Dynamic optimization method for primary and secondary fusion complete ring main unit cooling device

Abstract

The invention provides a primary and secondary fusion complete ring main unit cooling device dynamic optimization method, and belongs to the technical field of ring main unit cooling optimization based on deep learning. A dynamic monitoring data set comprising a ring main unit thermal field monitoring data set and a thermal coefficient monitoring data set is constructed, the thermal field monitoring data set is constructed through five-view infrared image splicing and high-density thermocouple calibration, and the thermal coefficient monitoring data set is constructed through combination of mechanical data such as vibration and stress and manual calibration; constructing a thermal coefficient sensing model, and evaluating a heat production coefficient and a heat dissipation coefficient of the ring main unit in real time based on force field data through three-branch feature extraction and adaptive weighted fusion; constructing a thermal field sensing model, and adopting an extractor-reconstructor architecture and a cross-layer feature fusion device to realize full-scene temperature sensing; and a cooling optimization control algorithm is designed, thermal field distribution and thermal coefficients are combined, self-adaptive adjustment is performed on seven cooling device parameters such as the fan rotating speed and the liquid cooling flow speed through multi-objective optimization, and refined cooling control is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of ring network box cooling optimization based on deep learning, and in particular relates to a dynamic optimization method for a primary and secondary fusion ring network box cooling device. Background Art

[0002] As a core component of distribution networks, integrated primary and secondary integrated ring mainframe boxes (RMBs) face long-term thermal anomalies (such as contact overheating and insulation aging) that directly impact the safe and stable operation of the distribution network. Traditional monitoring methods rely on single-mode, low-frequency detection methods such as single-point temperature measurement and manual inspections. These methods suffer from limitations such as delayed state perception, ambiguous fault location, and inefficient operation and maintenance, making them unable to meet the intelligent equipment requirements of the new power system. With the advancement of intelligent sensing technology and edge computing, the development of multimodal thermal and force field prediction and dynamic monitoring systems has become a key path to overcoming these bottlenecks. This system integrates multiple sensors, including infrared thermal imaging, vibration sensing, and stress monitoring, to construct a coupled thermal, mechanical, and electrical multiphysics model. This system enables comprehensive perception and in-depth analysis of the RMG's overall thermal field, addressing the challenge of traditional methods' inability to capture complex temperature field characteristics. On the basis of realizing automatic perception of the thermal field of the ring main box, the system further optimizes the control of the cooling equipment of the ring main box, assists in the intelligent upgrade of the distribution network, and provides key technical support for building a safe, reliable, efficient and low-carbon new power system. It has important engineering significance and industry promotion value for ensuring the stability of urban power supply and improving energy utilization efficiency.

[0003] At present, the multi-modal thermal field and force field prediction methods for the primary and secondary fusion ring network box are mainly concentrated in the following methods: (1) Traditional physical models and single-mode analysis methods: This type of method is based on basic theories such as heat transfer and mechanical vibration. It uses single-point sensors to collect single parameters such as temperature and vibration amplitude, and combines preset physical formulas and threshold logic to perform state assessment. Its core is to construct an analytical model through manually designed features. Its advantages are clear physical meaning and high computational efficiency, which are suitable for preliminary monitoring under steady-state conditions. However, when faced with complex scenarios such as multi-field coupling and nonlinear effects, single-mode data is difficult to fully characterize the equipment status, and its ability to identify thermal fields and force fields in complex scenarios is insufficient; (2) Multimodal data-driven fusion method: This method uses statistical analysis or shallow machine learning models to perform feature fusion and pattern recognition through multi-source sensor data. Its core is to convert multimodal data into a unified feature space and combine it with model analysis to achieve thermal field and force field prediction. However, this method relies too much on artificial feature engineering, has limited modeling capabilities for deep nonlinear laws, and the generalization of the model is limited by data quality and feature selection strategies; (3) Transfer learning and cross-task collaborative methods: This type of method transfers the thermal field prediction model of power equipment such as transformers and switchgear to the ring network box state prediction task, and effectively reduces the annotation cost by fine-tuning the key layer parameters to adapt to the thermal-mechanical characteristic distribution of specific equipment. However, when implementing cross-device migration, this type of method needs to solve the domain offset problem caused by differences in sensor layout and operating conditions, and the weight distribution of the multi-task loss function needs to be fine-tuned in combination with the physical characteristics of the equipment. The complexity and requirements of the model transplantation are too high. Summary of the Invention

[0004] In response to the above problems, the present invention proposes a dynamic optimization method for a primary and secondary fusion ring network box cooling device, which includes the following steps: S1, dynamically acquiring infrared perception images and mechanical characteristic data of the ring network box, wherein the infrared perception images of the ring network box are constructed as thermal field monitoring data of the ring network box as standard input of the model by splicing infrared images from five perspectives; the mechanical characteristic data of the ring network box are combined to obtain thermal coefficient monitoring data of the standard input of the model; S2: Input the thermal field monitoring data of the ring network box into the constructed and trained thermal field perception model of the ring network box, output the thermal field perception image, and calculate the temperature mean and temperature variance; input the thermal coefficient monitoring data into the constructed and trained thermal coefficient perception model of the ring network box, and output the heat generation coefficient and heat dissipation coefficient; S3, with maximizing the cooling efficiency of the ring network box, minimizing the cooling consumption of the ring network box, and maximizing the service life of the heat dissipation components as the model optimization objectives, takes the cooling equipment operating parameters during the ring network box cooling control process as the working variables, and combines the temperature mean, temperature variance, heat generation coefficient, and heat dissipation coefficient to construct a multi-objective optimization function; S4, solves the multi-objective optimization function and outputs refined cooling control parameters.

