A method and system for internal temperature inversion of a cable T-joint

Abstract

This invention discloses a method and system for inverting the internal temperature of a cable T-joint, relating to the field of power equipment condition monitoring technology. The method includes: establishing a three-dimensional electro-thermal coupled finite element model of the cable T-joint to obtain internal hot spot temperature and external surface candidate temperature measurement point data; using a nearest neighbor component analysis algorithm to screen external surface candidate temperature measurement points to obtain the optimal combination of temperature measurement points with the strongest correlation to the internal hot spot temperature; constructing a sample set under different operating conditions; constructing and training an improved particle swarm optimization algorithm-backpropagation neural network internal hot spot temperature inversion model, which includes an opposition learning strategy, a dynamic inertial weight strategy based on a cosine function, and a triggered population reset strategy; and using the trained model to invert the internal hot spot temperature. This invention can achieve objective optimization of temperature measurement point selection and high-precision inversion of internal hot spot temperature, and is applicable to the field of temperature monitoring of cable T-joints in ring main units.

Description

Technical Field

[0001] This invention relates to the technical field of power equipment condition monitoring and temperature inversion, and in particular to the technical field of non-invasive indirect measurement of hot spot temperature inside the T-joint of a ring main unit cable. Background Technology

[0002] Currently, monitoring the internal hotspot temperature of cable T-joints in ring main units is often difficult due to the challenge of directly measuring the location of the hotspot using sensors. Existing technologies mostly employ indirect temperature measurement methods, which involve placing several temperature measurement points on the outer surface of the joint, collecting surface temperature data, and establishing a mapping relationship between the surface temperature and the internal hotspot temperature to infer the internal hotspot temperature. Among existing technologies, one approach is the thermal circuit method based on the IEC 60287 standard, introducing concepts such as heat capacity and thermal resistance, listing corresponding nodal equations, and establishing a corresponding thermal circuit model to infer the internal conductor temperature from the surface temperature. Another approach combines optimization algorithms, neural networks, and other data-driven methods to estimate the internal hotspot temperature.

[0003] While the above solutions can achieve non-invasive temperature monitoring to some extent, they still have the following shortcomings: 1. The selection of temperature measurement points lacks objective optimization basis. In existing schemes, the selection of temperature measurement points on the outer surface usually relies on empirical point layout or limited trial and error. It is difficult to screen out the most sensitive and effective combination of measurement points for internal hot spot temperature from multiple candidate measurement points, which can easily lead to redundant input features or insufficient key information, thereby affecting the accuracy of subsequent temperature inversion. 2. Existing inversion models are insufficiently adaptable to T-joints. Cable T-joints have complex internal structures, with significant heating characteristics in the crimping area, and a strong nonlinear relationship exists between surface temperature and internal hot spot temperature. Existing conventional inversion models often suffer from slow convergence, susceptibility to local optima, and insufficient prediction accuracy and stability when dealing with such complex mapping relationships, making it difficult to meet the demand for high-precision inversion of the internal hot spot temperature of T-joints.

[0004] To address the aforementioned issues, this invention introduces Neighborhood Component Analysis (NCA) to evaluate the feature weights of candidate surface temperature measurement points, objectively and accurately selecting the two optimal measurement points with the strongest correlation to internal hotspots. This fundamentally solves the problem of feature redundancy or missing key information caused by subjective point selection. Secondly, a multi-strategy improved Particle Swarm Optimization (IPSO)-Back Propagation (BP) inversion model is constructed. Through three mechanisms—Opposition-Based Learning (OBL), cosine dynamic inertia weights, and Fitness-Monitoring-Driven Population Reset Mechanism (FM-PRM)—this effectively overcomes the bottlenecks of conventional models that easily fall into local optima, have insufficient prediction accuracy and stability when dealing with complex nonlinear thermal mapping of T-joints. Even under complex dynamic load conditions and interference from hardware measurement noise, it can still maintain high inversion accuracy. Summary of the Invention

[0005] The purpose of this invention is to solve the problems in the existing technology of lacking objective basis for the selection of temperature measurement points, easy redundancy of input features, and insufficient adaptability of existing inversion models to the complex nonlinear temperature mapping relationship of T-joints, as well as low inversion accuracy and stability. This invention proposes a method and system for inverting the internal temperature of cable T-joints, which can achieve objective optimization of temperature measurement point selection and high-precision inversion of internal hot spot temperature.

