Machine-learning-based adaptive distance protection relay to eliminate zone-3 protection under-reach problem on statcom-compensated transmission lines

There is impending distance relay (DR) zone-3 backup protection element safety compromise in a midpoint integrated STATCOM on the utility grid system. This impending protection limitation is due to the relay under-reach effect due to the infeed reactive current injection into the grid from the midpoint integrated STATCOM device during the far-end short circuit fault at the zone-3 element protection coverage boundary. The infeed injected current led to the wrong line impedance estimation from the relay location to the faulty line section. Such compensated power grid protection actualization is a critical concern to the power system protection engineer due to the involvement of the injected reactive current from the STATCOM in the apparent impedance fault loop used seen by the relay for every fault beyond the DR midpoint location for effective short circuit fault isolation. This nuisance current contribution from the midpoint integrated STATCOM device assists in the power system voltage stability but causes a protection compromise for the backup zone-3 protection element during the far-end short circuit faults at the relay protection boundary. The estimated fault impedance value by the zone-3 elements is slightly higher than the actual pre-fault estimated threshold value under normal operating conditions. Thereby locating the apparent impedance trajectory outside the preset protection coverage as if there was no fault in the system, leading to protection safety compromise. Several conventional adaptive distance relay (ADR) and computational based intelligent modifications presented to solve the impending compromise by using faulted line voltage and current parameters for the various protection relay controller modification, optimizing synchronized measurement to block or limit the fault current penetration into the grid. The computational complexity and mathematical formulation solutions are some limitations in optimizing the relay characteristic changes with changes in the system reactive power penetration for effective fault detection and isolations. The ADR schemes also presented high computational time due to communication channel breakdown, latency, and susceptibility to the cyberattack since the communication channel is used for the trip command transmission and considering the high cost of communication medium. The earlier intelligent approach presented an offline approach using only faulty line parameters for intelligent classifier model training to detect, classify and locate faults. The model limitation is in retraining for new knowledge with changes in the power system network topology and lacks robustness. This current study proposes an intelligent data mining approach for the Machine Learning- Adaptive Distance Relay (ML-ADR) fault classification model using novel extracted 1-cycle transient voltage and current signals hidden knowledge from both healthy and faulty lines parameters. The hybrid discrete wavelet multiresolution analyses and machine learning (DWMRA-ML) algorithm is deployed to discover the hidden useful knowledge extraction from the 1-cycle short circuit transient fault signals (voltage and current) from healthy and fault lines section. These parameters are used to develop a standalone intelligently machine learning adaptive distance relay (ML-ADR) modification. The intelligent algorithm ML-ADR fault classifier model could discriminate 10 different far-end short circuit fault types from two network topology changes with and without midpoint integrated STATCOM on the Matlab/Simulink power grid system model. Other system parameter variations are 4 different fault resistances (0.001 Ω, 10 Ω, 50 Ω, 100 Ω), and two inception angles (0 oC and 30 oC). The prior result from the Matlab model of the adaptive numerical distance relay connected on midpoint integrated STATCOM power grid system indeed establish the existence of the under-reach effect for the relay zone-3 elements ing far-end short circuit fault at the coverage boundary leading to wrong impedance estimation. The BayesNet provides the best integrated MLADR fault classifier model better at a 5 % significance level than other deployed algorithms in the intelligent supervised learning model realization. The BayesNet ML-ADR classifier model performance evaluation with the highest kappa statistic value of 0.991, the lowest mean absolute error value of 0.0009, weighted average precision values of 99.2 %, ROC area coverage of 100 %, the most down trip decision time of 10 ms better than the existing 20 ms for conventional ADR. The integrated BayesNet ML-ADR fault classifier model eliminates the under-reach effect compromise on the zone-3 backup protection element for accurate fault detection, classification, and trip decision time reduction during far-end boundary faults. This model satisfied and finally met the objectives of the desired ADR.

Text FK 2021 11 IR.pdf Download (1MB)

Actions (login required)

View Item