Research On Prediction Method Of Power Transformer Health State Based On Situational Parameter

Transformers is essential electrical apparatus in power system,whose reliability directly impacts the stability of of power system.Assessing and diagnosing the health state and fault of transformer is an important means to ensure its safe and stable operation.However,the existing state assessment and fault diagnosis methods for transformers focus on assessing current state,and cannot accurately predict future state and fault development trends,which makes it difficult to provide timely and accurate support for the operation,maintenance and power dispatching.Therefore,studying the prediction method of transformer health state,predicting the transformer health state in advance,and grasping the occurrence time of faults can reserve enough time for formulating outage maintenance and power dispatching plans.It is of great significance to ensure the safe and stable operation of transformers and to improve the level of intelligent operation and maintenance and intelligent dispatching of power grids.The existing transformer state prediction methods all adopt the "predict-diagnose"structure,that is,the state quantity of future time is predicted first,and then obtain the corresponding operation state based on the diagnosis method.This method separates the prediction and diagnosis process,and the state prediction result comes from the combination of the optimal state quantity prediction method and the optimal state diagnosis method,which is the superposition of the two local optimal methods,rather than the global optimization of the state prediction method.Therefore,there is an error superposition problem in the state prediction process,resulting in poor accuracy and the inability to accurately and timely predict the transformer health state.In addition,in the process of state quantity prediction,as the prediction time increases,it becomes more and more difficult to mine the change law of the state quantity time series,resulting in a significant decrease in the prediction accuracy over time,and it is impossible to achieve long-term prediction of the transformer health state.Transformer high-dimensional space situational quantity for health state prediction is proposed to solve the problem of error superposition in traditional "predict-diagnose"health state prediction methods.The changing relationship between the state quantities in the time series,the association relationship between the multiple state quantities,the mapping relationship between the state quantities and the actual operating state,the relationship between the state quantities and the assessment results in the traditional guideline,and the fuzzy relationship between the assessment results and the actual operating state are extracted from four types of information,including online monitoring,test,assessment and actual operation of the transformer.The transformer situational parameter is constructed by fusing the high-dimensional matrices that characterize the abovementioned relationships.The situational parameter includes the time series change rule of the transformer state quantity and the diagnosis rule between the state quantity and the state,etc.The process of predicting the situational quantity is the process of finding the global optimum of the state quantity prediction method and the state diagnosis method.After obtaining the optimal situational quantity in the future time,the transformer health state in the future time can be directly analyzed from it,which realizes the direct prediction of the health state and overcomes the error superposition of the traditional state prediction method.The SOA(Structural Optimization Attention)prediction method for short-term prediction of transformer health state and suitable for high-dimensional situational quantity is proposed to solve the problem that existing prediction models cannot be used to predict complex variables in multi-variable or high-dimensional space.In this method,the attention prediction model with high accuracy is used as the primary learning model,and the attention learning mechanism is constructed for each relationship in the situational quantity.The prediction of transformer situational quantity is realized by sharing attention.In the SOA prediction model,the error flows simultaneously between various relational attention prediction models and the entire situational quantity attention prediction model.By constraining the state prediction process with the diagnosis rules,the unification of the state quantity prediction process and the state diagnosis process is unified,the global optimization solution of state prediction is obtained,and the accuracy of short-term prediction of transformer health state is improved.Based on the actual case in field,the accuracy of SOA prediction model and the "predict-diagnose" prediction model is compared.The results show that the accuracy of SOA can reach 96.33%when the transformer health state is predicted for 15 days,which is 14.37%higher than the "predict-diagnose" prediction model.The false positive rate and false negative rate are 1.22%and 2.45%,4.90%and 12.53%lower than the "predict-diagnose" prediction model.SOA significantly improves the accuracy of direct prediction of transformer health state and solves the error superposition problem of conventional prediction model.The RL(Reinforcement Learning)prediction method,which integrates the chaotic prediction in phase space and the SOA prediction in Euclidean space,is proposed to solve the problem that the ability of the existing prediction model to mine the nonlinear regularity continues to decline over time,resulting in long-term prediction.The problem of significantly reduced accuracy.Based on the RL framework,the method combines the strong fitting ability of the SOA prediction model to mine nonlinear regularity in Euclidean space and the advantage of the chaos prediction model to mine its own chaotic characteristics in phase space without relying on the fitting model and sample numerical characteristics.The nonlinear regularity in the Euclidean space is supplemented by the nonlinear regularity in the phase space,which improves the longterm prediction accuracy of the transformer health state.When the long-term prediction of the case transformer health state is carried out for 60 days,the accuracy of the RL can reach 91.44%,which is 11.93%higher than the SOA prediction model;the false positive rate and false negative rate are 3.67%and 4.89%respectively,which are 6.42%and 5.51%lower than the SOA prediction model,showing significant advantages in all indicators.The RL prediction model solves the problem that the accuracy of the SOA prediction model decreases significantly with the increase of the prediction length.