Emergency Control Strategy Of Transient Instability Of Power System Based On Deep Learning

With the development of the "14th five year plan" power grid construction,the traditional power grid is gradually transforming to the energy internet.The grid environment is increasingly complex,and the penetration rate of distributed energy such as photovoltaic,wind power and electric vehicles is gradually increasing,which makes the safe and stable operation of the power system face new challenges.The traditional emergency control method based on physical model is difficult to adapt to the current large-scale interconnected power grid with high dimension,high randomness and high coupling.Big data of power grid provided by wide area measurement system and the deep learning theory provide a new idea for real-time emergency control after power system transient instability.Based on deep learning,a relatively complete transient instability emergency control architecture is proposed in this thesis.The main research contents are as follows:(1)In order to realize the fast prediction of power system transient stability,a dynamic time series transient stability assessment method based on CNN is proposed.Firstly,three kinds of input feature sets are extracted by different feature selection methods,and the trajectory cluster features with the best discrimination are selected as the input of the model.Secondly,the mapping relationship between the input features and transient stability level is mined,and the CNN evaluation model is established,which has high prediction accuracy,robustness and generalization ability,and can furtherly fit the degree of system instability.Finally,combined with sliding time window and reliability index,the dynamic time series transient stability assessment process is established to ensure the timeliness and accuracy of prediction.(2)In order to identify the severely disturbed generators after power system transient instability,a critical generator identification method based on LSTM power angle trajectory prediction model is proposed.Firstly,considering the time correlation of the unstable power angle trajectory,the LSTM trajectory prediction model is established,and the influence of different model parameters on the fitting effect is explored.Then,based on the coherency of generator cluster,the critical generator cluster of predicted trajectory is identified to provide alternative action space for emergency control.Finally,examples are given to verify the precision and rapidity of the LSTM trajectory prediction model and the accuracy of critical generator cluster identification.The proposed method can detect and visually divide the unstable generators in advance,leaving more decision-making time for emergency control.(3)In order to realize fast emergency control after transient instability,a closed-loop emergency control scheme based on improved Alex Net is proposed.Firstly,the optimization model of emergency control is established,and the sensitivity index of emergency control action is defined.Secondly,according to the characteristics of emergency control,the Alex Net algorithm is improved,and its structure is determined by layer by layer optimization.The sensitivity prediction model of emergency control based on improved Alex Net is established,which can accurately and quickly predict the effectiveness of emergency control action.The model has high robustness and generalization ability.Then,considering both rapidity and accuracy,combined with the transient stability assessment and trajectory prediction method proposed above,a closed-loop sequential emergency control scheme is proposed.Finally,simulation results verify the effectiveness of the proposed method.(4)In order to further solve the "curse of dimensionality" problem of action space in actual power grid,a generator tripping scheme based on deep reinforcement learning is proposed.Firstly,the architecture and algorithm of generator tripping after transient instability based on DQN are designed,and the dimension reduction of action space is realized by using the method of instability pattern clustering and multi-agent integrated modeling,which improves the training efficiency.Then,an integrated test platform for the interaction of agent and simulation environment is built,and the sequential generator tripping control is realized.Finally,several examples are given to show that the proposed method can make the system recover stable quickly.