[0005] Preferably, the method for constructing the data set for training the thermal field perception model of the ring grid box includes: Infrared images of the front, back, left, right and top of the cube ring network box were collected, and the infrared images collected from five different perspectives were stitched together to form an infrared perception image of the ring network box as a whole. ; A high-density thermocouple array is evenly arranged on the surface of the ring network box, and the thermocouples are fixed to the surface of the ring network box with high-temperature resistant tape. Then, the correspondence between the physical coordinates of the thermocouples and the pixel coordinates of the infrared image is established through a checkerboard calibration plate, and the temperature value actually collected by the thermocouple is used as the true value label of the corresponding pixel point. Then, the local weighted regression difference method is used to supplement the temperature label of the pixel points that lack the true value label, and finally the calibrated thermal field perception image of the ring network box is obtained. and use it as the output data for the thermal field prediction of the ring network box.

[0006] Preferably, the method for constructing the data set for training the ring network box thermal coefficient perception model includes: Acceleration sensors are deployed at the four corners of the switch operating mechanism connecting rod of the ring network cabinet, the busbar support insulator base, the cable terminal head, the gas insulation wall shell and the cabinet base to collect the corresponding vibration frequency data. And vibration amplitude data , deploy grating strain sensors to collect stress distribution Data, deploy tension and pressure sensors to collect corresponding tension data and pressure data , deploy resistance strain gauges to collect stress and deformation data The above collected mechanical characteristic data of the ring network box are combined to obtain a set of input data in the ring network box heat generation coefficient prediction data set. ; Predict input data for each group of heat production coefficients through manual judgment Added heat generation coefficient label and heat dissipation coefficient labels The two coefficient data are combined as the output data in the ring network box heat generation coefficient prediction data set .

[0007] Preferably, the thermal field perception model of the ring network box includes an infrared image feature extractor, a thermal field perception image reconstructor and a cross-layer feature fusion device; The overall feature extraction path of the thermal field perception model of the ring network box is composed of seven infrared image feature extractors stacked in sequence and connected front to back, and the thermal field image reconstruction path of the thermal field perception model is composed of six thermal field perception image reconstructors stacked in sequence and connected front to back; the outputs of the second extractor and the third extractor are sent to the first cross-layer feature fusion device for processing, and then the output of the feature fusion device is used as a supplementary input of the second reconstructor and the third reconstructor. At the same time, the outputs of the fourth extractor and the fifth extractor are sent to the second cross-layer feature fusion device for processing, and the output of the second feature fusion device is used as a supplementary input of the fourth reconstructor and the fifth reconstructor. The outputs of the sixth extractor and the seventh extractor are sent to the third cross-layer feature fusion device for processing, and the output of the third feature fusion device is used as a supplementary input of the sixth reconstructor and the seventh reconstructor.

[0008] Preferably, the specific structure of the infrared image feature extractor is: First, two convolutional layers with a convolution kernel size of 3*3 are used to extract preliminary features of the image, and the Relu activation function is added after the two convolutional layers to realize the nonlinear activation of the features; then the obtained features are sent to the global feature perception branch to focus on extracting the global thermal field features of the infrared image, and the features are sent to the detail feature perception branch to focus on extracting the local thermal field change features of the infrared image; in the global feature perception branch, three convolutional layers with a convolution kernel size of 5*5 are first connected in sequence, and then the normalization layer is used to perform feature standardization, and then the average pooling layer is used to complete the average pooling operation of the features; in the detail feature perception branch, four convolutional layers with a convolution kernel size of 3*3 are first used, and then the normalization layer is also used to perform feature standardization, and then the maximum pooling operation of the features is completed through the maximum pooling layer; then the feature fusion processing of the global thermal field features and the local thermal field features is realized through the fully connected layer, and the final feature extraction operation is performed through the convolution kernel 3*3 convolution layer, and finally the feature activation is realized using simgoid activation.

[0009] Preferably, the specific structure of the thermal field sensing image reconstructor is: Firstly, two deconvolution layers with a convolution kernel size of 3*3 are used to perform feature reconstruction upsampling, and two feature reconstruction pathways with different convolution kernel sizes are constructed. The first feature reconstruction pathway includes three sequentially connected deconvolution layers with a convolution kernel size of 5*5, followed by a BN normalization layer and a Dropout layer to enhance the image reconstruction robustness of the thermal field perception image reconstructor; the second feature reconstruction pathway includes three sequentially connected deconvolution layers with a convolution kernel size of 3*3, followed by a BN normalization layer and an anti-pooling layer, and finally the fully connected layer and the mish activation function are used to perform the final feature fusion processing on the features of the two feature reconstruction pathways.

[0010] Preferably, the cross-layer feature fusion device first realizes feature extraction processing for two sequentially connected 3*3 convolution layers, then follows the Relu activation function to realize nonlinear activation of features, and then realizes upsampling processing of features through two mutually connected 3*3 deconvolution layers, and finally uses the maximum pooling layer and Relu activation function to process again to obtain the fused ring network box thermal field perception image .

