Grounding Grids Fault Diagnosis Based On ELM And PNN

Ground ing gr ids are essentia l co mponents which ens ure the work ing groundingand protective ground ing in power stations and substations and pla y a vita l ro le forthe stable operation of power syste m. As bur ied underground, ground ing gr ids maysuffer corrosion which deteriorate the ir e lectrica l perfor mance parameters and affectthe safe operation of power system when the parameters get too bad. So it’s of greatsignifica nce to develop a method to dia gnose grounding gr id faults. The approachproposed in this paper takes ground ing gr ids as pure resista nce model, uses extremelearning machine (ELM) to locate ground ing gr id faults and utilizes probabilisticneural network (PNN) to identify grounding grids’ corrosion degrees..By adopting the steady-state ground ing gr id model of pure resista nce linearnetwork, this paper has proposed a method to dia gnose ground ing gr id fa ults based onradica l-based function (RBF) neura l network. The network is first trained by thetraining sa mples o f accessib le node voltages as subc ircuit breaks down, the n fed withthe accessib le node volta ges of grounding gr ids waiting to be diagnosed, andaccord ing to the outp ut, the fracture subcirc uit can be positioned. The simulationsreveal that this method is feasib le to position the fracture subc ircuit of ground inggr ids. On this basis, ELM is introduced to dia gnose ground ing gr id faults. Thediagnosis proves that this method is o f high accuracy and stability, which so lves theproble m that dia gnos is error is large whe n RBF neura l network is used for single faultdiagnosis, and not restricted by the fault location and ince ntive pos ition. Consider ingthe huge sca le of the training samp les for dua l fa ults and the inp ut error, this ELM hasbeen improved by sorting severa l most like ly fault s ubcircuits and introducing whitenoise d isturbance thro ugh observing the recognitio n rate. The diagnos is shows thatthe improved ELM can accurately position dua l breaks by only adopting s ingle fault’straining samp les instead of huge dual fa ult’s training samp les, which has greatlyimproved the diagnostic efficiency and the consistency to input error.A recognition method comb ining principa l component analys is (PCA) with PNNis presented to classify multip le fa ilure modes of grounding gr ids’ corrosio n. The PCAis tra ined with accessib le node volta ges under different fault modes, and the PNN istrained with the outputs o f the PCA, then the trained PCA is fed with accessib le nodevolta ges of ground ing gr ids waiting to be d iagnosed, and the trained PNN is fed withthe outputs of the trained PCA, fina lly the fa ilure mode of target ground ing gr id is recognized according to the output o f PNN. The results show that it’s an accurate andfast recognition method with higher recognitio n rate, fewer converge nce steps andshorter training time than back propagation (BP) ne utra l network approach to identifydifferent degrees of grounding grids’ corrosion.