A method for predicting the residual mechanical life of a low-voltage circuit breaker of a power system

Abstract

A kind of power system low-voltage circuit breaker residual mechanical life prediction method includes: obtaining the parameter of low-voltage circuit breaker, the parameter includes current, voltage and temperature;Extract the original feature of parameter, construct the behavior sequence of low-voltage circuit breaker, using random walk strategy to process behavior sequence, obtain behavior characterization;The original feature is input into LateAttentionLSTM module, and the input into DeepFM module in behavior characterization, obtain time sequence characterization and derived behavior characterization;Time sequence characterization and derived behavior characterization are input into TACrossBlock, and the joint characteristics of time sequence and behavior are output interactively;The joint characteristics of time sequence and behavior are input into downstream structure, and the residual mechanical life of low-voltage circuit breaker is obtained;The operating parameters of the collected low-voltage circuit breaker are modeled, and a large number of redundant data are hierarchically sampled using different time granularity, and the key effective sequence information is extracted.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of deep learning, and particularly relates to a residual mechanical life prediction method for a low-voltage circuit breaker of a power system. BACKGROUND

[0002] Low-voltage circuit breakers play a vital role in power systems, used to protect equipment from overload and short-circuit faults. Therefore, monitoring and predicting the state and life of the circuit breaker is a key link to ensure the reliable operation of the power system.

[0003] Compared with the traditional experience-based life evaluation, the amount of reference data is small, and the potential impact of various operating environments and working conditions on the circuit breaker cannot be fully considered. How to accurately predict the residual life of the low-voltage circuit breaker, so that decision makers can discover potential problems in advance, thereby reducing the risk of equipment failure, and maintenance personnel can carry out targeted maintenance and repair based on accurate prediction results, reducing maintenance costs and maximizing the availability of the power system. This is an ongoing optimization problem in the field of intelligent operation and maintenance anomaly monitoring. To solve this problem, the establishment of the residual mechanical life prediction scheme for the low-voltage circuit breaker of the power system greatly improves the prediction accuracy, thereby better serving decision makers and maintainers.

[0004] If the residual mechanical life of the low-voltage circuit breaker of the power system is predicted, the most critical indicator is how to reduce the error from the true life. In order to adapt to the intelligent operation and maintenance anomaly monitoring field with irregular and aperiodic data but with the need for small error, a more efficient, resource-saving, and error-reducing residual mechanical life prediction method for the low-voltage circuit breaker of the power system needs to be proposed. SUMMARY

[0005] To solve the problems existing in the prior art, the application provides a residual mechanical life prediction method for a low-voltage circuit breaker of a power system, which comprises the following steps: obtaining parameters of the low-voltage circuit breaker, the parameters comprising current, voltage and temperature; extracting original features of the parameters, constructing a behavior sequence of the low-voltage circuit breaker, and processing the behavior sequence by using a random walk strategy to obtain a behavior representation; inputting the original features into a LateAttentionLSTM module and inputting the behavior representation into a DeepFM module to obtain a time sequence representation and a derived behavior representation; inputting the time sequence representation and the derived behavior representation into a TACrossBlock to interactively output a joint feature of the time sequence and the behavior; and inputting the joint feature of the time sequence and the behavior into a downstream structure to obtain the residual mechanical life of the low-voltage circuit breaker.

[0006] The application has the following beneficial effects:

[0007] The application models the collected operating parameters of the low-voltage circuit breaker, uses different time granularities to hierarchically sample a large amount of redundant data, and extracts key effective sequence information; at the same time, the behavior sequence is constructed to initialize the embedded mode by using random walk, and an autonomous objective function is used in the training stage to maximize the reduction of error expectation in the training stage, which more effectively simulates various unstable conditions of parameters; a double-tower model is used to process the initialized time sequence representation and behavior representation in an end-to-end manner, a self-designed delayed attention mechanism is used for the time sequence input side, so that the model can pay more attention to the sequence information of the later input, and then an LSTM is used to enhance the understanding ability of the time sequence; the forward information is cross-processed in the behavior representation input end to derive the behavior characteristics, so that it can also pay attention to the forward information at its own time step; finally, the time sequence representation and the derived behavior representation are input into the TACrossBlock mutual correlation attention important joint block, and the joint features are input into the linear layer to obtain the remaining life prediction result. In addition, the whole process adopts distributed training, can input large batches and calculate parameters at the same time, speeds up the training and reasoning speed, and greatly improves the accuracy of the model scheme designed by the application. BRIEF DESCRIPTION OF DRAWINGS

[0008] Figure 1 A flow chart of a low-voltage circuit breaker residual mechanical life prediction method based on big data of a power system according to an embodiment of the application;

[0009] Figure 2 A model structure diagram of a low-voltage circuit breaker residual mechanical life prediction method based on big data of a power system according to an embodiment of the application. DETAILED DESCRIPTION

[0010] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the application.