[0011] Preferably, the ring network box thermal coefficient perception model includes a ring network box multi-layer feature perception path, a ring network box multi-layer feature fusion unit and a ring network box thermal coefficient generation head; The ring network box multi-layer feature perception path first converts the data The data is sent to three different feature extraction paths to complete the feature extraction operation. The network layer architecture of the first feature extraction path is three fully connected layers connected in sequence. Relu activation function is added after each fully connected layer. After the first feature extraction path is processed, the feature The network layer architecture of the second feature extraction path is still three fully connected layers connected in sequence, but each fully connected layer is added with Activation function, data After processing by the second feature extraction path, the feature The network layer architecture of the third feature extraction path is three fully connected layers connected in sequence, and each fully connected layer is added Activation function, data After processing through the feature extraction path three, the feature is obtained ; The ring network box multi-layer feature fusion unit is used to fuse the features extracted by the ring network box multi-layer feature perception path and adaptively assign weights to different features; The ring network box thermal coefficient generation head is composed of two consecutive fully connected layers to fusion features. Do further feature extraction processing, then pass The activation function realizes the nonlinear activation of the feature, and finally passes through The prediction results of the heat generation coefficient of the ring network box are obtained by activation function processing and heat dissipation coefficient prediction results .

[0012] Preferably, seven key control parameters in the ring network box cooling system are selected as working variables, including the number of fans working fg, fan speed fz, air outlet opening sj, coolant flow rate yn of the liquid cooling system, circulating pump working power wp, heat dissipation device start-stop threshold temperature difference wc, and the number of start-stop times of the circulating pump xc; The multi-objective optimization function is as follows:

[0013] in is the cooling efficiency objective function, minimizing this objective function is used to reduce the secondary heat additionally introduced when the ring network box cooling device is working; is the heat generation coefficient of the ring network box when it is working, is the heat dissipation coefficient of the ring network box when it is working, and All of them are calculated by the heat generation coefficient perception model of the ring network box. is the overall temperature of the ring network box at this time; represents the calculation function of the heat dissipation efficiency of the force field, Represents the calculation function of heat dissipation efficiency of thermal field factor; The objective function of cooling consumption is minimized to reduce the resources and energy consumption in the actual cooling process; represents the consumption cost coefficient; is the component service life objective function.

[0014] Preferably, the optimization range limiting operation is performed on the seven working variables, specifically including: , that is, the minimum value of the number of fans working fg is , the maximum value is ; , that is, the minimum fan speed is The maximum speed is change; , that is, the minimum opening angle of the air outlet is degrees, the maximum opening angle is , that is, the minimum flow rate of the coolant in the liquid cooling system is m / s, the maximum flow rate is m / s; , that is, the minimum working power of the circulation pump is Watts, the maximum operating power is watt. , that is, the minimum temperature difference between the start and stop thresholds of the heat dissipation device is Celsius, maximum degrees Celsius; , that is, the minimum number of starts and stops of the circulating pump is times, the maximum number of starts and stops is Second-rate.

[0015] Compared with the prior art, the innovative features of the present invention have the following beneficial effects: (1) Implementation of adaptive ring network box cooling control system: This invention realizes accurate monitoring of the heat generation of the ring network box from the perspectives of the ring network force field and the thermal field by constructing a specific dynamic monitoring data set and a ring network box thermal coefficient perception model and a ring network box thermal field perception model. It also optimizes the working parameters of the cooling device in real time through the designed cooling device optimization algorithm. This realizes a closed-loop system of high-precision thermal field perception-thermal coefficient prediction-cooling strategy optimization, significantly improving the automation and intelligence level of the ring network box heat dissipation control; (2) Design of thermal field perception model with cross-layer feature fusion: This paper designs a six-layer stacked extractor-reconstructor architecture to realize the overall framework design of the thermal field perception model. A dual-channel feature extraction path and a cross-layer feature fusion device are designed to take into account the image reconstruction capabilities of large-scale thermal areas and small-scale thermal areas, further improving the robustness and generalization ability of thermal field image reconstruction, ensuring that good thermal field image perception capabilities can be maintained in different ring network box working scenarios; (3) Design of a multi-objective optimization adaptive cooling control algorithm: This paper designs a multi-objective optimization function and comprehensively considers the actual thermal field conditions of the ring main box. Combined with ε-Pareto front screening and cross-mutation operations, it achieves global optimization of multiple working variables. This algorithm ensures optimization in terms of heat dissipation efficiency, heat dissipation energy consumption, and cooling device life, improves the selectivity of cooling control parameters, and further ensures the comprehensive cooling capacity of the ring main box cooling process. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flow chart of the overall technical route of the present invention.

[0017] Figure 2 This is the architecture diagram of the thermal field perception model of the ring network box of the present invention.

[0018] Figure 3 This is the architecture diagram of the infrared image feature extractor of the present invention.

[0019] Figure 4 This is the architecture diagram of the thermal field sensing image reconstructor of the present invention.

[0020] Figure 5 This is an example of the thermal field perception image of the ring network box.

[0021] Figure 6 This is an example diagram comparing the prediction accuracy of fan speed control parameters.

[0022] Figure 7 This is an example diagram comparing the prediction accuracy of the air outlet opening control parameters. DETAILED DESCRIPTION

[0023] The present invention proposes a dynamic optimization method for a primary and secondary fusion ring box cooling device, such as Figure 1 Specifically, it includes: constructing a dynamic monitoring dataset including a ring network box thermal field monitoring dataset and a thermal coefficient monitoring dataset, wherein the thermal field dataset is constructed by five-view infrared image stitching and high-density thermocouple calibration, and the thermal coefficient dataset is constructed by combining mechanical data such as vibration and stress with manual calibration; constructing a thermal coefficient perception model, and through three-branch feature extraction and adaptive weighted fusion, evaluating the heat generation coefficient and heat dissipation coefficient of the ring network box in real time based on force field data; constructing a thermal field perception model, adopting an extractor-reconstructor architecture and a cross-layer feature fusion device to achieve full-scene temperature perception; designing a cooling optimization control algorithm, combining the thermal field distribution and thermal coefficient, and adaptively adjusting seven cooling device parameters such as fan speed and liquid cooling flow rate through multi-objective optimization (cooling efficiency, energy consumption, component loss) to achieve refined cooling control.

