Research On Power System Security Dispatch And Control Based On Data-Driven Methods

With the construction of new power systems,the characteristics of AC-DC interconnected large power grid are gradually evolving: increased system uncertainty,complex cross-regional coupling mechanisms of power grids,and deep integration of cyber system and physics system.Power system security issues are becoming more and more prominent: the steady-state power flow mode is changeable,and the risk of power flow exceeding the limit is increasing;dynamic power grid characteristic events propagate quickly,have a wide range,and complex mechanisms,making event identification more difficult;transient stability analysis models are highly time-varying and difficult to analyze the continuously evolving power system.While the information communication system brings convenient data interaction,its security risk also has a significant impact on the security issues of the above-mentioned dispatch scenarios.Traditional power grid security control methods are often based on the technical route of "offline analysis,real-time matching" or "online rolling analysis,real-time matching",which is difficult to cope with the continuously evolving new power system.Therefore,the use of artificial intelligence(AI)for complex data online analysis and processing capabilities will help to achieve intelligent control of power grid security.However,there are still many deficiencies in the current data-driven power grid dispatch and control researches.It is worth exploring the deep mining of the spatio-temporal relationship of power grid data,the integration of existing domain knowledge into AI models,and the update of the AI model in time-varying scenarios.Based on this,under the support of the National Key Research and Development Program "High-performance Analysis and Situational Awareness Technology for Interconnected Large Power Grids" and the National Natural Science Foundation of China International Cooperation Project "Smart Grid Analysis and Control Technology Based on Artificial Intelligence",this thesis focuses on the security dispatch of new power system and contents are as following:(1)To minimize the risk of transmission line overload with high proportion of renewable energy,a transmission line overload preventive control strategy based on safe reinforcement learning algorithm is proposed.Firstly,a constrained Markov model for grid power flow dispatch is established for reinforcement learning agent training,and an optimization algorithm is used to find the theoretical optimal solution as a benchmark;secondly,considering the spatial distribution of power flow and the temporal characteristics of renewable energy,A spatiotemporal awareness network combining long-short-term memory network and edgeconditioned convolutional neural network is proposed;finally,in order to ensure the power flow control strategy meets the security constraints of power grid operation,safe reinforcement learning is used to conduct control agent policy training.It realizes the online dispatch of power flow,and solves the problem of transmission line overload.(2)To identify the cross-regional characteristic events in the AC-DC interconnected power grid,the real-time identification of power grid characteristic events and the prediction method of chain events are analyzed.Firstly,considering the diversity and uncertainty of the operation mode,the characteristic event chain data set of the AC-DC interconnected grid is constructed based on the time domain simulation;secondly,based on the physical mechanism of the event and the characteristics of the dynamic response,a physical-data integrated method is proposed;Then,according to the data characteristics of the coupling interval between events,the deep image recognition algorithm is used to discriminate the causal relationship of the event chain,and the evolution path of the power grid characteristic event is constructed.The real-time identification of dynamic characteristic events of the power grid and the prediction of chain events are realized,and accurate and fast event information is provided for the control center to formulate power grid control measures.(3)To address the mismatching problem between AI models for power grid transient stability and the time-varying power grid,an inheritance-based method for transient stability prediction is proposed.Compared with existing data-driven research on power grid stability prediction,this method is suitable for the transitional phase of power grids with incremental transient data or fundamental changes,filling the gap in existing researches.In order to address the time-varying nature of transient samples and power system stability characteristics,this thesis proposes a deep inheritance method that tuning the parameters of existing prediction models and a breadth inheritance method that considers the extendibility of sample and feature sets.The deep inheritance based on incremental learning uses newly added samples to reduce the training time required for model updates.The breadth inheritance based on transfer learning establishes a feature set and sample set transfer channel between the source system and the target system to construct a transient stability prediction model for the target system with limited number of samples.After implementing transient stability prediction,an emergency control based on a deep learning classification model is proposed to achieve intelligent control of power angle stability after large disturbances.The proposed method can reflect the dynamic changes of transient stability characteristics in time-varying power systems and has speed and accuracy advantages in the transitional phase of power system changes.(4)In the construction of new power system,more and more network devices are connected to the cyber physical power system.The dispatch of the power grid involves not only the risks of the physical power grid,but also the potential cyber security risks.For the above three typical security dispatch and control scenarios,this thesis researches the implementation of false data injection attacks in the mentioned scenarios and analyzes the impact on power grid control.Considering the scarcity of malicious data in the power grid,an anomaly data detection method based on a one-class neural network is proposed,which solves the problem of anomalous detection under the condition of imbalanced dataset.