A lightweight distribution cabinet service life loss prediction optimization method

Abstract

The invention discloses a method for optimizing the prediction of the service life loss of a lightweight distribution cabinet, which belongs to the technical field of loss prediction, and comprises the following steps: step one, collecting and recording the working environment data of the lightweight distribution cabinet on a large scale, including environmental parameters and operating parameters; step two, processing abnormal values ​​and missing values ​​in the working environment data, and scaling the key features in the data to a unified scale; dividing the processed data into a training set and a test set; step three, using the data of the training set to train an enhanced LSTM prediction model, wherein the enhanced LSTM prediction model comprises: optimizing the hyperparameter setting of a standard LSTM model by using an improved guided learning optimization algorithm, and using the hyperparameters in the LSTM prediction model to obtain the enhanced LSTM prediction model; and step four, inputting the test set of the working environment data of the lightweight distribution cabinet into the trained enhanced LSTM prediction model to realize the prediction of the service life loss of the lightweight distribution cabinet.

Description

Technical Field

[0001] The invention belongs to the technical field of loss prediction and relates to a method for optimizing the loss prediction of the service life of a lightweight power distribution cabinet. Background Art

[0002] As an important component of the power system, the stable operation of lightweight distribution cabinets is crucial to the reliability and stability of the entire power system. Through life loss prediction, potential faults can be discovered in time, and repairs or replacements can be carried out in advance, thereby avoiding the impact of sudden faults on the power system.

[0003] The service life loss of distribution cabinets is a typical time series prediction problem. For time series prediction problems, LSTM prediction models are often used. However, the performance of LSTM models depends largely on the selection of their hyperparameters. The selection of these hyperparameters needs to be adjusted according to specific tasks and data sets, and there is no fixed optimal solution. Therefore, when using LSTM models to predict the service life loss of distribution cabinets, sufficient experiments and verifications are required to find the best hyperparameter configuration.

[0004] The Guided Learning Strategy (GLS) is a new algorithm designed to balance the global exploration and local utilization stages in the global optimization process. Inspired by the feedback and guidance mechanism in the teaching process, the GLS algorithm uses the standard deviation to calculate the dispersion of individual historical positions to infer the current needs of the algorithm, that is, whether the algorithm is biased towards global exploration or local utilization; in the current GLS algorithm, C max It is a fixed parameter, which means that the GLS algorithm cannot automatically adjust the feedback interval according to its own performance and changes in optimization conditions during the entire operation process. ,This lack of adaptability in the design limits the performance of the GLS algorithm in ,optimizing the LSTM prediction model to realize the ,distribution cabinet service loss prediction problem. Summary of the invention

[0005] In view of the problems existing in the current service life loss prediction method of lightweight distribution cabinets and the problems existing in the guided learning optimization algorithm in the optimization of LSTM prediction model, the present invention proposes a service life loss prediction optimization method for lightweight distribution cabinets to improve the accuracy and efficiency of service life loss prediction of lightweight distribution cabinets.

[0006] The technical solution adopted by the present invention to solve the technical problem is: to guide the feedback interval of the learning optimization algorithm The improvement improves the global optimization accuracy and efficiency of the guided learning optimization algorithm, uses the improved guided learning optimization algorithm (CGLS) to optimize the hyperparameter tuning of the LSTM prediction model to obtain an enhanced LSTM prediction model, and uses the enhanced LSTM prediction model to realize the prediction of the service life loss of lightweight distribution cabinets.

[0007] Step 1: Large-scale collection and recording of working environment data of lightweight power distribution cabinets, wherein the working environment data includes: environmental parameters and operating parameters.

[0008] Step 2: Processing the abnormal values ​​and missing values ​​in the working environment data, scaling the working environment data to a uniform scale; dividing the processed working environment data into a training set and a test set.

[0009] Step 3: The training set is used to train an enhanced LSTM prediction model. The enhanced LSTM prediction model is specifically: using an improved guided learning optimization algorithm to optimize the hyperparameters of the standard LSTM model, and the hyperparameters are used in the LSTM prediction model to obtain an enhanced LSTM prediction model, wherein the hyperparameters include: the number of hidden units G, the maximum training cycle T, the initial learning rate K, and the regularization coefficient Y. The improved guided learning optimization algorithm includes: designing the feedback interval of the performance sensitivity component and the volatility adjustment component to the guided learning optimization algorithm Make improvements.

[0010] Step 4: Input the test set of the lightweight distribution cabinet working environment data into the trained enhanced LSTM prediction model to predict the service life loss of the lightweight distribution cabinet.

[0011] Specifically, the service life loss of lightweight distribution cabinets is mainly due to the use environment of the distribution cabinets. The working environment data has a total of six variable characteristics, including environmental parameters and operating parameters; among them, the environmental parameters include three variable characteristics, specifically including the temperature and humidity values ​​of the environment around the distribution cabinet and the air pressure value of the environment where the distribution cabinet is located. The operating parameters include three variable characteristics, specifically including the working current value and working voltage value during the use of the distribution cabinet and the temperature rise value of each switch inside the distribution cabinet; the working environment data of the lightweight distribution cabinet is collected and recorded on a large scale, wherein the recording method is to record according to each group of 6 different categories of variable values.