[0006] To achieve the above objectives, this invention proposes a method for inverting the internal temperature of a cable T-joint, comprising: S1. Establish a three-dimensional electro-thermal coupled finite element model of the cable T-joint to obtain internal hot spot temperature and external surface candidate temperature measurement point temperature data; S2. The nearest neighbor component analysis algorithm is used to screen the candidate temperature measurement points on the outer surface to obtain the optimal combination of temperature measurement points that has the strongest correlation with the internal hot spot temperature. S3. Construct sample sets under different operating conditions; S4. Construct and train an improved particle swarm optimization algorithm-backpropagation neural network internal hotspot temperature inversion model. The improved particle swarm optimization algorithm includes an opposition learning strategy, a dynamic inertial weight strategy based on the cosine function, and a triggered population reset strategy. S5. The internal hot spot temperature of the cable T-joint is inverted using the improved particle swarm optimization algorithm-backpropagation neural network model that has been trained.

[0007] Preferably, step S2 specifically includes: S21, arranging multiple candidate temperature measurement points on the outer surface of the cable T-joint, extracting the temperature values ​​of each candidate location under different operating conditions, and forming a temperature dataset of candidate temperature measurement points on the outer surface; S22, using the nearest neighbor component analysis algorithm to evaluate the feature weights of each candidate temperature measurement point; S23, selecting the two temperature measurement points with the highest weights as the optimal temperature measurement point combination, using the temperature data of the optimal temperature measurement point combination as the input feature quantity of the backpropagation neural network model, and using the internal hotspot temperature as the output label. This preferred scheme, through the feature weight evaluation mechanism of the nearest neighbor component analysis algorithm, achieves an objective quantitative evaluation of the candidate temperature measurement points, avoiding the subjectivity and blindness of manual experience in point selection, and at the same time, selecting the two temperature measurement points with the highest weights as input features effectively reduces input feature redundancy and improves the quality of input data for the subsequent inversion model.

[0008] Preferably, the candidate temperature measuring points are located near the rear plug of the cable T-joint, including a first set of measuring points equidistantly arranged along the transverse sleeve connection end, starting from the upper end of the rear plug, and a second set of measuring points equidistantly extending downward along the longitudinal cable body, starting from the inflection point inside the joint. This preferred scheme, considering the geometric characteristics of the T-joint formed by the intersection of transverse and longitudinal cylinders and the equivalence of its internal heat field in radial propagation, combined with the steady-state temperature distribution cloud map and the physical installation limitations of the right-side conductive rod being deeply embedded inside the insulating sleeve, locks the temperature measuring point layout area near the rear plug. This is both spatially closest to the internal heat-generating core and effectively avoids the transient thermal hysteresis effect caused by the thick insulation layer on the stress cone below the joint.

[0009] Preferably, in step S3, the different operating conditions include steady-state load conditions, single-step load conditions, multi-step load conditions, and equivalent actual load conditions; the sample data for each condition are divided into training and testing sets according to a stratified sampling principle. This preferred scheme, by covering multiple operating conditions such as steady-state, transient, and actual loads, ensures the diversity and comprehensiveness of the sample set, enabling the trained inversion model to adapt to various load changes that cable T-joints may encounter in actual operation, thereby improving the model's generalization ability and engineering applicability.

[0010] Preferably, the number of hidden layer neurons in the backpropagation neural network model is determined according to the formula... This was determined and confirmed through simulation testing, among which... It is the number of neurons in the input layer. It refers to the number of neurons in the output layer. It is a constant between [1, 10]; the activation function for neurons is the sigmoid function; and the mean squared error of the training set samples is used as the fitness function. This preferred scheme determines the number of hidden layer neurons by combining empirical formulas with simulation tests, taking into account both theoretical guidance and empirical optimization. At the same time, the use of the sigmoid function as the activation function and the mean squared error as the fitness function ensures the ability of the neural network model to fit the complex nonlinear temperature mapping relationship of the T-joint.

[0011] Preferably, in the opposition learning strategy, particles are randomly generated during initialization, and corresponding opposition particles are generated according to the opposition mapping formula. Particles with better fitness are selected from the original particles and opposition particles as the initial population; let the first... The search interval for the dimensional variable is Initial particles Opposite mapping vector It can be represented as: ,in, For the first The lower bound of the search interval for the dimensional variable. For the first The upper bound of the search interval for the dimensional variable. For the first The initial particle in the... The optimal scheme, by simultaneously considering both the original point and its counterpart in the initial population construction, improves the diversity and quality of the initial population from the perspectives of probability and search coverage, reduces the blindness caused by random initialization, and lowers the probability of the algorithm getting trapped in local optima.

[0012] Preferably, in the dynamic inertia weighting strategy based on the cosine function, the first... Inertia weights in the next iteration for: ,in, Let the current iteration algebra be... For the ultimate evolutionary generation, This represents the theoretical upper limit of the inertial weight. This represents the theoretical lower bound of the inertia weight. This preferred scheme constructs a nonlinear dynamic inertia weight update mechanism using a cosine function. It maintains a large inertia weight in the early stages of iteration to enhance global exploration capabilities, and gradually reduces the inertia weight in the later stages of iteration to improve local development accuracy. Compared with the traditional linear decreasing strategy, it has the advantages of a smooth change process and obvious stage characteristics, and can better coordinate the search needs of the particle swarm at different iteration stages.