[0024] The invention will be further described below with reference to specific embodiments.

[0025] 1. Construction of dynamic monitoring data set for ring network box This invention aims to dynamically monitor the thermal field of a ring mainframe (RMC) from a thermal perspective. Furthermore, the RMG corresponds to different heat generation coefficients under different mechanical conditions, which affect the RMG's overall heat dissipation performance. Due to the limited coupling between mechanical and thermal data, to enhance the subsequent optimization of the RMG's heat dissipation device parameters, during the RMG dynamic monitoring dataset construction phase, a RMG thermal field prediction dataset and a RMG heat generation coefficient prediction dataset were specifically constructed.

[0026] 1. Method for constructing ring network box thermal field prediction dataset To improve the spatial resolution and efficiency of subsequent thermal field prediction, this invention uses infrared images of the ring main box as input for the thermal field prediction dataset. This reduces the complexity of subsequent data calibration and improves the comprehensiveness of the ring main box temperature perception.

[0027] It is particularly noted that the present invention further performs image calibration processing on the collected infrared perception image of the ring network box, converts the infrared image of the ring network box into an actual thermal field perception image of the ring network box, and uses the converted thermal field perception image of the ring network box as the output data of the thermal field prediction data set of the ring network box; (1) Infrared image perception of ring network box: When collecting infrared images of ring network box, it is difficult to collect effective infrared perception images because the bottom surface of the ring network box is fixed and blocked by the ground. Therefore, infrared images of the front, back, left, right and top of the cube ring network box are collected, and the infrared images collected from five different perspectives are spliced ​​to form an infrared perception image of the whole ring network box. ; (2) Calibration of infrared images of ring network boxes: Since the ring network box is an electrical device, in order to ensure the accuracy of temperature data collection and the safety of the collection process, the present invention carries out the calibration operation of infrared images of the ring network box under laboratory conditions: First, a high-density thermocouple array is evenly arranged on the surface of the ring network box. To ensure the accuracy of temperature perception, the distance between any two thermocouples is less than 1 cm, and the thermocouples are fixed to the surface of the ring network box by high-temperature resistant tape. Then, the correspondence between the physical coordinates of the thermocouples and the pixel coordinates of the infrared image is established through the checkerboard calibration plate, and the temperature value actually collected by the thermocouple is used as the true value label of the corresponding pixel point. Then, the local weighted regression difference method is used to supplement the temperature labels of the pixels that lack the true value label, and finally the calibrated thermal field perception image of the ring network box is obtained. and use it as the output data for the thermal field prediction of the ring network box.

[0028] 2. Method for constructing a data set for predicting the heat generation coefficient of ring main box During the actual operation of the ring network cabinet, the vibration state of its internal structure, the stress and strain distribution, and the changes in the tension and resistance of the contact interface will cause changes in the contact resistance and thermal conductivity coefficient, thereby causing changes in the heat dissipation capacity and heat generation capacity of the ring network cabinet. This paper defines the heat dissipation coefficient To evaluate the heat dissipation capacity of the ring network box, the heat generation coefficient is defined To evaluate the heat generation capacity of the ring network box. and heat dissipation coefficient It cannot be directly measured by the equipment. In this invention, professional operation and maintenance personnel conduct in-depth analysis on the mechanical data of the ring network cabinet to obtain the corresponding heat generation coefficient. and heat dissipation coefficient .

[0029] (1) In this embodiment, acceleration sensors are deployed at the four corners of the switch operating mechanism connecting rod, busbar support insulator base, cable terminal head, gas insulation wall shell and cabinet base of the ring network cabinet to collect corresponding vibration frequency data. And vibration amplitude data , deploy grating strain sensors to collect stress distribution Data, deploy tension and pressure sensors to collect corresponding tension data and pressure data , deploy resistance strain gauges to collect stress and deformation data The mechanical characteristic data of the ring network box collected above are combined to obtain a set of input data in the ring network box heat generation coefficient prediction data set. ; (2) In view of the above-described ring network box mechanical data collection method, in order to ensure the accuracy of the subsequent evaluation of the heat generation coefficient of the ring network box, the ring network box mechanical operation status data collection is carried out in three different states: the ring network box overload operation state, the ring network box normal operation state and the ring network box underload operation state to ensure the integrity of the data collection; (3) Based on the data collection environment described above, actual data collection is carried out, and input data for each group of heat generation coefficient prediction is made through manual judgment by professionals. Add a heat generation coefficient tag and heat dissipation coefficient labels The two coefficient data are combined as the output data in the ring network box heat generation coefficient prediction data set .

[0030] Based on the described ring network box thermal field prediction data set construction method, this embodiment carries out ring network box thermal field prediction data collection under different outdoor temperature conditions, and a total of The thermal field prediction data of the ring network box is collected and the Combine the ring network box thermal field prediction data to complete the ring network box thermal field prediction data set 's construction.

[0031] Based on the described ring network box heat generation coefficient prediction data set construction method, the present invention collects a total of The heat generation coefficient prediction data of the ring network box is collected and the Combine the ring network box heat generation coefficient prediction data to complete the ring network box heat generation coefficient prediction data set 's construction.

[0032] The above-mentioned ring network box thermal field prediction data set and the heat generation coefficient prediction data set of the ring network box constructed Combine to complete the dynamic monitoring data set of the ring network box 's construction.

[0033] 2. Construction of thermal field perception model of ring network box In order to achieve high-standard generation of thermal field perception images of ring network boxes, the present invention designs an infrared image feature extractor-thermal field perception image reconstructor architecture in the thermal field perception model of ring network boxes. The overall architecture of the thermal field perception model of ring network boxes is shown in the figure below. Figure 2 shown.

