CN-122020021-A - Power system fault tracing method and system based on differential enhancement type LSTM network

CN122020021ACN 122020021 ACN122020021 ACN 122020021ACN-122020021-A

Abstract

A power system fault tracing method and system based on a differential enhanced LSTM network comprises the steps of inputting three-phase current and differential signals of a power line in a set period into a pre-trained differential enhanced LSTM network which is introduced with a differential gate and a multi-head attention mechanism, wherein the differential enhanced LSTM network consists of an input layer, a plurality of hidden layers and a full-connection layer, all neurons comprise forgetting gates, input gates, output gates and differential gates, all the neurons calculate output gate outputs, unit state values and differential gate outputs in the neurons, fusion is carried out to obtain an intermediate signal, the intermediate signal obtains an output signal of the neuron at the moment corresponding to the current layer through the multi-head attention mechanism, and the output of the neuron corresponding to the current moment of the last layer outputs a fault tracing position at the current moment through the full-connection layer. The method remarkably enhances the characterization capability of transient anomalies and generalization of the model.

Inventors

ZHANG WUYANG
LI ZILIANG
Li Mangyan
CHAI TIANYOU
WANG YIHE
LIU YUHENG
SONG SHUQI
DONGFANG
LU YAN
CHEN XINYU
ZHU TIANYI
JIN SHIXIN

Assignees

东北大学
国网辽宁省电力有限公司电力科学研究院

Dates

Publication Date: 20260512
Application Date: 20251230

Claims (10)

