Method and process of fault detection in power systems using machine learning feature extraction
Method and process of fault detection in power systems using feature extraction from the measurements of the voltages and currents in the power system, as well as machine learning methods. A computer processor, such as an industrial computer, is connected to various voltage and current measurement systems in various locations. This, in turn, is connected to a server computer and control system. Electrical power fault signals (features) propagating through the power system network are analyzed using various measurements and feature extraction methods. These can be used to detect, classify, and localize the fault. The present methods are particularly useful for hard-to-analyze faults such as high impedance faults and faults occurring due to inrush currents, bidirectional power flows, and islanding issues.
1 . A method of automatically detecting faults in at least one non-virtual electrical power network, said method comprising:
using a plurality of voltage and current measurement devices to obtain a plurality of voltage and current measurements at a plurality of times and locations in said non-virtual electrical power network, digitalizing and transmitting said plurality of voltage and current measurements via at least one communications interface as input to at least one monitoring computer system;
said monitoring computer system configured to use at least one artificial intelligence system to analyze said plurality of voltage and current measurements for a presence of at least one feature associated with at least one type of fault;
wherein, for at least some of said plurality of times and locations and types of said fault, said analysis includes analyzing said voltage and current measurements according to at least one sliding time window, producing sliding window raw measurements for at least some of said locations;
said faults comprising short-circuits, high-impedance faults, and at least one other type of fault chosen from the group consisting of inrush current faults, infeed currents, bidirectional power flow faults, isolated equipment failure, and cascading equipment failure;
said at least one artificial intelligence system comprising at least one machine learning or neural network trained with at least some information about layouts, components, and electrical properties of said at least one non-virtual electrical power network, as well as information about how at least one type of said fault alters at least some said plurality of voltage and current measurements; and
using said at least one artificial intelligence system to output information about any of the types, network locations, times, and severity of said fault.
2 . The method of claim 1 , further using said monitoring computer system to analyze said plurality of voltage and current measurements over at least some of said plurality of times and locations, and types of said fault by the steps of:
for at least some of said plurality of times and locations and types of said fault, said sliding window raw measurements comprise phases of voltage and currents as a function of time at said location and type of said fault;
for at least some said locations, analyzing said sliding window raw measurements using a Short-Time Fourier Transform (STFT), producing STFT frequency matrices, and using said STFT frequency matrices to determine their corresponding real component matrices and imaginary component matrices for at least some said locations and types of said faults; and
vectorizing said real component matrices and said imaginary component matrices, producing a plurality of real component vectors and imaginary components vectors for at least some said locations and types of said fault;
using at least some of said plurality of real component vectors and imaginary component vectors, and their corresponding locations, for any of training at least one said artificial intelligence system, and as input to said monitoring computer system.
3 . The method of claim 2 , wherein said electrical power network comprises a radial system topology.
4 . The method of claim 1 , further using said monitoring computer system to analyze said plurality of voltage and current measurements over at least some of said plurality of times and locations, and types of said fault by the steps of:
for at least some of said plurality of times and locations, and types of said fault, said sliding window raw measurements comprise phases of voltage and currents as a function of time at said location and type of said fault;
for at least some said locations, analyzing said sliding window raw measurements using a Kalman filter, said Kalman filter configured according to at least one state space model comprising at least one nonlinear power system model and a state space power system model;
using said Kalman filter to analyze said sliding window raw measurements by determining at least one innovation sequence for at least some said locations, and producing at least one innovation sequence vector comprising a plurality of said innovation sequences over a plurality of said locations;
and using at least some of said innovation sequence vectors, and their corresponding locations, for any of training at least one said artificial intelligence system, and as input to said monitoring computer system.
5 . The method of claim 4 , wherein said electrical power network is a meshed type network.
6 . The method of claim 1 , further pre-training said at least one artificial intelligence system by creating a computer model of said at least one non-virtual electrical power network, said computer model configured to mimic at least an electrical network topology, line types, and cable distances along said electrical network topology, and electrical equipment types that are used in said at least one non-virtual electrical power network;
said line types comprise any of main power grid lines, secondary busbar lines, feeder lines, overhead transmission lines, subtransmission lines, and underground transmission lines, and said electrical equipment types comprise any of transformers, circuit breakers, and switches;
using at least one computerized circuit simulator to simulating an impact of at least some electrical faults on at least some of said plurality of voltage and current measurements; said electrical faults chosen from the group consisting of short-circuits, high-impedance faults, inrush current faults, infeed currents, bidirectional power flow faults, isolated equipment failure, and cascading equipment failure, thus creating at least one failure simulation database;
and using said failure simulation database to train said artificial intelligence system.
7 . The method of claim 6 , wherein at least one said computerized circuit simulator uses any of MATLAB, EMTP, ATP, PSS, SIM600, Backend, and Python to simulate an impact of at least some electrical faults on at least some of said plurality of voltage and current measurements.
8 . The method of claim 6 , wherein said at least one computerized circuit simulator is configured to model an impact of at least some of said line types and at least some of said electrical equipment types, as well as at least some said electrical faults and at least some locations on at least some of said plurality of voltage and current measurements.