[0013] Preferably, the triggered population reset strategy defines a global extreme value monitoring function. When continuous During the iteration process, the change in the global optimal fitness is always less than the preset threshold. When the algorithm enters a stagnant state and triggers a population reset, the determination condition is expressed as follows: ,in, Indicates the length of the stagnation monitoring window. This represents the threshold for changes in the global optimal fitness. Indicates the first The fitness value of the globally optimal particle is used; after a reset is triggered, particles are sorted according to their fitness, and a reset ratio coefficient is introduced. Retain those with high fitness ranking Elite particles are left untreated, while those ranked lower are treated differently. Disadvantaged particles perform a position reset operation, simultaneously resetting their individual historical best positions. This preferred scheme, by monitoring the changing trend of the global optimal fitness in real time, triggers the selective reset of some disadvantaged particles when the algorithm shows significant stagnation. It balances the preservation of existing good solutions with breaking search stagnation, effectively restoring population diversity and enhancing the ability to escape local optima. Compared to restarting the entire population, it has better stability and search efficiency.

[0014] This invention also proposes a system for retrieving the internal temperature of a cable T-joint, comprising: a data acquisition module for establishing a three-dimensional electro-thermal coupled finite element model of the cable T-joint and acquiring internal hot spot temperature and external surface candidate temperature measurement point temperature data; a temperature measurement point screening module for using a nearest neighbor component analysis algorithm to screen the external surface candidate temperature measurement points and obtain the optimal combination of temperature measurement points with the strongest correlation to the internal hot spot temperature; a sample construction module for constructing sample sets under different operating conditions; a model construction and training module for constructing and training an improved particle swarm optimization algorithm-backpropagation neural network internal hot spot temperature inversion model, wherein the improved particle swarm optimization algorithm includes an opposition learning strategy, a dynamic inertial weight strategy based on a cosine function, and a triggered population reset strategy; and a temperature inversion module for using the trained improved particle swarm optimization algorithm-backpropagation neural network model to invert the internal hot spot temperature of the cable T-joint.

[0015] Preferably, in the triggered population reset strategy, when consecutive During the iteration process, the change in the global optimal fitness is always less than the preset threshold. Population reset is triggered when the population is reset, and those with the highest fitness ranking are retained. Elite particles, for those ranked lower The inferior particles are repositioned, and their individual historical best positions are simultaneously reset. This preferred scheme, through real-time monitoring and triggered repositioning of the global optimal fitness, effectively avoids the problem of the algorithm getting stuck in local optima due to the loss of population diversity in the later stages of iteration, thus improving the training stability and prediction accuracy of the inversion model.

[0016] The beneficial effects of this invention are: This invention utilizes a nearest-neighbor component analysis algorithm to optimize the features of candidate temperature measurement points on the outer surface of cable T-joints. This objectively selects the optimal combination of temperature measurement points that are more strongly correlated with the internal hot spot temperature, reducing the subjectivity and input feature redundancy caused by manual point selection and improving the scientific nature of temperature measurement point layout and the input effectiveness of the temperature inversion model. Simultaneously, this invention constructs an improved particle swarm optimization algorithm-backpropagation neural network inversion model more suitable for the complex nonlinear temperature mapping relationship of cable T-joints. By introducing an adversarial learning strategy, dynamic inertial weights, and a triggered population reset mechanism, the model's convergence speed, prediction accuracy, and stability are improved. This enables high-precision indirect inversion of the internal hot spot temperature of cable T-joints under various operating conditions, demonstrating significant engineering application value.

[0017] The features and advantages of the present invention will be described in detail through embodiments and in conjunction with the accompanying drawings. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the three-dimensional geometric structure of the cable T-joint of the present invention; Figure 2 This is a cloud map showing the internal temperature field distribution of the cable T-joint of the present invention; Figure 3 This is a schematic diagram of the arrangement of candidate temperature measuring points on the outer surface of the cable T-joint of the present invention; Figure 4 This is a graph showing the result of the feature weight selection of the temperature measurement points for nearest neighbor component analysis in this invention; Figure 5 This is a comparison chart of the fitness convergence curves of different optimization algorithms of this invention; Figure 6 This is a comparison chart of the temperature tracking curves retrieved from each model under the equivalent load condition of this invention.