[0012] Specifically, the working environment data of the lightweight distribution cabinet is processed, wherein the working environment data processing includes processing outliers and missing values ​​in a data set composed of six variable feature data, wherein the outlier value processing includes finding outliers and processing outliers, wherein the outliers are processed by deleting the outliers, and the missing value processing includes finding missing values ​​and processing missing values, wherein the missing values ​​are processed by deleting the missing values.

[0013] Specifically, the present invention creates a new method for discovering abnormal values. The specific method is: in the working environment data of the lightweight distribution cabinet, for the first Data points ,calculate arrive The distances of the nearest n data points around are calculated based on the distances The local density, the specific mathematical model of local density is:

[0014] (1);

[0015] In formula (1), for The local density of for The n nearest lightweight distribution cabinet working environment data points around, for The distance to the yth nearest data point of the working environment of the lightweight distribution cabinet, where n is the total number of the nearest data points of the working environment of the lightweight distribution cabinet; This is the working environment data of the lightweight distribution cabinet. data points;

[0016] Among them, according to Local density calculation If the local heterogeneous parameter value is greater than 1, then It is an outlier, otherwise it is not. The mathematical model of local outlier parameter value is:

[0017] (2);

[0018] In formula (2), for The local heterogeneous parameter value of is the local density of the yth nearest lightweight distribution cabinet working environment data point.

[0019] Specifically, the present invention creates a new method for discovering missing values. The specific method is: divide the lightweight distribution cabinet working environment data set into M clusters, including , where the mth cluster It is composed of a set of distribution cabinet working environment data points. Formula (3) is used to calculate the working environment data of the lightweight distribution cabinet. The uncertainty score of each data point;

[0020] (3);

[0021] In formula (3), for The uncertainty score is , d is the dimension of the data point, μ is the mean vector of the data set, R is the transposition symbol, is the covariance matrix, which represents the correlation between data sets in different dimensions;

[0022] Data points with uncertainty scores higher than 0.5 are considered missing values.

[0023] Specifically, the key features in the working environment data are scaled to the same scale so that the key features in the data have the same impact on the model training and avoid excessive prediction deviation due to different features. The specific method is: scale the working environment data by subtracting the minimum value and dividing by the difference between the maximum and minimum values.

[0024] Specifically, the working environment data is randomly divided into a training set and a test set according to a ratio of 7.5 to 2.5 to ensure the randomness of the data distribution. The training set is used for training the enhanced LSTM prediction model, and the test set is used to test the performance of the enhanced LSTM prediction model.

[0025] Specifically, the feedback interval before improvement is Controls how often the guided learning optimization algorithm receives feedback, the current iteration The cumulative number of iterations since the last feedback satisfy Greater than When , it enters the feedback phase and calculates the feedback result According to the feedback results, the learning optimization algorithm is guided to perform global search and local development of mathematical models to update the hyperparameters of the standard prediction model LSTM for the service life loss of lightweight distribution cabinets. The feedback interval Affects the adaptability and response speed of the guided learning optimization algorithm, fixed feedback interval Unable to adapt to dynamic changes in the optimization process, fixed feedback interval It is not possible to strike the best balance between efficiency and accuracy in guiding learning optimization algorithms.

[0026] Specifically, the smaller Provides faster response, allowing the guided learning optimization algorithm to find and use the effective optimal solution search area faster, larger The adjustment frequency is reduced, which helps to conduct deeper development in a larger search space, so that the global search and local development of the guided learning optimization algorithm can achieve a better balance. Furthermore, the feedback interval is dynamically adjusted by adaptively changing the feedback interval, so that the guided learning optimization algorithm can respond to the system changes in a timely manner, thereby improving the overall effectiveness and response speed of the system.

[0027] Specifically, for the feedback interval Improvements made, improved feedback intervals Includes: Performance sensitivity component and volatility adjustment component , the mathematical model is:

[0028] (4);

[0029] In formula (4), t is the current iteration number.

[0030] Specifically, we first design the performance sensitivity component , the feedback interval is adjusted by performance improvement. If the performance is significantly improved, the feedback interval is shortened to quickly utilize favorable exploration directions. Among them, performance improvement guides the fitness change of individuals in the learning optimization algorithm population. The mathematical model is:

[0031] (5);

[0032] In formula (5), and are the minimum and maximum values ​​of the feedback interval, To adjust the parameter of the curve steepness, is the average fitness value of the population in the tth iteration, is the minimum fitness value in the population at the tth iteration, is the minimum fitness value in the population at the t-1th iteration.

