Pipe-through cable thermal field distribution characteristic dynamic inversion method and system

Abstract

The present application belongs to the technical field of power system cable state monitoring, and particularly relates to a method and system for dynamically inverting the distribution characteristics of the thermal field of a pipe-through cable. The method of the present application obtains real-time operating parameters and environmental parameters of the pipe-through cable and generates a time-series feature data set, extracts current-carrying thermal shock, medium thermal diffusion and soil heat storage hysteresis characteristics using a multi-scale time-series convolution network architecture, and outputs an initial predicted state vector. A prediction residual is formed according to the outer surface temperature monitoring data of the sheath, boundary condition correction coefficients are identified, and the target channel weight matrix is scaled according to the physical interface mapping relationship. When the residual does not continue to decrease, learning rate decay and convolution kernel pruning are performed to obtain a time-domain dynamic inversion model. The spatial distribution parameters of the thermal field are then output and subjected to outlier detection and temperature threshold determination, and a graded warning signal is output. The present application is used to reduce the accumulation of thermal field inversion errors caused by changes in dynamic boundary conditions.

Description

Technical Field

[0001] This invention relates to the field of power system cable condition monitoring technology, specifically to a method and system for dynamic inversion of the thermal field distribution characteristics of cables in conduits. Background Technology

[0002] During operation, buried cables generate Joule heat as the current flows through the cable conductor. This heat is then transferred stepwise along the cable conductor, insulation layer, sheath, air inside the conduit, the outer wall of the conduit, and the surrounding soil. For 110kV and above high-voltage cables, the cables are typically laid inside high-density polyethylene conduits, with the conduit's exterior covered by non-uniform soil. This laying structure means that heat transfer from the cable conductor to the external environment is affected by changes in current carrying capacity, heat exchange within the conduit air, soil temperature, humidity of the surrounding medium, and the soil's heat storage capacity.

[0003] Current methods for analyzing the thermal field of cables in conduits typically employ finite element simulation, transient thermal circuit models, core temperature inversion models based on the outer surface temperature of the sheath, or temperature estimation models based on neural networks. Finite element simulation can describe the heat conduction process of a local structure, but it usually relies on preset material parameters and boundary conditions. Transient thermal circuit models are computationally fast, but when soil moisture content changes, air heat transfer within the conduit changes, or short-term overloads occur, the preset model parameters cannot reflect the actual heat transfer path changes in a timely manner. Inversion methods based on the outer surface temperature of the sheath can estimate the internal temperature using temperature measurement data, but the inversion relationship usually relies on a fixed mapping. Temperature estimation models based on neural networks can obtain temperature outputs using operational data, but if there is a lack of online correction paths corresponding to physical boundary conditions, model output errors can easily accumulate in multi-step predictions.

[0004] Dynamic inversion of the thermal field of cables in conduits is not a simple problem of estimating a single temperature point, but rather a time-varying boundary inversion problem formed by the combined effects of current-carrying heat generation, radial heat transfer, pipe heat exchange, soil heat storage, and measurement feedback. Sudden changes in current carrying capacity cause high-frequency thermal shocks, variations in the humidity of the surrounding medium cause slow changes in external thermal resistance, and diurnal variations in soil temperature cause low-frequency hysteresis. These three types of changes have different time scales. If feature extraction methods of the same scale or fixed boundary conditions are used, short-term thermal shocks and long-term heat storage will be mixed into the same source of error. The model will find it difficult to determine whether the error originates from current-carrying fluctuations or changes in the heat transfer boundary, leading to deviations in the spatial distribution parameters of the thermal field from the actual operating state.