[0011] A low-voltage circuit breaker residual mechanical life prediction method, as Figure 1As shown, the method includes: obtaining the parameters of the low-voltage circuit breaker, which include current, voltage and temperature; extracting the original features of the parameters, constructing the behavior sequence of the low-voltage circuit breaker, and processing the behavior sequence with a random walk strategy to obtain a behavior representation; inputting the original features into the LateAttentionLSTM module, and inputting the behavior representation into the DeepFM module to obtain a timing representation and a derived behavior representation; inputting the timing representation and the derived behavior representation into TACrossBlock, and interactively outputting the joint features of the timing and behavior; inputting the joint features of the timing and behavior into the downstream structure to obtain the remaining mechanical life of the low-voltage circuit breaker.

[0012] A method for predicting the remaining mechanical life of low-voltage circuit breakers in power systems based on big data includes: obtaining operating parameters of the low-voltage circuit breaker to be predicted, constructing a time series training sample input, and obtaining a prediction result from the model. The training of the remaining mechanical life prediction model for low-voltage circuit breakers in power systems includes the following steps:

[0013] S1: Obtain the operating parameters of the low-voltage circuit breaker, such as current, voltage, and temperature, and construct a time series training sample. Furthermore, the initial input process of the training sample includes:

[0014] Input(I t ,V t ,T t ,W t …)

[0015] Where, I t ,V t ,T t ,W t They respectively represent parameter record information such as current, voltage, temperature, and humidity, including but not limited to these four parameters.

[0016] S2: Extract the original features of the training samples, construct the behavior sequence of the low-voltage circuit breaker, and use random walk to obtain its behavior representation.

[0017] Further, such as Figure 2 The data preprocessing steps shown include:

[0018] Step 1: Perform random hierarchical sampling of different granularities on the original features according to different time steps, which ranges from [1min, 1day].

[0019] Step 2: Sampling in the same time step order forms a sequence and inputs it into the random walk model to train the behavior representation.

[0020] Step 3: Formulate a specific optimization objective function based on the previous and subsequent behaviors of the sequence, and maximize the objective function during the training process.

[0021] Further, the random hierarchical sampling rules of different granularities include:

[0022]

[0023] P(i∈S)=P(t(i)=T)·P(i∈S|t(i)=T)

[0024] where t(i) represents the time granularity of the data point, S is the set of the final sampling, P(t(i)=T) represents the probability of the data point being assigned to the time granularity T, and P(i∈S|t(i)=T) represents the probability of the data point being sampled under the condition of the time granularity T.

[0025] Further, the random walk adopts a specific optimization objective function to formulate the rules as follows:

[0026]

[0027] where max(·) represents the built-in maximization target, T represents the set time granularity, c represents the context window, P(i t+j |i t ) represents the conditional probability of t+1 at the t-th time step, represents the input vector of the time step i t , and represents the output vector of the time step i t+j , V represents the total number of time steps, and bias(·) represents the random bias.

[0028] Further, the calculation rules of the random walk for maximizing the objective function in training include:

[0029]

[0030] where max(·) represents the built-in maximization target, T represents the set time granularity, c represents the context window, represents the expected probability of negative sampling, K represents the number of negative samples, represents the probability distribution of the negative samples, represents the input vector of the time step i t .

[0031] S3: The original features are input into the LateAttentionLSTM module, and the behavior representation is input into the DeepFM for feature cross, and the output is the time series representation and the derived behavior representation.

[0032] Further, the calculation rules of the LateAttention mechanism include:

[0033]

[0034] where E x denotes the feature output with location information, U denotes the total time step, a denotes the moving exponent, F U-t denotes the feature representation in the last t time steps, SoftMax(·) denotes the activation function, W query denotes the learnable query weight matrix, d model denotes the model dimension.

[0035] Further, the behavior representation is derived from the FM output, and the time sequence representation and the derived behavior representation are input into the LSTM and the DeepFM module. The calculation rules of the modules include:

[0036]

[0037] T s +A s =LSTM(Time out )+DeepFM(X+Act out )

[0038] where Time out denotes the delayed attention time sequence output, Act out denotes the derived behavior feature output, ω0 denotes the bias term, ω i denotes the weight of the i-th feature, x i denotes the value of the i-th feature, v i denotes the hidden vector of the i-th feature, T s , A s denote the time sequence representation and the derived behavior representation, and X denotes the original input feature.

[0039] S4: input the time sequence representation and the derived behavior representation into the TACrossBlock, and output the joint features of the time sequence and the behavior.

[0040] Further, the calculation rules of the TACrossBlock include:

[0041] Q T =T s ·W Q Q A =A s ·W Q

[0042] K T =A s ·W K K A =T s ·W K

[0043] V T =As • W V V A = T s • W V

[0044]

[0045] where W Q ,W K ,W V denote the learnable weight matrices, T s ,A s denote the time-series representation and the derived behavior representation, F x ,F y denote the number of features at each time step, d model denote the dimension of the model, outputs denote the joint representation of time-series and behavior.