[0034] 1. Infrared image feature extractor structure design The infrared image feature extractor architecture is as follows Figure 3 As shown in the figure, the infrared image feature extractor is mainly used to extract the overall infrared perception image of the ring network box. The feature extraction is performed step by step to provide sufficient features for the subsequent thermal field perception image reconstruction. A dual-channel feature extraction path is introduced in this feature extractor to further improve the feature extraction capability of infrared images.

[0035] The specific architecture of the infrared image feature extractor includes: first, using two convolutional layers with a convolution kernel size of 3*3 to perform preliminary feature extraction on the image, and adding a Relu activation function after the two convolutional layers to realize nonlinear activation of the features. The obtained features are then sent to the global feature perception branch to focus on extracting the global thermal field features of the infrared image, and the features are sent to the detail feature perception branch to focus on extracting the local thermal field change features of the infrared image. In the global feature perception branch, three convolutional layers with a convolution kernel size of 5*5 are connected in sequence, followed by feature normalization using a normalization layer, and then the average pooling operation of the features is completed using an average pooling layer. In the detail feature perception branch, four convolutional layers with a convolution kernel size of 3*3 are first, followed by feature normalization using a normalization layer, and then the maximum pooling operation of the features is completed using a maximum pooling layer. Afterwards, a fully connected layer is used to realize the feature fusion processing of the global thermal field features and the local thermal field features, and a convolution layer with a convolution kernel of 3*3 is used to perform the final feature extraction operation. Finally, a simgoid activation is used to realize feature activation to complete the construction of the infrared image feature extractor.

[0036] 2. Thermal field sensing image reconstructor structure design Thermal field sensing image reconstructor architecture Figure 4 As shown in the figure, a thermal field perception image reconstructor is required to generate thermal field perception images in the construction of the ring network box thermal field perception model. The architecture of the thermal field perception image reconstructor is as follows: First, two deconvolution layers with a convolution kernel size of 3*3 are used to perform feature reconstruction upsampling. Two feature reconstruction pathways with different convolution kernel sizes are constructed. Feature reconstruction pathway one includes three sequentially connected deconvolution layers with a convolution kernel size of 5*5, followed by a batch normalization layer and a dropout layer to enhance the image reconstruction robustness of the thermal field perception image reconstructor. Feature reconstruction pathway two includes three sequentially connected deconvolution layers with a convolution kernel size of 3*3, followed by a batch normalization layer and an unpooling layer to complete the design of feature reconstruction pathway two. Finally, a fully connected layer and a mish activation function are used to perform the final feature fusion processing on the features of the two feature reconstruction pathways to complete the design of the thermal field perception image reconstructor.

[0037] 3. Cross-layer feature fusion structure design In order to further ensure the reconstruction accuracy of the thermal field reconstruction image, the present invention enhances the feature fusion processing capability of cross-layer features by constructing a cross-layer feature fuser. The structure of the cross-layer feature fuser is as follows: To ensure the consistency of the feature hierarchy, in the cross-layer feature fuser, two 3*3 convolutional layers connected in sequence are first used to implement feature extraction processing, followed by a Relu activation function to implement nonlinear activation of the features, and then two interconnected 3*3 deconvolution layers are used to implement feature upsampling processing. Finally, a maximum pooling layer and a Relu activation function are used to process again to complete the overall design of the cross-layer feature fuser.

[0038] The overall architecture design of the thermal field perception model of the ring grid box: the overall feature extraction path of the thermal field perception model of the ring grid box is composed of seven infrared image feature extractors stacked in sequence and connected front to back, and the thermal field image reconstruction path of the thermal field perception model is composed of six thermal field perception image reconstructors stacked in sequence and connected front to back. And in order to further establish a direct feature mapping relationship between the extractor and the reconstructor, the image resolution and accuracy of the thermal field reconstructed image are guaranteed. The present invention sends the output of extractor two and extractor three to the cross-layer feature fusion device one for processing, and then uses the output of the feature fusion device as a supplementary input to the reconstructor two and reconstructor three. Correspondingly: after the output of extractor four and extractor five is sent to the cross-layer feature fusion device two for processing, the output of the feature fusion device two is used as a supplementary input to the reconstructor four and reconstructor five. After the output of extractor six and extractor seven is sent to the cross-layer feature fusion device three for processing, the output of the feature fusion device three is used as a supplementary input to the reconstructor six and reconstructor seven.

[0039] An infrared perception image of the entire ring network box After being processed by the above constructed ring network box thermal field perception model, the thermal field perception image of the ring network box calculated by the model can be obtained. .

[0040] 3. Construction of thermal coefficient perception model of ring network box Data processing layer design: Prediction data set for heat generation coefficient of ring network box Input data for heat production coefficient prediction in , The data in the data have different dimensions and value ranges, which makes the prediction of the heat generation coefficient of the ring network box more difficult. In order to reduce the difficulty of predicting the heat generation coefficient of the ring network box and increase the prediction accuracy, the BN normalization layer and the data normalization layer are first designed in the ring network box heat generation coefficient perception model to perform the input data Implement data enhancement processing to obtain enhanced data .

[0041] Ring network box multi-layer feature perception channel design: the enhanced data The data is sent to three different feature extraction paths to complete the feature extraction operation. The network layer architecture of the feature extraction path 1 is three fully connected layers connected in sequence. A Relu activation function is added after each fully connected layer. After processing through the feature extraction path 1, the feature The network layer architecture of the feature extraction path 2 is still three fully connected layers connected in sequence, but each fully connected layer is added with Activation function, data After processing through the second feature extraction path, the feature , the network layer architecture of feature extraction path three is three fully connected layers connected in sequence, and each fully connected layer is added Activation function, data After processing through the feature extraction path three, the feature is obtained .