[0033] Specifically, the feedback interval is adjusted according to the volatility of performance, and the feedback frequency is increased when the volatility is high to stabilize the behavior of the guided learning optimization algorithm; where performance is the fitness value of the individuals in the guided learning optimization algorithm population; the mathematical model is:

[0034] (6);

[0035] In formula (6), β is the volatility adjustment coefficient, N is the population size, is the improvement of the performance of the ith individual after t consecutive iterations, is the mean of the population performance improvement after t consecutive iterations.

[0036] Specifically, the dynamic balance between global exploration and local development in the process of LSTM prediction model hyperparameter tuning by the improved guided learning optimization algorithm needs to be adjusted according to the optimization algorithm for the prediction effect of lightweight distribution cabinet service life loss under the current number of iterations, specifically: Calculate the feedback result , the feedback effect is used to analyze the dynamic balance required for global exploration and local development of the current improved guided learning optimization algorithm, where the required conditions are the individual fitness values ​​of the improved guided learning optimization algorithm in the process of tuning the LSTM prediction model hyperparameters, where the LSTM prediction model hyperparameters are the mapping of the individual positions of the improved guided learning optimization algorithm; feedback results The mathematical model is:

[0037] (7);

[0038] In formula (7), std function is the standard deviation function, is the minimum fitness value in the population in the tth iteration, ub is the upper bound of the search space, and lb is the lower bound of the search space.

[0039] Specifically, the hyperparameter tuning of the standard LSTM model is optimized using the improved guided learning optimization algorithm. The specific steps are as follows:

[0040] S31, encoding the number of hidden units G, the maximum training cycle T, the initial learning rate K and the regularization coefficient Y of the standard LSTM prediction model into a space vector, and establishing a position mapping between the space vector and the individual position of the improved guided learning optimization algorithm, wherein the problem dimension Dim of the improved guided learning optimization algorithm is 4;

[0041] S32, initializing the parameters of the improved guided learning optimization algorithm including the maximum number of iterations Tmax, the population size N, the problem dimension Dim, the search space upper bound ub and the search space lower bound lb, the feedback threshold α, and the minimum and maximum values ​​of the feedback interval;

[0042] S33, Individual positions of the population of the guided learning optimization algorithm improved by random initialization , check the individual position using formula (8), and if it exceeds the boundary [lb, ub], set the individual position on the boundary;

[0043] (8);

[0044] S34, determine whether the current number of iterations t satisfies t=Tmax. If so, output the individual position corresponding to the global minimum fitness value and parse it into the number of hidden units G, the maximum training cycle T, the initial learning rate K, and the space vector of the regularization coefficient Y, which are recorded as the optimal hyperparameters of the standard LSTM prediction model; otherwise, execute step S35;

[0045] S35. Calculate the feedback interval after the current iteration improvement value and current iteration number The cumulative number of iterations since the last feedback ,judge Is it greater than If so, calculate the feedback result value; otherwise, execute S37;

[0046] S36, if the feedback result If it is greater than the equilibrium threshold α, the local development phase of the improved guided learning optimization algorithm is carried out, and the new position of the individual is updated with the current global best individual position as the center; otherwise, the global exploration phase of the improved guided learning optimization algorithm is carried out, and the new position of the individual is updated with the Levy flight strategy introduced with the current position of the individual as the center;

[0047] (9);

[0048] In formula (9), Individual new positions for improved guided learning optimization algorithms, is the best individual position of the improved guided learning optimization algorithm, Levy is the Levy flight function, is the i-th individual position of the guided learning optimization algorithm improved in the t-th iteration, and R is a random number between 0 and 1;

[0049] S37, using the objective function to calculate the individual position fitness value of the current iteration-improved guided learning optimization algorithm population, and record the minimum fitness value in the t-th iteration population as , and the minimum fitness value in the population at the t-1th iteration Compare and take the smaller fitness value of the two as the current best fitness value;

[0050] S38, execute t+1 for the current iteration number t, and return to step S34.

[0051] Specifically, the objective function uses the root mean square error RMSE, which measures the mean of the square root of the difference between the predicted value of the service life loss of the lightweight distribution cabinet and the true value of the service life loss of the lightweight distribution cabinet. The mathematical model of RMSE is:

[0052] (10);

[0053] In formula (10), is the true value of the service life loss of the j-th lightweight distribution cabinet, is the predicted value of the service life loss of the j-th lightweight distribution cabinet, and s is the number of working environment data of the lightweight distribution cabinet.

