Power System Post-disturbance Dynamic Frequency Feature Prediction Based On Deep Learning

The frequency stability of the power system is an important indicator for the safe and stable operation of the power grid.However,the interconnection of the areas and the integration of the large amount of renewable energy bring new risks and challenges to the frequency stability of the modern power grid.On the one hand,the power system areas are interconnected by ultra-high voltage AC and DC transmission line,which greatly increases the complexity of the system,and also increases the risk caused by the fault of the transmission corridors and the trip of the large-capacity generators.On the other hand,a large number of new energy units are connected to the grid,which enhances the system's stochastic fluctuation and safety risks,and weakens the system's inertia and frequency regulation capability The power system frequency prediction methods based on physical model have contradiction between accuracy and real-time requirement,while the traditional shallow machine learning method is difficult to adapt to modern power system analysis with complex nonlinear features.In recent years,the trend of deep learning has brought new ideas to the security and stability analysis of power systems.The deep learning algorithm represented by convolutional neural networks has been successfully applied in the processing of image data with spatial features.It also shows its application potential in processing power system data with spatially distributed features.Focus on power system frequency prediction,the power system post-disturbance dynamic frequency feature prediction algorithm based on convolutional neural network is carried out in this thesis.Taking advantages of physical model and machine learning method,the power system post-disturbance dynamic frequency feature prediction algorithm based on physics-information combination is studied.The main research contents of the thesis are as follows:The basic concepts of dynamic frequency of power system are expounded.The structure and principle of several machine learning models commonly used in power system frequency prediction are introduced and summarized,including support vector regression,artificial neural network and multi-layer perceptron.The basic concepts,structure and training process of convolutional neural networks are elaborated.Through comparison with several machine learning models,the applicability and advantages of convolutional neural networks in the applications of power system frequency prediction are pointed out.Aiming at the application of convolutional neural network in power system frequency prediction,a convolutional neural network input features screening and system operation states database generation method are first proposed.Based on the dynamic frequency process of power system under power disturbance,the important operation data of power system is screened and used as the input features of the convolutional neural network,and the lowest frequency of the system after disturbance is used as the output,the power system operation state database is automatically generated by the PSS/E.Based on the power system operation state data,a convolutional neural network tensor input construction method is proposed.The electrical distance is used to describe the high-dimensional spatial position of the power system nodes,and the t-distributed stochastic neighbor embedding algorithm is used for dimensionality reduction and mapping the system nodes to the two-dimensional plane,and realizes the reconstruction from the original vector operation state to the tensor data,and retains the spatial information of the system node state data.The tensor sample data is used to train the network parameters and obtained the power system post-disturbance dynamic frequency feature prediction model based on convolutional neural network.The method is validated on an improved New England 39-node system with wind farm integration and the actual power grid of South Carolina,which proves the superiority and instantaneity of the proposed method.Aiming at the advantages and disadvantages of physical model and machine learning model for power system frequency prediction,a physical-information combined power system post-disturbance dynamic frequency feature prediction algorithm is proposed.Firstly,a power system equivalent physical model that takes into account of the wind farm is proposed to establish the key physical relationship between power system disturbance and dynamic frequency.Then,the convolutional neural network is used to establish the machine learning model based on the operating state information of the power system.The adaptive neuro-fuzzy inference system is used to organically combine the frequency prediction results of the two sub-models to obtain the lowest frequency prediction result after physical-information combination.The improved New England 39-node system with wind farm is used to verify the accuracy of the proposed method.