[0046] S5: The joint representation of time-series and behavior is passed through a simple downstream structure to output the remaining mechanical life of the low-voltage circuit breaker.

[0047] Further, the simple downstream structure Output computation process includes:

[0048]

[0049] where x denotes the i-th data output, W ij denotes the weight matrix, connecting the weights of x j and b i denotes the bias term of the output.

[0050] The above examples further illustrate the purpose, technical solutions and advantages of the present application. It should be understood that the above examples are only preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made to the present application within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A method for predicting the remaining mechanical life of a low-voltage circuit breaker in a power system, characterized in that: include: Obtaining parameters of a low-voltage circuit breaker, including current, voltage, and temperature; Extract the original features of the parameters, construct the behavior sequence of the low-voltage circuit breaker, and use the random walk strategy to process the behavior sequence to obtain the behavior representation; The original features are input into the LateAttentionLSTM module, and the behavior representation is input into the DeepFM module to obtain the time series representation and the derived behavior representation. The time series representation and the derived behavior representation are input into the TACrossBlock, which interactively outputs the joint features of time series and behavior. The combined characteristics of timing and behavior are input into the downstream structure to obtain the remaining mechanical life of the low-voltage circuit breaker; The LateAttentionLSTM module and DeepFM module process the input data including: T s +A s =LSTM(Time out )+DeepFM(X+Act out ) Among them, Time out Indicates delayed attention timing output, Act out represents the derived behavioral feature output, ω0 represents the bias term, ω i represents the weight of the i-th feature, x i represents the value of the i-th feature, v i represents the latent vector of the i-th feature, T s ,A s represents the temporal representation and the derived behavior representation, and X represents the features of the original input; TACrossBlock processes timing representation and derived behavioral representation including: Q T =T s ·W Q Q A =A s ·W Q K T =A s ·W K K A =T s ·W K V T =A s ·W V V A =T s ·W V Among them, W Q ,W K ,W V represents the learnable weight matrix, T s ,A s represents temporal representation and derived behavioral representation, F x ,F y represents the number of features at each time step, d model Represents the dimension of the model, and outputs represents the joint representation of timing and behavior.

2. A method for predicting the remaining mechanical life of a low-voltage circuit breaker in a power system according to claim 1, characterized in that: Obtaining behavioral representation includes: performing random hierarchical sampling of different granularities on the original features at different time steps, where the time step range is [1 minute, 1 day]; forming a sequence according to the sampling order of the same time step, and inputting the sequence into the random walk model for training, and formulating an optimization objective function based on the behavior before and after the sequence. At the same time, the objective function is maximized during the training process to obtain the optimal behavioral representation.

3. The method for predicting the remaining mechanical life of a low-voltage circuit breaker in a power system according to claim 2, characterized in that: Random stratum sampling rules of different granularities include: P(i∈S)=P(t(i)=T)·P(i∈S|t(i)=T) Where t(i) represents the time granularity, U is the time synchronization set, T is the set time granularity, S is the final sampled set, P(t(i) = T) represents the probability that a data point is assigned to the time granularity T, and P(i∈S|t(i) = T) represents the probability that a given data point is sampled under the condition of time granularity T.

4. The method for predicting the remaining mechanical life of a low-voltage circuit breaker in a power system according to claim 2, characterized in that: The optimization objective function is: Among them, max(·) represents the built-in maximization target, T represents the time granularity of the setting, c represents the context window, and P(i t+j |i t ) represents the conditional probability of t+1 at the tth time step, represents time step i t The input vector, represents time step i t+j The output vector of , V represents the sum of time steps, and bias(·) represents the random addition of bias.

5. The method for predicting the remaining mechanical life of a low-voltage circuit breaker in a power system according to claim 2, characterized in that: The calculation rules for maximizing the objective function include: Among them, max(·) represents the built-in maximization target, T represents the set time granularity, and c represents the context window. represents the expected probability of negative sampling, K represents the number of negative sampling, represents the probability distribution of negative samples, represents time step i t The input vector.

6. The method for predicting the remaining mechanical life of a low-voltage circuit breaker in a power system according to claim 1, characterized in that: The formula for the delayed attention mechanism is: Among them, E x represents the feature output that incorporates position information, U represents the total time step, ɑ represents the movement index, and F U-t represents the feature representation at the last t time steps, SoftMax(·) represents the activation function, and W query represents the learnable query weight matrix, d model Represents the model dimension.

7. The method for predicting the remaining mechanical life of a low-voltage circuit breaker in a power system according to claim 1, characterized in that: Joint features of downstream structural computation timing and behavior include: in, Indicates the i-th data output, W ij represents the weight matrix, b i Represents the bias term of the output.

Images