[0042] Ring network box multi-layer feature fusion unit design: based on the obtained features 、 、 A multi-layer feature fusion unit of the ring network box is designed to obtain fused perception features ;

[0043] in, represents the normalization operation function, Adaptively assign weights to three different features, Denoted as feature scaling factor.

[0044] Design of thermal coefficient generator for ring network box: based on the obtained fusion perception features , the features Send it to the ring network box coefficient generator for further processing to obtain the predicted result of the ring network box heat generation coefficient and heat dissipation coefficient prediction results The overall architecture of the heat generation coefficient generator of the ring network box is as follows: It first integrates the features through two consecutive fully connected layers. Do further feature extraction processing, then pass The activation function realizes further nonlinear activation of the feature and finally passes through After activation function processing, the design process of the thermal coefficient generating head of the ring network box is completed.

[0045] 4. Implementation of Ring Network Box Cooling Control Optimization Algorithm In order to achieve high efficiency optimization of the cooling device of the ring network box, if the thermal field perception image is directly sent to the optimization algorithm for direct processing, the calculation complexity of the optimization calculation will be too high. Therefore, the present invention uses the thermal field image data processing algorithm to obtain the thermal field perception image of the ring network box. The mean temperature and temperature variance . And use this temperature mean and temperature variance As a reference condition parameter for subsequent optimization algorithms; The present invention takes maximizing the cooling efficiency of the ring network box, minimizing the cooling consumption of the ring network box, and maximizing the service life of the heat dissipation components as the model optimization objectives, and uses the cooling equipment working parameters in the ring network box cooling control process as working variables to establish a multi-objective optimization problem for the control parameters of the ring network box cooling equipment. The specific process is as follows: Determination of working variables: To optimize the key control parameters of the ring main box's adaptive cooling process, this paper selects seven key control parameters in the ring main box's cooling system as working variables. These include: number of fans operating fg, fan speed fz, air outlet opening sj, coolant flow rate yn in the liquid cooling system, circulating pump operating power wp, heat sink start / stop threshold temperature difference wc, and circulating pump start / stop times xc. Multi-objective optimization function design: The multi-objective optimization function designed by the present invention is as follows:

[0046] in The objective function of cooling efficiency is minimized, which is beneficial to reducing the secondary heat introduced by the cooling device of the ring network box when it is working, and dynamically adjusting the control parameters of the cooling device according to the current thermal field state of the ring network box, thereby improving the cooling efficiency of the ring network box during actual cooling. is the heat generation coefficient of the ring network box when it is working, is the heat dissipation coefficient of the ring network box when it is working, and All of them can be calculated by the heat generation coefficient perception model of the ring network box. It is the overall temperature of the ring network box at this time. represents the calculation function of the heat dissipation efficiency of the force field, Represents the calculation function of thermal field factor heat dissipation efficiency.

[0047] The objective function of cooling consumption is to minimize the objective function, which is beneficial to reduce the resources and energy consumption in the actual cooling process and improve the economic efficiency of the cooling operation. Represents the consumption cost coefficient. is the component service life objective function. Minimizing this objective function is beneficial to reducing the component loss of the cooling device during the actual cooling process of the ring network box, thereby further extending the service life of the ring network box cooling device and ensuring the robustness and stability of the cooling operation.

[0048] Limiting the value range of working variables: In order to ensure the subsequent optimization space and optimization efficiency of the proposed working parameters, the optimization range of the seven proposed working variables is limited in combination with the actual ring network box heat dissipation working system. Specifically: , that is, the minimum value of the number of fans working fg is , the maximum value is ; , that is, the minimum fan speed is The maximum speed is change; , that is, the minimum opening angle of the air outlet is degrees, the maximum opening angle is , that is, the minimum flow rate of the coolant in the liquid cooling system is m / s, the maximum flow rate is m / s; , that is, the minimum working power of the circulation pump is Watts, the maximum operating power is watt. , that is, the minimum temperature difference between the start and stop thresholds of the heat dissipation device is Celsius, maximum degrees Celsius. , that is, the minimum number of starts and stops of the circulating pump is times, the maximum number of starts and stops is Second-rate.

[0049] Initialization of ring network box working variable population: Based on the value interval of each working variable described in detail, the Bayesian sampling method is used to initialize the working variable population of the ring network box. Sampling within its value range Obtained Each working variable set corresponds to a ring network box working variable population. The population is obtained in the population initialization stage. , .

[0050] Working variable population mutation iteration: (1) Combined with the heat generation coefficient obtained by monitoring and heat dissipation coefficient and monitored , using the three objective functions defined Calculate the fitness of each ring box variable population. When calculating for the first time, the selected population is the initial population. ; (2) According to the fitness of each working variable population calculated in the above (1), all populations are sorted according to their fitness, and the population located at the first frontier of ε-Pareto is selected. and discard the remaining population; (3) For populations Perform crossover and mutation operations to obtain the mutated population ; (4) The mutated population Send it into the process (1) as a new initialized population to calculate the population fitness; (5) Iterate process (1) to process (4) k times to allow the population to fully mutate and compete, and The population obtained by the iteration As the best Pareto solution set for the ring network box cooling control parameter optimization problem, the working variables of the population individual with the smallest fitness variance value are selected from the solution set as the cooling control parameters actually used in the final ring network box.

[0051] 5. Experimental Results In order to clarify the image reconstruction effect of the ring network box thermal field perception model of the invention, the intercepted ring network box thermal field reconstruction image is as follows Figure 5 shown.

