Research On Generalized Load Modeling Based On Optimized Neural Network

Under the guidance of China’s new energy development strategy,distributed generation technology,power electronics technology and intelligent control technology have been rapidly improved.Due to the integration of distributed generation systems,the miscellaneous load area is further increased,which forces the modeling work to develop towards the generalized load direction.How to build the generalized load model effectively and improve the simulation accuracy is an urgent problem in the field of power system research.In this paper,two new methods of generalized load modeling based on optimized neural network are proposed.To solve the stochastic and complex problem of generalized load,a modeling method based on Auto-encoder fused Extreme Learning Machine(AE-ELM)was proposed.The generalized load system is regarded as a whole external system,and the voltage and power data at the load nodes of the system are extracted as the input and output of the neural network to construct the generalized load model based on the auto-encoder fusion extreme learning machine.Firstly,the feature value can be extracted by the characteristic of auto-encoder reducing the dimension of input data,and then the structure of the auto-encoder is obtained by its characteristic of minimizing the reconstruction error.The structure is taken as the input structure of the extreme learning machine,an extreme learning machine structure with optimized number of hidden nodes is obtained.Then,the network is guaranteed to converge to the optimal value through adjusting the weight from hidden layer to output layer according to the supervised learning method of the extreme learning machine.The modeling test of the generalized comprehensive load area containing the doubly-fed induction wind power system and the battery energy storage system is carried out,and the accuracy and feasibility of the proposed method are verified by comparing the fitting results with the commonly used neural network training.On the basis of the above research,this paper explores how to establish the generalized load model which is suitable for different load components at the same time,and a modeling method based on SOA-LSTM-Da NN was proposed.Firstly,the generalized load system established previously is taken as the source domain model,and a new distributed generation system is added to form the target domain model.LSTM network(Long Short-Term Memory)is pre-trained by source domain data set.Meanwhile,in order to further improve the approximation ability of the model,Seagull Optimization Algorithm(SOA)is used to optimize the weight and threshold of the hidden layer module of LSTM.Then,a domain-adversarial Training of Neural Network(Da NN)was added after the LSTM Network to form the LSTM-Da NN Network,and the distribution distance between the source Domain data and the target Domain data was calculated by using the maximum mean difference method.As the measurement index of transfer learning,parameters of the pre-training model were transferred to LSTM-Da NN,fine-tuning training was carried out to minimize the loss,and the generalized load model of the target domain was finally obtained.By comparing the fitting results of RNN,LSTM,SOA-LSTM and SOA-LSTM-Da NN under different degrees of three-phase short circuit faults and different fan capacities,the effectiveness of the modeling method was verified.