[0054] The beneficial effects of the present invention are as follows: by improving the feedback interval of the guided learning optimization algorithm, the present invention can improve the accuracy and efficiency of the service life loss prediction of lightweight distribution cabinets. This improvement enables the guided learning optimization algorithm to more accurately adjust the hyperparameters of the LSTM prediction model, thereby obtaining an enhanced LSTM prediction model, further improving the accuracy of the prediction results; secondly, the improved guided learning optimization algorithm dynamically adjusts the feedback interval and adaptively changes the guidance strategy according to the needs of the actual optimization process and the response of the system, which enables the guided learning optimization algorithm to more effectively balance the global exploration and local utilization stages, thereby improving the overall effectiveness and response speed of the service life loss prediction of lightweight distribution cabinets. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 A step diagram of a method for predicting and optimizing the service life loss of a lightweight power distribution cabinet according to the present invention;

[0056] Figure 2 Diagram of the steps for optimizing the hyperparameters of the standard LSTM model for the improved guided learning optimization algorithm;

[0057] Figure 3 Comparison chart of fitness values ​​for the improved guided learning optimization algorithm and the guided learning optimization algorithm to tune the LSTM prediction model;

[0058] Figure 4 The result diagram of parameter setting of the number of hidden units G, the maximum training cycle T, the initial learning rate K and the regularization coefficient Y of the lightweight distribution cabinet service life loss prediction model;

[0059] Figure 5 The prediction results of the training set for the life loss prediction of lightweight distribution cabinets;

[0060] Figure 6 The prediction results of the lightweight distribution cabinet life loss prediction test set;

[0061] Figure 7 This is a comparison chart of the prediction results of the service life loss of lightweight distribution cabinets by various models;

[0062] Figure 8 This is a comparison chart of the prediction errors of each model for the service life loss of lightweight distribution cabinets. DETAILED DESCRIPTION

[0063] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0064] The present invention provides a technical solution: to guide the feedback interval of the learning optimization algorithm Improvement: improve the global optimization accuracy and efficiency of the guided learning optimization algorithm, use the improved guided learning optimization algorithm to optimize the hyperparameters of the LSTM prediction model to obtain an enhanced LSTM prediction model, and use the enhanced LSTM prediction model to predict the service life loss of lightweight distribution cabinets. Figure 1 shown.

[0065] Step 1: Large-scale collection and recording of working environment data of lightweight power distribution cabinets, wherein the working environment data includes: environmental parameters and operating parameters.

[0066] Furthermore, the working environment data has a total of six variable characteristics, mainly including environmental parameters and operating parameters; wherein, the environmental parameters include three variable characteristics, specifically including the temperature and humidity values ​​of the environment surrounding the distribution cabinet and the air pressure value of the environment where the distribution cabinet is located; the operating parameters include three variable characteristics, specifically including the working current value and the working voltage value during the use of the distribution cabinet and the temperature rise value of each switch inside the distribution cabinet; the working environment data of the lightweight distribution cabinet is collected and recorded on a large scale, wherein the recording method is to record according to the method of each group of 6 different categories of variable characteristic values, and a total of 100 groups of data are counted, and finally the working environment data of the lightweight distribution cabinet is counted in an xlsx format table.

[0067] Step 2: Process the outliers and missing values ​​in the working environment data, and scale the key features in the working environment data to a uniform scale; divide the processed working environment data into a training set and a test set.

[0068] Furthermore, the specific method for finding abnormal values ​​is as follows: in the working environment data of the lightweight distribution cabinet, for the first Data points ,calculate arrive The distances of the nearest n data points around are calculated based on the distances The local density, the specific mathematical model of local density is:

[0069] (1);

[0070] In formula (1), for The local density of for The n nearest lightweight distribution cabinet working environment data points around, for The distance to the yth nearest data point of the working environment of the lightweight distribution cabinet, where n is the total number of the nearest data points of the working environment of the lightweight distribution cabinet; This is the working environment data of the lightweight distribution cabinet. data points;

[0071] Among them, according to Local density calculation If the local heterogeneous parameter value is greater than 1, then It is an outlier, otherwise it is not. The mathematical model of local outlier parameter value is:

[0072] (2);

[0073] In formula (2), for The local heterogeneous parameter value of is the local density of the yth nearest lightweight distribution cabinet working environment data point.

[0074] Furthermore, the specific method for finding missing values ​​is as follows: the lightweight distribution cabinet working environment data set is divided into M clusters, including , where the mth cluster It is composed of a set of distribution cabinet working environment data points. Formula (3) is used to calculate the working environment data of the lightweight distribution cabinet. The uncertainty score of each data point;

[0075] (3);

[0076] In formula (3), for The uncertainty score is , d is the dimension of the data point, μ is the mean vector of the data set, R is the transposition symbol, is the covariance matrix, which represents the correlation between data sets in different dimensions;

[0077] Data points with uncertainty scores higher than 0.5 are considered missing values.

[0078] Furthermore, the working environment data of the lightweight distribution cabinet is preprocessed, and the preprocessing includes processing outliers and missing values ​​in a data set composed of six variable feature data, wherein the outlier value processing includes finding outliers and processing outliers, wherein the outliers are processed by deleting the outliers, and the missing value processing includes finding missing values ​​and processing missing values, wherein the missing values ​​are processed by deleting the missing values.

[0079] Furthermore, the key features in the working environment data are scaled to the same scale so that the key features in the data have the same impact on the model training and avoid excessive prediction deviation due to different features. The specific method is to scale the working environment data by subtracting the minimum value and dividing by the difference between the maximum and minimum values. The mathematical model is:

[0080] ;

[0081] In the formula, is the original working environment data point, and are the maximum and minimum values ​​of the original working environment data points, is the normalized working environment data point, and its value is in the interval [0, 1].

