Frequency Stability Emergency Control Based On Deep Learning

The frequency stability is an important indicator for the safe and stability of the power grid.Due to the increasing maturity of HVDC transmission technology and the commissioning of the DC asynchronous interconnection project,the large power grid has a tendency to be divided into multiple asynchronous small power grids.In addition,renewable energy power generation is increasing year by year in China.Under this trend,the reserve capacity and inertia of the system are shrinking,which may cause the system to have more serious frequency stability problems after severe disturbance.Therefore,the research on frequency stability emergency control after large disturbances of DC asynchronous interconnected power grids is important.Traditional power system model prediction methods are generally difficult to meet real-time requirements which have high accuracy.However,the booming development of deep learning provides a new direction for power system safety and stability analysis.The deep learning algorithm represented by convolutional neural network has achieved good results in processing image data containing spatial features,and is very suitable for the prediction of power systems with spatial distribution characteristics.This thesis has carried out frequency-stabilized emergency DC power support and automatic load shedding control prediction based on convolutional neural network.The main research contents are summarized as follows:This thesis introduces the concept of power system dynamic frequency.Then it introduces the theory of shallow machine learning models that have been applied in the field of power systems,such as support vector machines.At the same time,it introduces Convolutional Neural Networks,one of the representative algorithms of deep learning,including its theory,structure and training methods which shows the application potential in the field of power system forecasting.This thesis proposes an emergency control strategy for frequency stability after power system disturbance based on deep learning.This method uses the operating state data of the power system before and after the disturbance as the input features of the CNN to quickly and accurately predict the amount of DC emergency power support and load-shedding required by the dangerous grid.When making samples for CNN,the spatial distribution of power system nodes and the CNN input image format were taken into consideration,so that the t-SNE algorithm was used to reduce the dimensionality of the system nodes and to project nodes onto a two-dimensional plane.Realize graphical input.When obtaining the output control amount,this thesis uses dynamic simulation combined with linearization model algorithm to generate sufficient sample output data for emergency control,thereby simplifying the difficulty of generating samples and preventing sample imbalance.In view of the advantages and characteristics of deep learning that can output multi-dimensional variables,this thesis innovatively changes the output variables from the previous single-input mode to the matrix form,so that multiple machine learning models are not required,and the frequency prediction is unnecessary.After the sample is made,the example of a DC asynchronous interconnection system built by two New England 39-node systems explains in detail how to build a convolutional neural network and how to realize parameter tuning.A series of prediction results are compared to prove that the proposed CNN network for DC emergency power support and automatic load shedding control after power system disturbances can excellently increase the steady-state frequency of the target area to make sure frequency safety and stability.Comparing with traditional shallow machine learning,this method is faster and more accurate,and maintains more detailed power system information.