Research On Operation State Prediction And Risk Preventive Control Of Power Grids Based On Machine Learning

With the continuous growth of the proportion of new energy grid-connected and the gradual improvement of the degree of electric power electronization,the uncertainty and complexity of power system operation and control increase significantly,which poses a huge threat to the safe and stable operation of the system.Most of the existing regulation methods of power grid operation usually adopt model analysis or rely on expert experience to make decisions,which cannot well meet the requirements of variable system operating conditions.In recent years,with the great progress of science and technology such as algorithm performance,computing equipment and data accumulation,as well as the increase of national and industrial policy support,artificial intelligence technology has become a research hotspot again,and has been widely used in many industrial fields.In the long-term operation of power grid,the dispatching center has accumulated massive operation monitoring data,which lays a foundation for the application of artificial intelligence technology in the dispatching operation of power grid.As a typical artificial intelligence technology,machine learning methods such as deep learning and reinforcement learning can realize fast and accurate scheduling and operation decisions through effective mining of massive data generated by power grid operation,thus ensuring the safe and stable operation of the system.Power grid operation states prediction and risk prevention and control are the core content of system scheduling decisions.Machine learning technology is used to achieve accurate prediction,efficient evaluation,and optimization of operation states,while quickly and accurately screening out key links that can promote the spread of cascading faults for key monitoring or upgrading,and further conducting timely and effective safety correction and control of unsafe operation status of the power grid,It is of great significance to reduce the risk of cascading failures and major blackouts.This paper studies the method of power grid operation state prediction and risk prevention and control based on machine learning.The main work includes the following aspects:(1)A prediction method for future operation state based on spatio-temporal similarity mining of power grid operation sections is propoesd.Fully considering the characteristics of power grid operation time continuity and spatial distribution correlation,GraphSAGE algorithm is used to conduct deep unsupervised learning for power grid topology and its attribute information,and attribute vectors representing the spatial characteristics of running sections are extracted.The sliding time window algorithm is used to divide the spatial feature vectors corresponding to the historical running sections into multiple windows according to different periods,and the characteristic vectors of the running sections considering the spatio-temporal features are obtained,which enhances the representativeness of the system operation characteristics.The similarity of corresponding samples between different windows is calculated from the aspects of direction and distance.A group of continuous sections that are most similar to the sections in the current period is obtained,and the sections at subsequent moments of this group of historical sections are taken as the reference for the future state of the current power grid operation.Compared with the traditional historical running section similarity matching method,the error between the most similar historical section matched by the proposed method and the actual section is smaller,and the reference value is higher.(2)An evaluation and optimization method for power grid static voltage stability state based on deep learning and reinforcement learning is proposed.Firstly,the graph attention network(GAT)algorithm and long short-term memory(LSTM)algorithm for voltage stability margin prediction are introduced,and the working principle and flow of GAT-LSTM algorithm are summarized.Secondly,a static voltage stability margin prediction model based on GAT-LSTM algorithm is constructed,and the influence of new energy volatility is fully considered in the process of model training.Finally,the transmission network structure is optimized and adjusted based on Deep Q learning(DQN)algorithm to improve the static voltage stability margin.The proposed static voltage stability margin prediction method can better balance calculation speed and result accuracy than the existing methods,and can sense the future voltage stability trend.The proposed voltage stability margin enhancement method has a faster calculation speed than traditional methods,and can timely avoid dangerous deterioration of the system state.(3)A key link identification method for power grid cascading failure based on graph deep learning algorithm is proposed.Firstly,the propagation mechanism of cascading failure and the characteristics of critical lines are analyzed,and the cascading failures are simulated with the operational reliability model.Secondly,the fault probability of the line and the impact of the fault on the system are calculated,and the critical line is defined according to the risk theory.Then,the GAT algorithm is used to establish the critical line identification model.Considering the influence of calculation stability and sample balance on the identification results,the multi-attention mechanism and weighted cross entropy loss function are designed to improve and optimize the model.Finally,the operation scenario-critical line sample set is constructed to train,verify and test the GAT identification model.Compared with the traditional critical line identification method,the proposed method significantly improves the accuracy and speed of critical line identification in the new energy scenario,and can better adapt to the actual scenario of power grid operation,which has certain application prospects.(4)A static active power security correction control method for cascading failure blocking based on deep reinforcement learning is proposed.Firstly,under the condition of satisfying the static security constraints of the system,the active power safety correction control model is established with the goal of minimizing the output adjustment of adjustable components and ensuring the highest overall operation safety of the system.Secondly,the deep reinforcement learning framework for active power correction is constructed,and the reward function considering the objective and constraint,the observed state reflecting the power system operation,the regulation action that can change the system state,and the agent based on the improved TD3 algorithm are defined.Then,the historical system overload scenario considering the uncertainty of the source and load is constructed,and the agent is continuously interactive trained with the help of deep reinforcement learning model to obtain good decision effect.Finally,the online application simulation test is carried out,and the possible value of the source and load in the future is taken into account,and the optimal correction control scheme for line overload elimination is quickly obtained.The proposed method can effectively eliminate the line overload in the system and avoid crossing the limit again in a short time.Compared with the traditional method,it has obvious advantages in calculation speed,correction effect,etc.