[0082] Step 3: The training set is used to train an enhanced LSTM prediction model. The enhanced LSTM prediction model is specifically: using an improved guided learning optimization algorithm to optimize the hyperparameters of the standard LSTM model, and the hyperparameters are used in the LSTM prediction model to obtain an enhanced LSTM prediction model, wherein the hyperparameters include: the number of hidden units G, the maximum training cycle T, the initial learning rate K, and the regularization coefficient Y. The improved guided learning optimization algorithm includes: designing the feedback interval of the performance sensitivity component and the volatility adjustment component to the guided learning optimization algorithm Make improvements.

[0083] Furthermore, 75% of the working environment data is used as the training set to train the enhanced LSTM prediction model, and the following is obtained: Figure 5 The training set lightweight distribution cabinet life loss prediction result diagram is shown in the figure, and then 25% of the working environment data is used as the test set to use the trained enhanced LSTM prediction model to predict, and the following is obtained: Figure 6 The figure shows the prediction results of the service life loss of lightweight distribution cabinets in the test set.

[0084] Furthermore, the feedback interval Improvements made, improved feedback intervals Includes: Performance sensitivity component and volatility adjustment component , the mathematical model is:

[0085] (4);

[0086] In formula (4), t is the current iteration number.

[0087] Furthermore, we first design the performance sensitivity component , the feedback interval is adjusted by performance improvement. If the performance is significantly improved, the feedback interval is shortened to quickly utilize favorable exploration directions. Among them, performance improvement guides the fitness change of individuals in the learning optimization algorithm population. The mathematical model is:

[0088] (5);

[0089] In formula (5), and are the minimum and maximum values ​​of the feedback interval, To adjust the parameter of the curve steepness, is the average fitness value of the population in the tth iteration, is the minimum fitness value in the population at the tth iteration, is the minimum fitness value in the population at the t-1th iteration.

[0090] Furthermore, the feedback interval is adjusted according to the volatility of the performance. When the volatility is high, the feedback frequency is increased to stabilize the behavior of the guided learning optimization algorithm; where the performance is the fitness value of the individuals in the guided learning optimization algorithm population; the volatility adjustment component The mathematical model is:

[0091] (6);

[0092] In formula (6), β is the volatility adjustment coefficient, N is the population size, is the improvement of the performance of the ith individual after t consecutive iterations, is the mean of the population performance improvement after t consecutive iterations.

[0093] Step 4: Input the test set of the lightweight distribution cabinet working environment data into the trained enhanced LSTM prediction model to predict the service life loss of the lightweight distribution cabinet.

[0094] Furthermore, the working environment data are randomly divided into training set and test set according to the ratio of 7.5 to 2.5 to ensure the randomness of data distribution. The training set is used for training the enhanced LSTM prediction model, and the test set is used for predicting the service life loss of lightweight distribution cabinets using the enhanced LSTM prediction model.

[0095] Furthermore, the feedback result is calculated , the feedback effect is used to analyze the dynamic balance required for global exploration and local development of the current improved guided learning optimization algorithm, where the required conditions are the individual fitness values ​​of the improved guided learning optimization algorithm in the process of tuning the LSTM prediction model hyperparameters, where the LSTM prediction model hyperparameters are the mapping of the individual positions of the improved guided learning optimization algorithm; feedback results The mathematical model is:

[0096] (7);

[0097] In formula (7), std function is the standard deviation function, is the minimum fitness value in the population in the tth iteration, ub is the upper bound of the search space, and lb is the lower bound of the search space.

[0098] Furthermore, if Figure 2 As shown in the figure, the hyperparameter tuning and optimization of the standard LSTM model are optimized using the improved guided learning optimization algorithm. The specific steps are:

[0099] S31, encoding the number of hidden units G, the maximum training cycle T, the initial learning rate K and the regularization coefficient Y of the standard LSTM prediction model into a space vector, and establishing a position mapping between the space vector and the individual position of the improved guided learning optimization algorithm, wherein the problem dimension Dim of the improved guided learning optimization algorithm is 4;

[0100] S32, initializing the parameters of the improved guided learning optimization algorithm including the maximum number of iterations Tmax, the population size N, the problem dimension Dim, the search space upper bound ub and the search space lower bound lb, the feedback threshold α, and the minimum and maximum values ​​of the feedback interval;

[0101] S33, Individual positions of the population of the guided learning optimization algorithm improved by random initialization , check the individual position using formula (8), and if it exceeds the boundary [lb, ub], set the individual position on the boundary;

[0102] (8);

[0103] S34, determine whether the current number of iterations t satisfies t=Tmax. If so, output the individual position corresponding to the global minimum fitness value and parse it into the number of hidden units G, the maximum training cycle T, the initial learning rate K, and the space vector of the regularization coefficient Y, which are recorded as the optimal hyperparameters of the standard LSTM prediction model; otherwise, execute step S35;