[0019] In the diagram: 1-bolt, 2-rear plug, 3-cable conductor, 4-insulation layer, 5-stress cone, 6-main insulation of connector, 7-outer shielding layer, 8-outer semiconductive layer, 9-conductive rod, 10-insulating sleeve, 11-crimp terminal. Detailed Implementation

[0020] (I) Finite element modeling and data acquisition See Figure 1, in the present invention, the T-shaped cable terminal matched with 10 kV ring main unit is selected as the research object, and a three-dimensional geometric model is established according to its actual structure. The model comprises main structures such as conductor, insulating layer, stress cone, main joint insulation, outer shielding layer, conductive rod, insulating sleeve and crimping terminal. Since the actual structure of the T-shaped cable terminal is relatively complex, and in consideration of the computing power of the computer and the convergence of the simulation model, the present invention appropriately equivalently simplifies the structure of the T-shaped cable terminal: the main function of the stress cone is to disperse the electric field stress concentration at the stripping position of the outer semiconducting layer of the cable, which is simplified in this model; since the inner semiconducting layer has a relatively thin thickness and little influence on the numerical calculation of the temperature field, the inner semiconducting layer is equivalently incorporated into the insulating layer, so as to shorten the solving time on the premise of ensuring the calculation accuracy; on the premise of ensuring that the simulation results are not affected, the model is established as regular as possible.

[0021] Corresponding material parameters are set according to the material properties of each component of the T-shaped cable terminal, the main material parameters are shown in Table 1, and boundary conditions, initial conditions and load conditions are set in combination with actual operating conditions, so as to establish an electro-thermal coupling finite element model of the T-shaped cable terminal.

[0022] Table 1 Main material simulation parameters

[0023] Based on the established finite element model, temperature field simulation is performed on the T-shaped cable terminal to obtain the hot spot temperature in the internal crimping area and the temperature distribution at each position of the outer surface under different working conditions. Refer to Figure 2 , it can be concluded from the cloud chart of the internal temperature distribution of the T-shaped joint that: the temperature difference of each structure in the internal area of the joint is obvious. Due to the existence of contact resistance, the temperature at the crimping terminal of the cable is the highest, and the temperature of the joint generally shows a downward trend along the radial direction, but the temperature field distribution of different radial sections varies greatly; the temperature change trend is first high and then low along the vertical upward direction, the temperature at the crimping terminal is the highest, and the maximum temperature difference is about 2.5°C; affected by the contact connection between the conductive rod and the crimping terminal, the temperature change trend is first high and then low along the horizontal right direction, but the change range is small; at the section near the cable body, the radial temperature gradient is small, and the temperature difference between the inside and outside is about 10°C, while at the section of the joint crimping terminal and the end of the conductive rod, the radial temperature gradient is relatively obvious, and the maximum temperature difference between the conductor and the outer surface exceeds 20°C.

[0024] (II) Optimal screening of temperature measurement points based on NCA Refer to Figure 3 , a plurality of candidate temperature measurement points are arranged on the outer surface of the T-shaped cable terminal, and by extracting the temperature values of each candidate position under different working conditions, a temperature data set of candidate temperature measurement points on the outer surface is formed. Considering the geometric feature that the T-shaped cable joint is formed by the intersection of horizontal and vertical cylinders, its internal thermal field has high equivalence in radial propagation. Combined with Figure 2Based on the steady-state temperature distribution cloud map and considering the physical installation constraint that the right-side conductive rod is deeply embedded inside the insulating sleeve, this invention focuses the feature selection area on the vicinity of the rear plug of the T-joint. This area is spatially closest to the internal heat-generating core; simultaneously, it effectively avoids the transient thermal hysteresis effect caused by the thick insulation layer on the stress cone below the joint. Based on this logic, a system is constructed as follows... Figure 3 The 20-point surface feature array shown: The first group of measuring points starts from the upper end T1 of the plug after the joint, and is evenly distributed along the transverse sleeve connection end at spatial steps of 10mm to T. 10 The second set of measuring points uses the inflection point T inside the joint as the starting point. 11 Starting from this point, extend downwards along the longitudinal cable body at equal intervals of 10mm increments to T. 20 .

[0025] The nearest neighbor component analysis (NRM) algorithm was used to evaluate the feature weights of each candidate temperature measurement point, quantifying the contribution of each candidate point to the inversion of internal hot spot temperature, and selecting the optimal combination of temperature measurement points accordingly. By changing simulation conditions such as load current, ambient temperature, and convective heat transfer coefficient, multiple new datasets of surface temperature and internal hot spot temperature of cable T-joints were generated. The feature weight results of these datasets after selection by the NRM algorithm are shown below. Figure 4 As shown.

[0026] Nearest Neighbor Component Analysis (NNCOM) is a supervised feature selection method based on distance metric learning. Unlike unsupervised methods such as Principal Component Analysis (PCA), NNCOM does not simply focus on preserving the overall variance of the data. Instead, it learns the weights of each feature to make the transformed feature space more conducive to subsequent prediction tasks. Therefore, NNCOM can highlight features that are more important to the prediction target while preserving the local neighborhood relationships of samples, thus exhibiting better interpretability.

