Interpretable Neural Network Power Grid Fault Diagnosis Mode

The concept of "energy Internet" was proposed,which has brought with the larger scale of power systems and the more complex topology of power systems.At the same time,it has put forward higher requirements on the level of intelligent digitalization of power systems and energy production index.When a fault occurs,massive power systems fault data and alarm information will be collected in the power dispatching center.And it is difficult for the dispatchers to judge the fault components efficiently from the massive data.Power systems fault diagnosis based on intelligent algorithm is an important mean to assist the dispatchers in judging fault components quickly and accurately,which is conducive to the restoration of power,and is of great significance to ensure the safe and stable operation of power systems.The thesis devotes to the study of power system fault diagnosis,and the specific contents are as follows:Firstly,a divisional power system fault diagnosis model based on structure adaptive hierarchical extreme learning machine(HELM)is proposed to solve the problem that large-scale power system fault diagnosis is prone to dimension disaster.Based on the idea of divisional diagnosis,the diagnosis process consists of two parts:divisional fault diagnosis based on HELM and the overlapping line fault diagnosis based on Choquet fuzzy integral.Moreover,considering that manual configuration of divisional diagnosis sub-modules in large-scale power system will waste a lot of manpower and material resources,differential evolution algorithm is used to select the optimal value of the control parameters of the HELM,search the optimal sub-module structure of each sub-module,and realize the adaptive construction of the diagnosis sub-modules.Experiments show that the models have strong adaptability to large-scale power system,and can diagnose fault sections in complex situations.Then,the unexplained "black box" problem of power system fault diagnosis based on neural network is studied.In order to improve the explanatory ability of the NN-based fault diagnosis model,a universal transparent artificial neural network-based fault section diagnosis models for power systems is studied.The diagnosis model is oriented to the types of sections.By establishing the mapping relationship between the NN’s internal structure and the operation logic of section-protective relay-circuit breaker in the fault removal process,the model constructs the hidden layer connection of diagnosis model reasonably to obtain an interpretative diagnostic process.In addition,differential evolution algorithm is used to optimize the network parameters of the diagnosis models to improve the models’ fault tolerance rate.The simulation results show that the diagnosis processes of the models have explanatory ability,which can provide explainable diagnosis output for the dispatchers’ decision,and have strong adaptive ability to the power system topology change.Finally,as a derivative of neural network in membrane computing,the spiking neural P system(SNPS)method has strong explanatory ability for diagnosis process,but poor adaptive ability to the power system topology change.By extending the application of universal transparent artificial neural network-based fault section diagnosis models for power systems to the SNPS-based fault diagnosis method,a universal fault diagnosis model based fuzzy reasoning SNPS is studied.The simulation results show that the diagnosis models have simple structure,strong explanatory ability,and can effectively adapt to the power systems with frequent topological changes.And the diagnosis models can diagnose the fault effectively even if there are error alarm information.