[0104] S35. Calculate the feedback interval after the current iteration improvement value and current iteration number The cumulative number of iterations since the last feedback ,judge Is it greater than If so, calculate the feedback result value; otherwise, execute S37;

[0105] S36, if the feedback result If it is greater than the equilibrium threshold α, the local development phase of the improved guided learning optimization algorithm is carried out, and the new position of the individual is updated with the current global best individual position as the center; otherwise, the global exploration phase of the improved guided learning optimization algorithm is carried out, and the new position of the individual is updated with the Levy flight strategy introduced with the current position of the individual as the center;

[0106] (9);

[0107] In formula (9), Individual new positions for improved guided learning optimization algorithms, is the best individual position of the improved guided learning optimization algorithm, Levy is the Levy flight function, is the i-th individual position of the guided learning optimization algorithm improved in the t-th iteration, and R is a random number between 0 and 1;

[0108] S37, using the objective function to calculate the individual position fitness value of the current iteration-improved guided learning optimization algorithm population, and record the minimum fitness value in the t-th iteration population as , and the minimum fitness value in the population at the t-1th iteration Compare and take the smaller fitness value of the two as the current best fitness value;

[0109] S38, execute t+1 for the current iteration number t, and return to step S34.

[0110] Furthermore, the objective function uses the root mean square error RMSE, which measures the mean of the square root of the difference between the predicted value of the service life loss of the lightweight distribution cabinet and the true value of the service life loss of the lightweight distribution cabinet. The mathematical model of RMSE is:

[0111] (10);

[0112] In formula (10), is the true value of the service life loss of the j-th lightweight distribution cabinet, is the predicted value of the service life loss of the j-th lightweight distribution cabinet, and s is the number of working environment data of the lightweight distribution cabinet.

[0113] Furthermore, in order to verify the applicability of the service life loss prediction method of the lightweight distribution cabinet proposed in the present invention, the standard LSTM prediction model and the LSTM prediction model optimized by the standard guided learning optimization algorithm (GLS-LSTM) are compared with the enhanced LSTM prediction model (CGLS-LSTM) proposed in the present invention. Based on the above implementation methods and steps, the standard LSTM prediction model, the GLS-LSTM prediction model and the CGLS-LSTM prediction model are written in MATLAB to predict the working environment data of the lightweight distribution cabinet. The specific steps of the standard LSTM prediction model are as follows: first, the xlsx format data of the working environment data of the lightweight distribution cabinet is imported into MATLAB, recorded as the prediction sample, and the number of prediction samples is calculated. Secondly, the missing values ​​and outliers of the prediction samples are processed, and then the prediction sample data set is reconstructed, 75% of which is used as the training set and 25% as the test set. Finally, the prediction sample data points are normalized in [0, 1]; construct a standard LSTM prediction model, including: input layer, LSTM layer, Relu activation layer and regression layer, set hyperparameter values, including the number of hidden units G=5, the maximum training cycle T=20, the initial learning rate K=10, and the regularization coefficient Y=1, the gradient threshold is 1, and the learning rate adjustment factor is 0.2.

[0114] Furthermore, the GLS-LSTM prediction model and the CGLS-LSTM prediction model, based on the standard LSTM prediction model, respectively add the codes of the standard guided learning optimization algorithm (GLS) and the improved guided learning optimization algorithm (CGLS); set the maximum number of iterations Tmax=50, the population size N=50, the problem dimension Dim=4, the upper boundary ub=[150,80,0.05,1e-3] and the lower boundary lb=[30,50,0.001,1e-4] of the individual position, the feedback threshold α=0.5, and the minimum value of the feedback interval is 0.01 and the maximum value is 1; based on the above-mentioned standard guided learning optimization algorithm (GLS) and improved guided learning optimization algorithm (CGLS) mathematical model implementation methods, the codes of the standard guided learning optimization algorithm and the improved guided learning optimization algorithm are designed in MATLAB.

[0115] Furthermore, the MATALB code is executed in the order of the standard LSTM prediction model, the GLS-LSTM prediction model, and the CGLS-LSTM prediction model to output the number of hidden units G, the maximum training cycle T, the initial learning rate K, and the regularization coefficient Y setting result diagram, the GLS and CGLS optimized LSTM individual position fitness value comparison diagram, the lightweight distribution cabinet service life loss prediction result comparison diagram, the lightweight distribution cabinet service life loss prediction error comparison diagram, and the prediction set lightweight distribution cabinet service life loss prediction result diagram and the training set lightweight distribution cabinet service life loss prediction result diagram of the method of the present invention.