[0027] The basic idea of ​​nearest neighbor component analysis (NNComponent Analysis) is to redefine the distance between samples by assigning different weights to each input feature. If a feature is more closely related to the prediction target, its corresponding weight will be larger; conversely, its weight will be smaller. In this way, in the new feature space, features more relevant to the target value will play a more important role in measuring sample similarity. For a given training sample set, let any two sample points in the feature space... and The weighted distance between them Defined as: (1) In the formula, The input feature dimension, For the first The weight parameters of each feature, and Representing samples respectively With sample In the The values ​​of each feature are determined by its weight. As shown in equation (1), the contribution of different features to the distance between samples is determined by their weights; the larger the weight, the more important the feature is in representing sample differences. Based on this distance metric, NCA uses a Softmax-based random nearest neighbor assignment mechanism to describe the proximity relationships between samples. For each sample... ,sample The probability of being selected as its reference nearest neighbor It can be represented as: (2) in and This refers to the sequence number of the training samples traversed during the calculation process. Representative with The base is . As shown in equation (2), the smaller the weighted distance between samples, the greater the probability of selecting the nearest neighbor. NCA continuously optimizes the feature weight parameters to make the resulting structure more conducive to the prediction of the target variable. After training, features with larger weights indicate that they contribute more to the prediction target, while features with smaller weights indicate that they have a weaker effect or are redundant.

[0028] In this invention, each input feature corresponds to the temperature value of different candidate temperature measurement points. The feature weights obtained through NCA learning can quantitatively evaluate the contribution of each temperature measurement point to the prediction of internal hotspot temperatures, thereby eliminating redundant measurement points and retaining key temperature measurement points.

[0029] Depend on Figure 4 It can be seen that the measurement points T5 and T6 have the strongest correlation with the temperature of the T-joint. The optimal temperature measurement points selected are used as the input features of the temperature inversion model to reduce the influence of irrelevant features and redundant information on model training, while improving the representativeness and effectiveness of the model input features.

[0030] (III) Sample set construction The two temperature measurement points T5 and T6 with the highest weights were selected as the input feature quantities of the inversion model, and the internal hot spot temperature T0 was used as the output label. Based on the finite element model, steady-state load conditions, single-step load conditions, multi-step load conditions, and equivalent actual load conditions were set to simulate the temperature change process of the cable T-joint under different operating conditions.

[0031] The specific settings and physical meanings of the four operating conditions, taking into account the dynamic characteristics of cable load changes, are as follows: (1) Steady-state load conditions: under steady-state parameterized scanning, the load current ranges from 200 A to 900 A, with 100 A intervals; the ambient temperature ranges from 5 ℃ to 40 ℃, with 5 ℃ intervals; the convective heat transfer coefficient is 4-12 (W / (m³)). 2·K)), 1 (W / (m 2 ·K)) represents the interval. This simulates the steady-state distribution of cable lines under different operating conditions.

[0032] (2) Single step load condition: Apply load current of 300 A-700 A, with 100 A intervals, for a total of 5 transient simulations, with a sampling time step of 3 min. Simulate sudden change from zero load to different loads to examine the transient response characteristics of each layer of material in the T-joint during heat conduction.

[0033] (3) Multi-step load condition: The sampling time step is 2 min, and the load current is applied with multiple step changes at irregular time intervals. The model's ability to track continuous dynamic thermal fluctuations is tested during periods of frequent and drastic fluctuations in daily electricity consumption under different seasons or extreme weather conditions.

[0034] (4) Actual load conditions: Apply piecewise function load current at 1-hour intervals, with a sampling time step of 2 minutes, and continue the simulation for 24 hours. Simulate the daily electricity load curve of a residential area.

[0035] Subsequently, the sample data for each working condition were divided into training and testing sets in an 8:2 ratio according to the stratified sampling principle. The collection details are shown in Table 2.

[0036] Table 2 Data Collection Status

[0037] (iv) Construction and training of IPSO-BP inversion model The two temperature measurement points T5 and T6 with the highest weights were selected as the input features of the inversion model, and the actual temperature T0 was used as the output label to establish a BP neural network temperature inversion model. However, conventional BP neural networks suffer from problems such as random initial weights and thresholds, susceptibility to local optima during training, and insufficient convergence speed and stability.

[0038] In this embodiment, an improved particle swarm optimization algorithm is used to optimize the BP neural network, and it is compared with the PSO, GWO, and SSA optimization algorithms. The population size of all algorithms is set to 50, and the maximum number of iterations is set to 150. The number of hidden layers in the back propagation neural network (BPNN) is 1. The nonlinear sigmoid activation function is selected for the neurons. The number of hidden layer neurons h is determined to be 8 based on the empirical formula (3) and after multiple sets of simulation tests. The mean squared error (MSE) of the training set samples is selected as the fitness function, as shown in formula (4).