1. The power system fault tracing method based on the differential enhancement type LSTM network is characterized by comprising the following steps of: inputting three-phase current and differential signals of a power line in a set period into a pre-trained differential enhanced LSTM network which is introduced into a differential gate and a multi-head attention mechanism, wherein the set period comprises a current time and a plurality of times before the current time; the differential enhanced LSTM network consists of an input layer, a plurality of hidden layers and a full-connection layer, wherein each hidden layer of each layer comprises a plurality of neurons, and each neuron of the same layer corresponds to one moment in a set period; For neurons of a first layer, the input signals are three-phase currents of a power line at corresponding moments, and for neurons of other layers, the input signals are output signals of neurons corresponding to the corresponding moments of the previous layer; for neurons of a first layer, the differential enhancement signals are differential signals at corresponding moments, for neurons of other layers, if the differential enhancement signals are neurons corresponding to earliest moments in a set period, the differential enhancement signals are output signals of neurons corresponding to the corresponding moments of a previous layer minus initial neuron output set by the previous layer, otherwise, the differential enhancement signals are output signals of neurons corresponding to the corresponding moments of the previous layer minus output signals of neurons corresponding to the previous moment of the previous layer; And fusing the output gate output, the unit state value and the differential gate output to obtain an intermediate signal, obtaining an output signal of the neuron of the current layer at the corresponding moment by the intermediate signal through a multi-head attention mechanism, and outputting the output of the neuron of the last layer at the current moment to the fault tracing position of the current moment through the full-connection layer.
2. The differential enhancement type LSTM network-based power system fault tracing method according to claim 1, wherein the method is characterized by: the differential signal specifically comprises: The differential signal is a differential signal of three-phase currents, and the differential signal at the corresponding moment of each phase current is the current at the corresponding moment of the corresponding phase minus the current at the previous moment.
3. The differential enhancement type LSTM network-based power system fault tracing method according to claim 2, wherein the method is characterized by: The input signal is output through a forgetting gate, an input gate and an output gate to obtain output gate output and unit state values, and the input signal is specifically: Splicing the input signals and the initial neuron outputs of the current layer for the neurons corresponding to the earliest moment in the set period; for neurons corresponding to other moments, splicing the input signals with the output of the neurons corresponding to the moment before the current layer; Multiplying the spliced signals with the weight matrix of the trained output gate, adding the offset matrix of the trained output gate, and inputting the added result into a sigmoid function to obtain output of the output gate; multiplying the spliced signals with the weight matrix of the trained forgetting gate, adding the bias matrix of the trained forgetting gate, and inputting the added result into a sigmoid function to obtain forgetting gate output; multiplying the spliced signals with the weight matrix of the trained input gate, adding the bias matrix of the trained input gate, and inputting the added result into a sigmoid function to obtain the output of the input gate; Multiplying the spliced signals by the weight matrix of the trained candidate state, adding the bias matrix of the trained candidate state, and inputting the added result into the hyperbolic tangent function to obtain a candidate state value; And fusing the forget gate output, the input gate output and the candidate state value to obtain the unit state value.
4. The differential enhanced LSTM network-based power system fault tracing method according to claim 3, wherein: The unit state value is obtained by fusing the forget gate output, the input gate output and the candidate state value, and specifically comprises the following steps: For a neuron corresponding to the earliest moment in a set period, calculating a Hadamard product of forgetting gate output in the neuron corresponding to the moment corresponding to the current layer and an initial unit state value of the set current layer as a first Hadamard product; for neurons corresponding to other moments, calculating a Hadamard product of a unit state value calculated by a neuron corresponding to the moment of the current layer and a neuron corresponding to the moment of the previous layer to be output by a forgetting gate in the neurons corresponding to the moment of the current layer as a first Hadamard product; And calculating the Hadamard product of the candidate state value calculated by the neuron corresponding to the moment corresponding to the current layer and output by an input gate in the neuron corresponding to the moment corresponding to the current layer as a second Hadamard product; and adding the first Hadamard product and the second Hadamard product to obtain a unit state value calculated by the neuron corresponding to the moment corresponding to the current layer.
5. The differential enhancement type LSTM network-based power system fault tracing method according to claim 2, wherein the method is characterized by: The differential enhancement signal is input into a differential gate to obtain a differential gate output, specifically: And multiplying the absolute value of the differential enhancement signal by the trained weight matrix of the differential gate, adding the trained bias matrix of the differential gate, and inputting the added result into a sigmoid function to obtain the output of the differential gate.
6. The differential enhancement type LSTM network-based power system fault tracing method according to claim 1, wherein the method is characterized by: the output gate output, the unit state value and the differential gate output are fused to obtain an intermediate signal, which is specifically: Calculating the Hadamard product of the sum of the output gate output and the differential gate output and the result of inputting the unit state value into the hyperbolic tangent function to obtain an intermediate signal; the output gate output, the differential gate output, the unit state value and the intermediate signal are all corresponding signals of neurons corresponding to the current moment of the current layer.
7. The differential enhancement type LSTM network-based power system fault tracing method according to claim 6, wherein the method is characterized by: the intermediate signal obtains the output signal of the neuron at the moment corresponding to the current layer through a multi-head attention mechanism, specifically: The method comprises the steps of receiving the input signals of the neurons at the corresponding moment of the current layer, sending the input signals into a plurality of parallel attention heads in the middle state, obtaining the output of the attention heads after the attention weighted calculation in each attention head, splicing all the attention output, multiplying the spliced attention output with the trained mapping matrix, adding the multiplied result into the middle state, and inputting the added result into a normalization function to obtain the output signals of the neurons at the corresponding moment of the current layer.
8. The power system fault tracing system based on the differential enhanced LSTM network by using the method of any one of claims 1 to 7, comprising an acquisition module, a differential enhanced LSTM network construction module and a fault tracing module, and is characterized in that: The acquisition module is used for acquiring three-phase current and differential signals of the power line in a set period, wherein the set period comprises a current time and a plurality of times before the current time; the differential enhanced LSTM network construction module is used for constructing a differential enhanced LSTM network, wherein the differential enhanced LSTM network consists of an input layer, a plurality of hidden layers and a full-connection layer, each hidden layer of each layer comprises a plurality of neurons, each neuron of the same layer corresponds to one moment in a set period respectively, and all the neurons comprise a forgetting gate, an input gate, an output gate and a differential gate; The fault tracing module is used for obtaining output gate output and unit state values through forgetting gates, input gates and output gates by input signals, obtaining differential gate output by inputting differential enhancement signals into the differential gates for neurons of a first layer, obtaining differential gate output by the differential enhancement signals for neurons of the first layer, obtaining differential signals of corresponding moments for neurons of other layers, obtaining differential enhancement signals for neurons of the first layer by an attention mechanism, subtracting the initial neuron output set by the previous layer from the output signals of the neurons corresponding to the previous layer if the differential enhancement signals are the neurons corresponding to the earliest moments in a set period, otherwise, obtaining intermediate signals by integrating the output gate output, the unit state values and the differential gate output, obtaining the output signals of the neurons corresponding to the current layer by the intermediate signals, and obtaining the output signals of the neurons corresponding to the current layer by the attention mechanism.
9. An apparatus comprising a memory, a processor and a computer program stored in the memory and executable on the processor, the processor performing the steps of using the method of any one of claims 1-7.
10. A computer readable storage medium storing a computer program which when executed by a processor uses the steps of the method of any of claims 1-7.