[0116] Furthermore, if Figure 3 As shown in the figure, the individual position fitness value comparison chart of GLS and CGLS optimized LSTM shows that the individual position fitness value of CGLS is smaller than that of GLS in the early stage of iteration, indicating that the mean of the square root of the difference between the predicted value of the service life loss of the lightweight distribution cabinet and the true value of the service life loss of the lightweight distribution cabinet in the early stage of iteration of CGLS optimized LSTM is smaller, indicating that GLS-LSTM is more accurate in predicting the working environment data of the lightweight distribution cabinet. After the setting is stable, the individual position fitness value of GLS is 0.0409822; the individual position fitness value of CGLS is 0.0242603, which further shows that the output hidden unit number G, the maximum training cycle T, the initial learning rate K and the regularization coefficient Y hyperparameter values ​​of the CGLS standard LSTM prediction model are more accurate. The optimal hyperparameter setting changes of LSTM are shown in the figure. Figure 4 shown.

[0117] Furthermore, if Figure 7 As shown in the figure, a comparison chart of the prediction results of the service life loss of lightweight distribution cabinets by the standard LSTM prediction model, the GLS-LSTM prediction model, and the CGLS-LSTM prediction model. Compared with the actual value of the working environment data of the lightweight distribution cabinet, the standard LSTM prediction model has a large volatility, and the prediction result is relatively rough, showing more peaks and valleys; the GLS-LSTM prediction model has a relatively small volatility, but there are still some sharp peaks; the CGLS-LSTM prediction model shows the smoothest curve with the smallest volatility, indicating that the method of the present invention performs best in terms of prediction accuracy. Compared with the other two models, the CGLS-LSTM model has higher stability and accuracy in predicting the service life loss of lightweight distribution cabinets.

[0118] Furthermore, if Figure 8 As shown, a comparison chart of the prediction errors of the service life loss of lightweight distribution cabinets by the standard LSTM prediction model, the GLS-LSTM prediction model, and the CGLS-LSTM prediction model is shown. It can be clearly found in the figure that among the three model methods, the CGLS-LSTM prediction model proposed in the present invention has the smallest prediction error for the service life loss of lightweight distribution cabinets, which is within ±0.15; followed by the GLS-LSTM prediction model, which has a prediction error of the service life loss of lightweight distribution cabinets within ±0.25; the standard LSTM prediction model with the largest error predicts the service life loss of lightweight distribution cabinets, with an error within ±0.4.5.

Claims

1. A lightweight distribution cabinet service life loss prediction optimization method, characterized in that: The enhanced LSTM prediction model is used to predict the service life loss of lightweight distribution cabinets. The specific steps are as follows: Step 1: Large-scale collection and recording of working environment data of lightweight power distribution cabinets, wherein the working environment data includes: environmental parameters and operating parameters; Step 2: Processing abnormal values ​​and missing values ​​in the working environment data, scaling the working environment data to a uniform scale; dividing the processed working environment data into a training set and a test set; Step 3: The training set is used to train an enhanced LSTM prediction model. The enhanced LSTM prediction model is specifically: using an improved guided learning optimization algorithm to optimize the hyperparameters of the standard LSTM model, and the hyperparameters are used in the LSTM prediction model to obtain an enhanced LSTM prediction model, wherein the hyperparameters include: the number of hidden units G, the maximum training cycle T, the initial learning rate K, and the regularization coefficient Y. The improved guided learning optimization algorithm includes: designing the feedback interval of the performance sensitivity component and the volatility adjustment component to the guided learning optimization algorithm Make improvements; the specific steps are: S31, encoding the number of hidden units G, the maximum training cycle T, the initial learning rate K and the regularization coefficient Y of the standard LSTM prediction model into a space vector, and establishing a position mapping between the space vector and the individual position of the improved guided learning optimization algorithm, wherein the problem dimension Dim of the improved guided learning optimization algorithm is 4; S32, initializing the parameters of the improved guided learning optimization algorithm including the maximum number of iterations Tmax, the population size N, the problem dimension Dim, the search space upper bound ub and the search space lower bound lb, the feedback threshold α, and the minimum and maximum values ​​of the feedback interval; S33, Individual positions of the population of the guided learning optimization algorithm improved by random initialization , check the individual position using formula (8), and if it exceeds the boundary [lb, ub], set the individual position on the boundary; （8）； S34, determine whether the current number of iterations t satisfies t=Tmax. If so, output the individual position corresponding to the global minimum fitness value and parse it into the number of hidden units G, the maximum training cycle T, the initial learning rate K, and the space vector of the regularization coefficient Y, which are recorded as the optimal hyperparameters of the standard LSTM prediction model; otherwise, execute step S35; S35. Calculate the feedback interval after the current iteration improvement value and current iteration number The cumulative number of iterations since the last feedback ,judge Is it greater than If so, calculate the feedback result value; otherwise execute S37; feedback result The mathematical model is: （7）； In formula (7), std function is the standard deviation function, is the minimum fitness value in the population at the tth iteration, ub is the upper bound of the search space, and lb is the lower bound of the search space; S36, if the feedback result If it is greater than the equilibrium threshold α, the local development phase of the improved guided learning optimization algorithm is carried out, and the new position of the individual is updated with the current global best individual position as the center; otherwise, the global exploration phase of the improved guided learning optimization algorithm is carried out, and the new position of the individual is updated with the Levy flight strategy introduced with the current position of the individual as the center; （9）； In formula (9), Individual new positions for improved guided learning optimization algorithms, is the best individual position of the improved guided learning optimization algorithm, Levy is the Levy flight function, is the i-th individual position of the guided learning optimization algorithm improved in the t-th iteration, and R is a random number between 0 and 1; S37, using the objective function to calculate the individual position fitness value of the current iteration-improved guided learning optimization algorithm population, and record the minimum fitness value in the t-th iteration population as , and the minimum fitness value in the population at the t-1th iteration Compare the two and take the smaller fitness value as the current optimal fitness value; the objective function RMSE mathematical model is: （10）； In formula (10), is the true value of the service life loss of the j-th lightweight distribution cabinet, is the predicted value of the service life loss of the jth lightweight distribution cabinet, and s is the number of working environment data of the lightweight distribution cabinet; S38, the current iteration number t is executed t+1, and then the process returns to step S34; Step 4: Input the test set of the lightweight distribution cabinet working environment data into the trained enhanced LSTM prediction model to predict the service life loss of the lightweight distribution cabinet.