[0039] (3) (4) For equations (3) and (4) It is the number of neurons in the input layer. It refers to the number of neurons in the output layer. It is a constant between [1, 10]. This is the number of samples in the training set; among which, The training sample number. , The predicted internal temperature derived from the model inversion. This represents the actual measured internal hotspot temperature.

[0040] The basic idea of ​​the PSO algorithm is as follows: A certain number of initial solution particles are randomly selected within the possible solution space. Then, the fitness of each solution particle is calculated according to the fitness objective function. The historical optimal solution position of each solution particle and the global optimal solution position of all solution particles are recorded. Based on the recorded solution particle positions, the current position of each solution particle is iteratively updated until a solution particle converges to a position with optimal fitness or the number of iterations reaches the upper limit. Assume that in a D-dimensional target search space, there are N particles forming a community. Each particle is represented by a D-dimensional vector: (5) in Represents particles The first dimension of the variable, Represents particles The second dimension of the variable.

[0041] No. The "flight" velocity of a particle is also a D-dimensional vector, denoted as . (6) in, Represents particles The velocity component corresponding to the first dimension parameter Represents particles The velocity component corresponding to the second dimension parameter.

[0042] No. The optimal position found by each particle is the individual extreme value, denoted as . (7) in, Represents particles The individual extreme value component corresponding to the first dimension parameter Represents particles The individual extreme value component corresponding to the second dimension parameter.

[0043] The optimal position found by the entire particle swarm is the global extremum, denoted as . (8) in, Represents particles The global extremum component corresponding to the first dimension parameter Represents particles The second dimension parameter corresponds to the global extremum component. After defining the above process, the particle's velocity and position are updated using the following equation: (9) (10) in This represents the training sample number of the particle. Indicates the number of dimensions. .For example (5) refers to the velocity component corresponding to the fourth dimension parameter of the third particle in the fifth iteration. In the above formula, and These respectively represent individual cognitive learning factors and social group learning factors. When When the value is high, the particle will tend to move closer to its historical local optimum position during the optimization trajectory; while Increasing the value will drive the particle to converge toward the globally optimal position explored by other members of the population. Variable , The random numbers are uniformly distributed in the interval [0,1], which are intended to introduce random perturbations into the optimization mechanism, thereby improving the robustness of the algorithm in escaping local optima. Defined as the inertia weighting coefficient, it measures the degree to which a particle inherits its velocity from the previous moment; a higher weighting coefficient indicates better performance. The value gives particles a stronger inertia to maintain their original flight direction, which helps to expand the search radius and explore the unknown global solution space in depth.

[0044] To address the problem of spatial distribution blind spots caused by traditional initialization methods based on pseudo-random sequences, this invention introduces an opposition learning mechanism into the improved particle swarm optimization algorithm. From the perspectives of probability and search coverage, considering both the original point and its opposition point simultaneously offers a greater chance of finding a better solution than considering only the original point. The working logic of this mechanism is that during initialization, not only are the initialization points randomly generated... Each particle is further divided into two pairs, and an opposite particle is generated for each particle according to the opposite mapping formula. The process is as follows: (11) Let the first The search interval for the dimensional variable is Initial particles Opposite mapping vector It can be represented as: (11) Finally, we obtained the following: We select 10 candidate individuals, then calculate their fitness, and retain the better ones. This can be used as the initial population. This can improve the diversity and quality of the initial population, reduce the blindness of random initialization, and reduce the probability of getting trapped in local optima without significantly increasing computational complexity.

[0045] A dynamic inertia weight mechanism is introduced into the IPSO algorithm. The inertia weight w in the particle update equation directly affects the retention of the velocity term and is an important parameter determining the balance between the global search capability and local exploitation capability of the particle swarm. Although the traditional linear decreasing inertia weight strategy is simple to implement, its change process is fixed and it is difficult to adapt to the dynamic requirements of search step size and search intensity at different iteration stages in complex nonlinear optimization problems. Therefore, this invention constructs a nonlinear dynamic inertia weight update mechanism based on the cosine function: (12) In the formula: For the first Inertia weights in the next iteration and They respectively refer to the current algebra and the limiting evolutionary algebra; in the formula and These represent the theoretical upper and lower limits of the inertial weight, respectively.

[0046] From the above formula, it can be seen that in the early stage of iteration, when When the inertia weight is small, Values ​​close to Particles can maintain a strong ability to inherit velocity, thereby expanding the search range and enhancing global exploration capabilities; as iterations proceed, The inertial weights gradually decrease according to the cosine law, causing the particle swarm to transition from broad search to local exploitation; in the later stages of iteration, the inertial weights gradually approach... This allows particles to perform a fine search near the current good solution, thereby improving the convergence accuracy near the optimal solution.

[0047] Compared with the traditional linear decreasing strategy, this cosine evolution mechanism has the advantages of smooth change process and obvious stage characteristics, which can better coordinate the search needs of particle swarm at different iteration stages and improve the convergence performance and optimization stability of the algorithm.