9 . The method of claim 1 , further training said at least one artificial intelligence system by creating a computer model of said at least one non-virtual electrical power network, said computer model configured to mimic at least an electrical network topology, line types, and cable distances along said electrical network topology, and electrical equipment types that are used in said at least one non-virtual electrical power network;
said line types comprise any of main power grid lines, secondary busbar lines, feeder lines, overhead transmission lines, subtransmission lines, and underground transmission lines, and said electrical equipment types comprise any of transformers, circuit breakers, and switches;
creating a previous experience record of said at least one non-virtual electrical power network, said previous experience record comprising a database consisting of at least some previously observed electrical faults, fault times, and fault network locations, as well as previously observed voltage and current measurements at said fault times, thus creating at least one previous experience database;
and using said computer model and said previous experience database to further train said artificial intelligence system.
10 . The method of claim 1 , wherein said output information comprises any of trip signals to protection equipment and signals to human system operators.
11 . The method of claim 1 , wherein said at least some information about said layout, components, and electrical properties of said at least one non-virtual electrical power network, as well as information about how at least one type of said fault alters at least some said plurality of voltage and current measurements comprises any of:
(a) Information encoded in said machine learning or neural network; and
(b) At least one computerized circuit simulator operating upon at least one computer model of at least a portion of said layout, components, and electrical properties of said at least one non-virtual electrical power network.
12 . The method of claim 1 , wherein said at least one machine learning or neural network comprises any of at least one large language model, convolutional neural network, generative adversarial network, other generative AI, and other trainable neural network hardware and software.
13 . The method of claim 1 , wherein said layout of said at least one non-virtual electrical power network comprises at least some underground transmission lines, and wherein said machine learning or neural network is specifically trained with a plurality of either simulated or non-virtual high-impedance faults located at various locations on said least some underground transmission lines.
14 . A method of automatically detecting faults in at least one non-virtual electrical power network, said method comprising:
using a plurality of voltage and current measurement devices to obtain a plurality of voltage and current measurements at a plurality of times and locations in said non-virtual electrical power network, digitalizing and transmitting said plurality of voltage and current measurements via at least one communications interface as input to at least one monitoring computer system;
said monitoring computer system configured to use at least one artificial intelligence system to analyze said plurality of voltage and current measurements for a presence of at least one feature associated with at least one type of fault;
wherein at least one of said at least one artificial intelligence system was previously pre-trained by creating a computer model of said at least one non-virtual electrical power network, said computer model configured to mimic at least an electrical network topology, line types, and cable distances along said electrical network topology, and electrical equipment types that are used in said at least one non-virtual electrical power network;
said faults comprising short-circuits, high-impedance faults, and at least one other type of fault chosen from the group consisting of inrush current faults, infeed currents, bidirectional power flow faults, isolated equipment failure, and cascading equipment failure;
said at least one artificial intelligence system comprising at least one machine learning or neural network trained with at least some information about layouts, components, and electrical properties of said at least one non-virtual electrical power network, as well as information about how at least one type of said fault alters at least some said plurality of voltage and current measurements; and
using said at least one artificial intelligence system to output information about any of the types, network locations, times, and severity of said fault.
15 . A method of automatically detecting faults in at least one non-virtual electrical power network, said method comprising:
using a plurality of voltage and current measurement devices to obtain a plurality of voltage and current measurements at a plurality of times and locations in said non-virtual electrical power network, digitalizing and transmitting said plurality of voltage and current measurements via at least one communications interface as input to at least one monitoring computer system;
said monitoring computer system configured to use at least one artificial intelligence system to analyze said plurality of voltage and current measurements for a presence of at least one feature associated with at least one type of fault;
said faults comprising short-circuits, high-impedance faults, and at least one other type of fault chosen from the group consisting of inrush current faults, infeed currents, bidirectional power flow faults, isolated equipment failure, and cascading equipment failure;
said at least one artificial intelligence system comprising at least one machine learning or neural network trained with at least some information about layouts, components, and electrical properties of said at least one non-virtual electrical power network, as well as information about how at least one type of said fault alters at least some said plurality of voltage and current measurements; and
using said at least one artificial intelligence system to output information about any of the types, network locations, times, and severity of said fault;
further training said at least one artificial intelligence system by creating a computer model of said at least one non-virtual electrical power network, said computer model configured to mimic at least an electrical network topology, line types, and cable distances along said electrical network topology, and electrical equipment types that are used in said at least one non-virtual electrical power network;
said line types comprise any of main power grid lines, secondary busbar lines, feeder lines, overhead transmission lines, subtransmission lines, and underground transmission lines, and said electrical equipment types comprise any of transformers, circuit breakers, and switches;
creating a previous experience record of said at least one non-virtual electrical power network, said previous experience record comprising a database consisting of at least some previously observed electrical faults, fault times, and fault network locations, as well as previously observed voltage and current measurements at said fault times, thus creating at least one previous experience database;
and using said computer model and said previous experience database to further train said artificial intelligence system.