Research On Online Situation Of Power System Frequency Based On Machine Learning

With the introduction of new energy and the pressure of energy saving and emission reduction,my country’s power system is undergoing a series of changes and challenges.The power system frequency is an important operational index reflecting the global power surplus and deficit of the AC power grid.The traditional research methods of power system frequency(such as time domain analysis method,direct method,equivalent model method,etc.)are based on the idea of physical theory analysis,and there are problems that the construction of the model is complicated,and the calculation amount and calculation accuracy cannot be taken into account.Some artificial intelligence-based methods(such as neural networks,decision trees,and Bayesian networks)have problems with non-linear power-oriented problems,such as insufficient generalization ability and low accuracy.Traditional machine learning methods include gradient lifting decision tree(GBDT)and support vector machine(SVM).GBDT has strong generalization ability.SVM can solve decision problems in high-dimensional space.Both are widely used in prediction and classification tasks.Deep learning is a type of machine learning.The frequency dynamics of the power system after being disturbed is a time series.Since deep learning LSTM and RBF can capture long-distance associations in the sequence,it is suitable for processing sequence prediction problems.Based on the basic concept of power system frequency,this paper analyzes the static frequency characteristics and dynamic frequency characteristics of the power system.Us-ing PSASP based on the time domain method to build a network model for power system transient stability simulation,and introducing CEPRI36 and IEEE39 standard node exam-ples to analyze the frequency response characteristics after disturbance.The consistency of the data and the frequency characteristic curve of the two calculation examples can in-dicate that the disturbance has a strong correlation with the frequency change,and has a weak correlation with the voltage and power angle difference.Therefore,it is proposed to learn these characteristics by machine learning and use them to predict the frequency of the power system.According to the time series of the frequency after the disturbance,the traditional machine learning model GBDT and SVM model are constructed respectively to predict the frequency.Since neural networks are suitable for processing sequence problems,LSTM and RBF models for deep learning are constructed for frequency prediction.Theoretical analysis and prediction results show that the frequency after power sys-tem disturbance can be studied and analyzed through frequency time series.Both machine learning method GBDT and deep learning method LSTM can be used to study the power system frequency situation.The LSTM prediction model has the highest accuracy,and the training and prediction process is relatively stable? the prediction speed of the GBDT model is faster than the LSTM model.The SVM model has the problem of poor prediction accuracy,and the RBF model has the worst frequency prediction effect.Both the training set loss and the test set loss are relatively large,and the model is not easy to train.