[0048] A triggered population reset mechanism is introduced into the improved particle swarm optimization algorithm. By monitoring the changing trend of the global optimal fitness in real time, a partial particle reset operation is triggered when the algorithm stagnates significantly, in order to restore population diversity and enhance the ability to escape local optima.

[0049] Define a global extremum monitoring function When continuous During the iteration process, the change in the global optimal fitness is always less than the preset threshold. When the algorithm is considered to have entered a stagnant state, the population reset mechanism is triggered, and the determination condition is expressed as follows: (13) in, Indicates the length of the stagnation monitoring window. This represents the threshold for changes in the global optimal fitness. Indicates the first The fitness value of the globally optimal particle. The threshold of FM-PRM. This value was determined experimentally. When the relative improvement of the optimal solution is less than 0.1% and there is no improvement for 20 consecutive generations, the algorithm is considered to have stalled and a reinitialization is triggered. This value is the result of repeated testing: a value that is too large (0.01) will trigger too frequently; a value that is too small (0.0001) will make it difficult to escape stalling in time. It performs best in the multi-step test of condition 3, and is also applicable to other conditions.

[0050] when The triggering of FM-PRM indicates that the current particle swarm has not achieved significant optimization over several generations, and the algorithm may have gotten stuck in a local optimum. After triggering reinitialization, particles are sorted according to their fitness, and a reset scaling factor is introduced. Retain those with high fitness rankings. Elite particles are left untreated, while those ranked lower are treated differently. Disadvantaged particles undergo a position reset operation. This strategy redistributes particles with weaker search capabilities while preserving information about current good solutions.

[0051] Furthermore, taking into account the individual historical best position of particles in standard PSO This will have a continuous pulling effect on the subsequent search trajectory. If only the current position is perturbed without updating the individual's historical best information, the particle may still return to its original local area under the attraction of the historical best position, making it difficult to achieve a true escape. Therefore, while resetting the position of the inferior particle, its individual historical best position is simultaneously reset, specifically as follows: (14) in, Indicates the reset particle position. This represents the updated historical best position of the individual. This indicates that a new location is randomly generated within the boundaries of the search space.

[0052] Through this mechanism, the algorithm can selectively reset some inferior particles when the population stagnates in the later stages of iteration, thereby restoring population diversity and improving its ability to escape local optima. Compared to restarting the entire population, this method balances the preservation of existing good solutions with breaking search stagnation, thus exhibiting better stability and search efficiency.

[0053] (v) Model performance verification To intuitively evaluate the optimization performance of the proposed IPSO algorithm in the temperature inversion problem, it is compared with three other optimization algorithms: PSO, GWO, and SSA. To more accurately evaluate the adaptability of each optimization algorithm to the transient thermal characteristics of the T-joint, multi-step load data under operating condition 3 is used as the training set, and the fitness convergence curves of the four algorithms under operating condition 3 are obtained as follows: Figure 5 As shown in Table 3, the evaluation indicators of the inversion performance of each model are summarized. The smaller the values ​​of MAE (Mean Absolute Error), MSE (Mean Square Error), and MAPE (Mean Absolute Percentage Error), the better the R... 2 The closer the value is to 1, the better the model. This is used to verify the superiority of the method of this invention in terms of training effect and prediction performance.

[0054] Table 3 Evaluation Indicators for Inversion Results of Each Model

[0055] Plot the inversion temperature tracking curve under equivalent load conditions as follows: Figure 6 All models effectively reflected the temperature change trend of the pressing process, with the IPSO-BP model showing the best agreement with the actual temperature curve, exhibiting smaller errors during the heating, peak, and cooling phases. This demonstrates that the IPSO optimization strategy can effectively improve the temperature inversion accuracy of the BP network under complex dynamic load conditions.

[0056] The results above show that the method of the present invention can accurately invert the internal hot spot temperature of the cable T-joint, and is superior to the conventional inversion model in terms of prediction accuracy, stability and anti-interference ability. It can also better adapt to the complex nonlinear temperature mapping relationship of the T-joint.

[0057] The above embodiments are illustrative of the present invention and are not intended to limit the present invention. Any simple modifications to the present invention are within the scope of protection of the present invention.

Claims

1. A method for inverting the internal temperature of a cable T-joint, characterized in that, include: S1. Establish a three-dimensional electro-thermal coupled finite element model of the cable T-joint to obtain internal hot spot temperature and external surface candidate temperature measurement point temperature data; S2. The nearest neighbor component analysis algorithm is used to screen the candidate temperature measurement points on the outer surface to obtain the optimal combination of temperature measurement points that has the strongest correlation with the internal hot spot temperature. S3. Construct sample sets under different operating conditions; S4. Construct and train an improved particle swarm optimization algorithm-backpropagation neural network internal hotspot temperature inversion model. The improved particle swarm optimization algorithm includes an opposition learning strategy, a dynamic inertial weight strategy based on the cosine function, and a triggered population reset strategy. S5. The internal hot spot temperature of the cable T-joint is inverted using the improved particle swarm optimization algorithm-backpropagation neural network model that has been trained.

