is used to train the neural network.Using the deep learning framework PyTorch,the programming language Python builds a recurrent neural network model optimized without whale algorithm,a cyclic neural network model based on standard whale algorithm,and a cyclic neural network model based on improved whale algorithm.Through the three network prediction and test ensemble errors of five typical days,it is verified that the cyclic neural network optimized by the improved whale algorithm is not easy to fall into the local optimal state,and the prediction accuracy is high.