2. A lightweight power distribution cabinet service life loss prediction and optimization method according to claim 1, characterized in that: In the step one, the working environment data has a total of six variable characteristics, including environmental parameters and operating parameters; wherein the environmental parameters include three variable characteristics, specifically including the temperature and humidity values ​​of the environment around the distribution cabinet and the air pressure value of the environment where the distribution cabinet is located; the operating parameters include three variable characteristics, specifically including the working current value and the working voltage value during the use of the distribution cabinet and the temperature rise value of each switch inside the distribution cabinet; the working environment data is collected and recorded on a large scale, wherein the recording method is to record in the form of 6 different categories of variable values ​​per group.

3. A lightweight power distribution cabinet service life loss prediction and optimization method according to claim 2, characterized in that: The step 2 is to process the abnormal values ​​and missing values ​​in the working environment data, wherein when processing the abnormal values, the specific method for discovering the abnormal values ​​is: in the working environment data of the lightweight distribution cabinet, for the first Data points ,calculate arrive The distances of the nearest n lightweight distribution cabinet working environment data points are calculated based on the distances The local density, the specific mathematical model of local density is: （1）； In formula (1), for The local density of for The n nearest lightweight distribution cabinet working environment data points around, for The distance to the yth nearest data point of the working environment of the lightweight distribution cabinet, where n is the total number of the nearest data points of the working environment of the lightweight distribution cabinet; This is the working environment data of lightweight distribution cabinet. data points; Among them, according to Local density calculation If the local heterogeneous parameter value is greater than 1, then is an outlier, otherwise it is not. The mathematical model of local outlier parameter value is: （2）； In formula (2), for The local heterogeneous parameter value of is the local density of the yth nearest lightweight distribution cabinet working environment data point.

4. A lightweight power distribution cabinet service life loss prediction and optimization method according to claim 3, characterized in that: In the missing value processing of step 2, the specific method of finding the missing value during the missing value processing is as follows: the lightweight distribution cabinet working environment data set is divided into M clusters, including , where the mth cluster It is composed of a set of distribution cabinet working environment data points. Formula (3) is used to calculate the working environment data of the lightweight distribution cabinet. The uncertainty score of each data point; （3）； In formula (3), for The uncertainty score is , d is the dimension of the data point, μ is the mean vector of the data set, R is the transposition symbol, is the covariance matrix, which represents the correlation between data sets in different dimensions; Data points with uncertainty scores higher than 0.5 are considered missing values.

5. A lightweight power distribution cabinet service life loss prediction and optimization method according to claim 1, characterized in that: The improved guided learning optimization algorithm is specifically: Improvements made, improved feedback intervals Includes: Performance sensitivity component and volatility adjustment component , the mathematical model is: （4）； In formula (4), t is the current iteration number; Establishing Performance Sensitivity Components , the feedback interval is adjusted by performance improvement. If the performance is significantly improved, the feedback interval is shortened to quickly utilize favorable exploration directions. Among them, performance improvement guides the fitness change of individuals in the learning optimization algorithm population. The mathematical model is: （5）； In formula (5), and are the minimum and maximum values ​​of the feedback interval, To adjust the parameter of the curve steepness, is the average fitness value of the population in the tth iteration, is the minimum fitness value in the population at the tth iteration, is the minimum fitness value in the population at the t-1th iteration; the feedback interval is adjusted according to the volatility of the performance, and the feedback frequency is increased when the volatility is high to stabilize the behavior of the guided learning optimization algorithm; where the performance is the fitness value of the individual in the guided learning optimization algorithm population; the mathematical model is: （6）； In formula (6), β is the volatility adjustment coefficient, N is the population size, is the improvement of the performance of the ith individual after t consecutive iterations, is the mean of the population performance improvement after t consecutive iterations.

Images