[0052] In order to verify that the cooling device optimization control algorithm proposed in this invention has a good control effect, the algorithm performance comparison test was carried out under four different working conditions: underload, overload, normal operation, and minor fault. The comparison algorithms were fuzzy control algorithm and BP neural network-based control algorithm. First, the optimal cooling device control parameters under the above working conditions were determined by professionals. , and record the control parameters predicted by each algorithm , and then calculate the parameter prediction accuracy , Indicates the absolute value operation.

[0053] The algorithm comparison and verification were carried out using the air outlet opening control parameters and fan control parameters as examples. Figure 6 and Figure 7 As shown in the figure, the verification results show that the algorithm proposed in this paper has the highest prediction accuracy for the optimal control parameters under four different ring main box operating modes. This is because the cooling device optimization control algorithm proposed in this paper can fully explore the optimization space of control parameters and select the optimal cooling device control parameters based on this full exploration of the control parameter optimization space, thus ensuring the optimality of the control parameters.

[0054] The above description is merely a preferred embodiment of the present application and is not intended to limit the present application. Various modifications and variations are possible for those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present application shall be included within the scope of protection of the present application.

[0055] Although the above describes the specific implementation methods of the present invention, it does not limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A dynamic optimization method for a primary and secondary fusion ring network box cooling device, characterized in that: The following processes are included: S1, dynamically acquiring infrared perception images and mechanical characteristic data of the ring network box, wherein the infrared perception images of the ring network box are constructed as thermal field monitoring data of the ring network box as standard input of the model by splicing infrared images from five perspectives; the mechanical characteristic data of the ring network box are combined to obtain thermal coefficient monitoring data of the standard input of the model; S2: Input the thermal field monitoring data of the ring network box into the constructed and trained thermal field perception model of the ring network box, output the thermal field perception image, and calculate the temperature mean and temperature variance; input the thermal coefficient monitoring data into the constructed and trained thermal coefficient perception model of the ring network box, and output the heat generation coefficient and heat dissipation coefficient; S3, with maximizing the cooling efficiency of the ring network box, minimizing the cooling consumption of the ring network box, and maximizing the service life of the heat dissipation components as the model optimization objectives, takes the cooling equipment operating parameters during the ring network box cooling control process as the working variables, and combines the temperature mean, temperature variance, heat generation coefficient, and heat dissipation coefficient to construct a multi-objective optimization function; S4, solves the multi-objective optimization function and outputs refined cooling control parameters.

2. The method for dynamic optimization of a primary and secondary fusion ring network box cooling device according to claim 1, characterized in that: The dataset used to train the thermal field perception model of the ring grid enclosure is constructed in the following ways: Infrared images of the front, back, left, right and top of the cube ring network box were collected, and the infrared images collected from five different perspectives were stitched together to form an infrared perception image of the ring network box as a whole. ; A high-density thermocouple array is evenly arranged on the surface of the ring network box, and the thermocouples are fixed to the surface of the ring network box with high-temperature resistant tape. Then, the correspondence between the physical coordinates of the thermocouples and the pixel coordinates of the infrared image is established through a checkerboard calibration plate, and the temperature value actually collected by the thermocouple is used as the true value label of the corresponding pixel point. Then, the local weighted regression difference method is used to supplement the temperature label of the pixel points that lack the true value label, and finally the calibrated thermal field perception image of the ring network box is obtained. and use it as the output data for the thermal field prediction of the ring network box.

3. The method for dynamic optimization of a primary and secondary fusion ring network box cooling device according to claim 1, characterized in that: The dataset used to train the ring network box thermal coefficient perception model is constructed in the following ways: Acceleration sensors are deployed at the four corners of the switch operating mechanism connecting rod of the ring network cabinet, the busbar support insulator base, the cable terminal head, the gas insulation wall shell and the cabinet base to collect the corresponding vibration frequency data. And vibration amplitude data , deploy grating strain sensors to collect stress distribution Data, deploy tension and pressure sensors to collect corresponding tension data and pressure data , deploy resistance strain gauges to collect stress and deformation data The above collected mechanical characteristic data of the ring network box are combined to obtain a set of input data in the ring network box heat generation coefficient prediction data set. ; Predict input data for each group of heat production coefficients through manual judgment Added heat generation coefficient label and heat dissipation coefficient labels The two coefficient data are combined as the output data in the ring network box heat generation coefficient prediction data set .

4. The method for dynamic optimization of a primary and secondary fusion ring network box cooling device according to claim 1, characterized in that: The thermal field perception model of the ring network box includes an infrared image feature extractor, a thermal field perception image reconstructor and a cross-layer feature fusion device; The overall feature extraction path of the thermal field perception model of the ring network box is composed of seven infrared image feature extractors stacked in sequence and connected front to back, and the thermal field image reconstruction path of the thermal field perception model is composed of six thermal field perception image reconstructors stacked in sequence and connected front to back; the outputs of the second extractor and the third extractor are sent to the first cross-layer feature fusion device for processing, and then the output of the feature fusion device is used as a supplementary input of the second reconstructor and the third reconstructor. At the same time, the outputs of the fourth extractor and the fifth extractor are sent to the second cross-layer feature fusion device for processing, and the output of the second feature fusion device is used as a supplementary input of the fourth reconstructor and the fifth reconstructor. The outputs of the sixth extractor and the seventh extractor are sent to the third cross-layer feature fusion device for processing, and the output of the third feature fusion device is used as a supplementary input of the sixth reconstructor and the seventh reconstructor.