Description

Power system fault tracing method and system based on differential enhancement type LSTM network Technical Field The invention belongs to the technical field of industrial artificial intelligence, and particularly relates to a power system fault tracing method and system based on a differential enhanced LSTM network. Background The electric power system is a key national energy infrastructure, and the safe, stable and reliable operation of the electric power system is of great importance. However, the system inevitably suffers from disturbances of various types of faults, such as short circuits, wire breaks, and ground faults. If the faults cannot be traced and cleared timely and accurately, large-scale power failure accidents can be caused, and huge economic and social losses are caused. Therefore, the development of the efficient and accurate fault tracing technology has important significance for the research of the field of power system protection and control. In recent years, with the popularization of wide area measurement systems (Wide Area Measurement System, WAMS) and advanced metering architecture (ADVANCED METERING, information ‌, AMI), data-driven deep learning algorithms, particularly Long Short-Term Memory (LSTM), have revealed great potential in power system state sensing and fault analysis. LSTM can effectively capture the dynamic evolution rule of fault current and voltage signals due to the excellent time sequence dependency modeling capability. However, conventional data-driven approaches often treat the input data as a purely mathematical sequence that fails to adequately integrate into the physical mechanisms inherent in the power system, which limits to some extent the further improvement and interpretability of their performance. Disclosure of Invention In order to solve the defects existing in the prior art, the invention provides a power system fault tracing method and system based on a differential enhanced LSTM network. The invention adopts the following technical scheme. The first aspect of the invention provides a power system fault tracing method based on a differential enhancement type LSTM network, which comprises the following steps: inputting three-phase current and differential signals of a power line in a set period into a pre-trained differential enhanced LSTM network which is introduced into a differential gate and a multi-head attention mechanism, wherein the set period comprises a current time and a plurality of times before the current time; The differential enhanced LSTM network consists of an input layer, a plurality of hidden layers and a full-connection layer, wherein each hidden layer of each layer comprises a plurality of neurons, and each neuron of the same layer corresponds to one moment in a set period; All neurons include forget gates, input gates, output gates, and differential gates; for neurons of a first layer, the input signals are three-phase currents of a power line at corresponding moments, and for neurons of other layers, the input signals are output signals of neurons corresponding to the corresponding moments of the previous layer; for neurons of a first layer, the differential enhancement signals are differential signals at corresponding moments, for neurons of other layers, if the differential enhancement signals are neurons corresponding to earliest moments in a set period, the differential enhancement signals are output signals of neurons corresponding to the corresponding moments of a previous layer minus initial neuron output set by the previous layer, otherwise, the differential enhancement signals are output signals of neurons corresponding to the corresponding moments of the previous layer minus output signals of neurons corresponding to the previous moment of the previous layer; And fusing the output gate output, the unit state value and the differential gate output to obtain an intermediate signal, obtaining an output signal of the neuron of the current layer at the corresponding moment by the intermediate signal through a multi-head attention mechanism, and outputting the output of the neuron of the last layer at the current moment to the fault tracing position of the current moment through the full-connection layer. Preferably, the differential signal is specifically: The differential signal is a differential signal of three-phase currents, and the differential signal at the corresponding moment of each phase current is the current at the corresponding moment of the corresponding phase minus the current at the previous moment. Preferably, the input signal is output through a forgetting gate, an input gate and an output gate to obtain an output gate output and a unit state value, specifically: Splicing the input signals and the initial neuron outputs of the current layer for the neurons corresponding to the earliest moment in the set period; for neurons corresponding to other moments, splicing the input signals with the output of the