Online Transient Stability Assessment And Control Of Power Systems Based On Artificial Intelligence

With the gradual construction of the inter-area power grid with ultra-high voltage AC-DC tie-lines,the power system in China has significantly promoted the nationwide energy efficiency.However,the large-scale interconnected power grid has also proposed several challenges to the system operators.Firstly,due to the nonlinear nature and the complexity of the dynamic power system,it is difficult to identify the critical factors that are related to power system stability.Thus it is not straightforward when making decision to prevent system collapse.Secondly,the dimension of the mathematical model of the dynamic power system has increased drastically.As a result,the traditional model-driven strategy for power system stability analysis is incapable for real-time situation awareness,which further leads to the delayed reaction to instability prevention.Therefore,it is of-urgent necessity to develop the more effective method to enable online stability analysis and protective control.Advanced techniques in the field of big data and artificial intelligence have provided the alternative yet prospective approaches to address the problems in the electricity industry.In this thesis,artificial intelligence(AI)technologies are used to develop the AI-based predictors by learning from the big data generated by power system simulation.These AI-based predictors will then be used for real-time assessment and protective control against power system transient instability.The proposed techniques provide a novel and systematic approach for real-time assessment and protective control in the emerging smart grids.The specific research works and results are as follows:Firstly,the interpretable predictor for transient stability assessment is proposed by using nonparametric statistics.To address the nonlinear nature of power system transient stability problem,the nonparametric additive model is proposed as the regressor against the critical clearing time of fault contingencies.The nonparametric independence screening approach is used to identified the critical factors that are related to transient stability and to determine the functional structure of the additive model.Then the nonparametric statistics-based regressor is well-learned by group Lasso.Numerical results shows that the proposed method can be used to enable realtime evaluation of transient stability and also to identify the critical generation,which assists the decision making in preventive control.Secondly,the probabilistic predictor for transient stability assessment is proposed by using Bayesian deep learning.The variational inference and the Bayes by Backprop algorithm are used together to develop the probabilistic predictor.Numerical results shows that the Bayesian neural network-based predictor can classify the stability status of the power system more accurately and also adaptively evaluate the credibility of the prediction so as to avoid the mis-classification of stability status when the power system operates in the critical condition.Thirdly,a systematical approach for power system online preventive control is proposed by using artificial intelligence and Bayesian optimization.The transient stability constrained optimal power flow(TSCOPF)model is proposed by incorporating the Bayesian neural network-based transient stability predictor as the security constraints.As the Bayesian neural network and the resulting TSCOPF model is non-analytical,the Bayesian optimization is used to compute the optimal strategy of generation rescheduling in real-time manner to prevent the transient instability caused by the credible contingencies.Fourthly,a novel approach based on PMU measurements and attentive relation network is proposed for post-fault transient stability emergency control.In the proposed framework,the gated recurrent unit(GRU)is firstly used to learn the state embeddings of generators from the post-fault PMU measurements.The attention mechanism is used to extract the state embedding of the system without the dependence of the rank of generators.On this basis,the multi-layer perception is used for real-time prediction against transient instability.If the power system is predicted as unstable,the amount of generator shedding for emergency control is further estimated by the attentive relation network.The proposed approach has fulfilled the PMU-based real-time decision making for emergency control against transient instability.