5. The method for dynamic optimization of a primary and secondary fusion ring network box cooling device according to claim 4, characterized in that: The specific structure of the infrared image feature extractor is: First, two convolutional layers with a convolution kernel size of 3*3 are used to extract preliminary features of the image, and the Relu activation function is added after the two convolutional layers to realize the nonlinear activation of the features; then the obtained features are sent to the global feature perception branch to focus on extracting the global thermal field features of the infrared image, and the features are sent to the detail feature perception branch to focus on extracting the local thermal field change features of the infrared image; in the global feature perception branch, three convolutional layers with a convolution kernel size of 5*5 are first connected in sequence, and then the normalization layer is used to perform feature standardization, and then the average pooling layer is used to complete the average pooling operation of the features; in the detail feature perception branch, four convolutional layers with a convolution kernel size of 3*3 are first used, and then the normalization layer is also used to perform feature standardization, and then the maximum pooling operation of the features is completed through the maximum pooling layer; then the feature fusion processing of the global thermal field features and the local thermal field features is realized through the fully connected layer, and the final feature extraction operation is performed through the convolution kernel 3*3 convolution layer, and finally the feature activation is realized using simgoid activation.

6. The method for dynamic optimization of a primary and secondary fusion ring network box cooling device according to claim 4, characterized in that: The specific structure of the thermal field sensing image reconstructor is: Firstly, two deconvolution layers with a convolution kernel size of 3*3 are used to perform feature reconstruction upsampling, and two feature reconstruction pathways with different convolution kernel sizes are constructed. The first feature reconstruction pathway includes three sequentially connected deconvolution layers with a convolution kernel size of 5*5, followed by a BN normalization layer and a Dropout layer to enhance the image reconstruction robustness of the thermal field perception image reconstructor; the second feature reconstruction pathway includes three sequentially connected deconvolution layers with a convolution kernel size of 3*3, followed by a BN normalization layer and an anti-pooling layer, and finally the fully connected layer and the mish activation function are used to perform the final feature fusion processing on the features of the two feature reconstruction pathways.

7. The method for dynamic optimization of a cooling device for a primary and secondary fusion ring network box according to claim 4, characterized in that: The cross-layer feature fusion device first implements feature extraction processing for two sequentially connected 3*3 convolution layers, then follows the Relu activation function to implement nonlinear activation of features, and then implements feature upsampling processing through two mutually connected 3*3 deconvolution layers. Finally, the maximum pooling layer and Relu activation function are used to process again to obtain the fused ring network box thermal field perception image. .

8. The method for dynamic optimization of a cooling device for a primary and secondary fusion ring network box according to claim 1, characterized in that: The ring network box thermal coefficient perception model includes a ring network box multi-layer feature perception path, a ring network box multi-layer feature fusion unit and a ring network box thermal coefficient generation head; The ring network box multi-layer feature perception path first converts the data The data is sent to three different feature extraction paths to complete the feature extraction operation. The network layer architecture of the first feature extraction path is three fully connected layers connected in sequence. Relu activation function is added after each fully connected layer. After the first feature extraction path is processed, the feature The network layer architecture of the second feature extraction path is still three fully connected layers connected in sequence, but each fully connected layer is added with Activation function, data After processing by the second feature extraction path, the feature , the network layer architecture of the third feature extraction path is three fully connected layers connected in sequence, and each fully connected layer is added Activation function, data After processing through the feature extraction path three, the feature is obtained ; The ring network box multi-layer feature fusion unit is used to fuse the features extracted by the ring network box multi-layer feature perception path and adaptively assign weights to different features; The ring network box thermal coefficient generation head is composed of two consecutive fully connected layers to fusion features. Do further feature extraction processing, then pass The activation function realizes the nonlinear activation of the feature, and finally passes through The prediction results of the heat generation coefficient of the ring network box are obtained by activation function processing and heat dissipation coefficient prediction results .

9. The method for dynamic optimization of a primary and secondary fusion ring network box cooling device according to claim 1, characterized in that: Seven key control parameters in the ring network box cooling system were selected as working variables, including the number of fans working fg, fan speed fz, air outlet opening sj, coolant flow rate yn of the liquid cooling system, circulating pump working power wp, heat dissipation device start-stop threshold temperature difference wc, and the number of start-stop times of the circulating pump xc; The multi-objective optimization function is as follows: in is the cooling efficiency objective function, minimizing this objective function is used to reduce the secondary heat additionally introduced when the ring network box cooling device is working; is the heat generation coefficient of the ring network box when it is working, is the heat dissipation coefficient of the ring network box when it is working, and All of them are calculated by the heat generation coefficient perception model of the ring network box. is the overall temperature of the ring network box at this time; represents the calculation function of the heat dissipation efficiency of the force field, Represents the calculation function of heat dissipation efficiency of thermal field factor; The objective function of cooling consumption is minimized to reduce the resources and energy consumption in the actual cooling process; represents the consumption cost coefficient; is the component service life objective function.

10. The method for dynamic optimization of a primary and secondary fusion ring network box cooling device according to claim 9, characterized in that: Perform optimization range limiting operations on seven working variables, including: , that is, the minimum value of the number of fans working fg is , the maximum value is ; , that is, the minimum fan speed is The maximum speed is change; , that is, the minimum opening angle of the air outlet is degrees, the maximum opening angle is , that is, the minimum flow rate of the coolant in the liquid cooling system is m / s, the maximum flow rate is m / s; , that is, the minimum working power of the circulation pump is Watts, the maximum operating power is watt. , that is, the minimum temperature difference between the start and stop thresholds of the heat dissipation device is Celsius, maximum degrees Celsius; , that is, the minimum number of starts and stops of the circulating pump is times, the maximum number of starts and stops is Second-rate.

Images