2. The method according to claim 1, characterized in that, Step S2 specifically includes: S21. Arrange multiple candidate temperature measurement points on the outer surface of the cable T-joint, extract the temperature values ​​of each candidate position under different working conditions, and form a temperature dataset of candidate temperature measurement points on the outer surface. S22. The nearest neighbor component analysis algorithm is used to evaluate the feature weights of each candidate temperature measurement point; S23. Select the two temperature measurement points with the highest weights as the optimal temperature measurement point combination, use the temperature data of the optimal temperature measurement point combination as the input feature of the backpropagation neural network model, and use the internal hotspot temperature as the output label.

3. The method according to claim 2, characterized in that, The candidate temperature measurement points are located in the area near the rear plug of the cable T-joint, including a first set of measurement points that are equidistantly arranged along the transverse sleeve connection end starting from the upper end of the rear plug of the joint, and a second set of measurement points that are equidistantly extended downward along the longitudinal cable body starting from the inflection point inside the joint.

4. The method according to claim 1, characterized in that, In step S3, the different operating conditions include steady-state load condition, single-step load condition, multi-step load condition, and equivalent actual load condition; the sample data of each condition are divided into training set and test set according to the hierarchical sampling principle.

5. The method according to claim 1, characterized in that, The number of hidden layer neurons in the backpropagation neural network model is based on the formula... This was determined and confirmed through simulation testing, among which... It is the number of neurons in the input layer. It refers to the number of neurons in the output layer. It is a constant between [1, 10]; the neuron activation function is a sigmoid function; the mean square error of the training set samples is used as the fitness function.

6. The method according to claim 1, characterized in that, In the aforementioned opposition learning strategy, particles are randomly generated during initialization, and corresponding opposition particles are generated according to the opposition mapping formula. Particles with better fitness are selected from the original particles and opposition particles as the initial population. Let the first The search interval for the dimensional variable is Initial particles Opposite mapping vector It can be represented as: ,in, For the first The lower bound of the search interval for a dimensional variable. For the first The upper bound of the search interval for the dimensional variable. For the first The initial particle in the... The value that a dimension can take.

7. The method according to claim 1, characterized in that, In the aforementioned dynamic inertia weighting strategy based on the cosine function, the first... Inertia weights in the next iteration for: ,in, Let the current iteration algebra be... For the ultimate evolutionary generation, This represents the theoretical upper limit of the inertial weight. This represents the theoretical lower bound of the inertia weight.

8. The method according to claim 1, characterized in that, In the triggered population reset strategy, a global extreme value monitoring function is defined. When continuous During the iteration process, the change in the global optimal fitness is always less than the preset threshold. When the algorithm enters a stagnant state and triggers a population reset, the determination condition is expressed as follows: ,in, Indicates the length of the stagnation monitoring window. This represents the threshold for changes in the global optimal fitness. Indicates the first The fitness value of the globally optimal particle; After a reset is triggered, particles are sorted according to their fitness, and a reset scaling factor is introduced. Retain those with high fitness ranking Elite particles are left untreated, while those ranked lower are treated differently. The inferior particle performs a position reset operation and simultaneously resets its individual historical best position.

9. A temperature inversion system for the internal structure of a cable T-joint, characterized in that, include: The data acquisition module is used to establish a three-dimensional electro-thermal coupling finite element model of the cable T-joint and acquire internal hot spot temperature and external surface candidate temperature measurement point temperature data. The temperature measurement point screening module is used to screen the candidate temperature measurement points on the outer surface using a nearest neighbor component analysis algorithm to obtain the optimal combination of temperature measurement points that has the strongest correlation with the internal hot spot temperature. The sample construction module is used to build sample sets under different operating conditions; The model building and training module is used to build and train an improved particle swarm optimization algorithm-backpropagation neural network internal hotspot temperature inversion model. The improved particle swarm optimization algorithm includes an opposition learning strategy, a dynamic inertial weight strategy based on the cosine function, and a triggered population reset strategy. The temperature inversion module is used to invert the internal hot spot temperature of the cable T-joint using a trained improved particle swarm optimization algorithm-backpropagation neural network model.

10. The system according to claim 9, characterized in that, In the triggered population reset strategy, when consecutive During the iteration process, the change in the global optimal fitness is always less than the preset threshold. Population reset is triggered when the population is reset, and those with the highest fitness ranking are retained. Elite particles, for those ranked lower The inferior particle performs a position reset and simultaneously resets its individual historical best position.

Images