[0005] Therefore, there is a need for a dynamic inversion method and system that can integrate real-time operating parameters, environmental parameters, and sheath outer surface temperature monitoring data into the dynamic inversion process, and convert the prediction residuals into boundary condition correction coefficients, so that the spatial distribution of the thermal field of the cable in conduit can be updated with changes in current carrying state and external heat dissipation conditions. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for dynamic inversion of the thermal field distribution characteristics of cable in conduit. The technical problem to be solved by this invention is that the existing thermal field analysis of cable in conduit relies on fixed boundary conditions or fixed mapping relationships in complex laying environments. It cannot convert the prediction residual formed by the temperature monitoring data of the outer surface of the sheath into the boundary condition correction amount of the physical interface, resulting in the cumulative deviation of the spatial distribution parameters of the thermal field under the conditions of sudden current change and soil moisture change.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0008] A method for dynamically inverting the thermal field distribution characteristics of cables in conduits is provided, including:

[0009] The real-time operating parameters and environmental parameters of the cable in the conduit are obtained, and the real-time operating parameters and environmental parameters are synchronously corrected and timestamp aligned to generate a time-series feature dataset.

[0010] The temporal feature dataset is input into a multi-scale temporal convolutional network architecture. The current-carrying thermal shock features, medium thermal diffusion features, and soil heat storage hysteresis features in the temporal feature dataset are extracted through the first, second, and third temporal convolutional layers with increasing receptive fields, and the initial predicted state vector is output.

[0011] Acquire the temperature monitoring data of the outer surface of the sheath, calculate the prediction residual between the temperature monitoring data of the outer surface of the sheath and the predicted temperature dimension of the interface between the outer surface of the sheath and the air inside the pipe in the initial prediction state vector, and identify the boundary condition correction coefficient based on the prediction residual.

[0012] The network weights corresponding to the physical interface in the multi-scale temporal convolutional network architecture are scaled and adjusted using the boundary condition correction coefficients. When the absolute value of the prediction residual does not decrease for a continuous number of consecutive sampling periods, the learning rate decay and convolutional kernel pruning are performed to obtain the temporal dynamic inversion model.

[0013] The time-domain dynamic inversion model is used to output the spatial distribution parameters of the thermal field within the future predicted time domain. Outlier detection and temperature threshold determination are performed on the spatial distribution parameters of the thermal field, and a graded early warning signal is output.

[0014] Furthermore, generating the time-series feature dataset includes: collecting cable conductor current and conductor voltage to form the real-time operating parameters; collecting air temperature, soil temperature and ambient medium humidity in the pipeline to form the environmental parameters; adding absolute timestamps to the real-time operating parameters and the environmental parameters, removing data segments with timestamp deviations exceeding the synchronization deviation threshold, and concatenating the aligned parameters according to the sampling period to form the time-series feature dataset covering the historical time window.

[0015] Furthermore, the first temporal convolutional layer outputs a first feature map with a first convolutional kernel length and a first step length; the second temporal convolutional layer outputs a second feature map with a second convolutional kernel length greater than the first convolutional kernel length and a second step length greater than the first step length; the third temporal convolutional layer outputs a third feature map with a third convolutional kernel length greater than the second convolutional kernel length and a third step length greater than the second step length; and the third feature map is mapped through a fully connected layer to generate the initial prediction state vector containing the predicted temperatures of multiple physical interfaces.

[0016] Furthermore, time-dimension alignment is performed on the first feature map, the second feature map, and the third feature map to make their time series lengths consistent. The time-dimension aligned first feature map, second feature map, and third feature map are then concatenated along the channel dimension and input into the residual connection unit for cross-layer fusion. The fused features are used to generate the initial prediction state vector.

[0017] Furthermore, the boundary condition correction coefficients identified based on the predicted residuals include:

[0018] When the absolute value of the prediction residual exceeds the tolerance dead zone, the regression vector formed by the time series feature dataset and the initial prediction state vector, the measurement output formed by the prediction residual, the parameter estimation vector of the previous time step, and the covariance matrix of the previous time step are extracted, and the recursive least squares method is used to update the parameter estimation vector of the current time step.

[0019] The regression vector includes the predicted temperature characteristics of the interface between the cable conductor and the insulation layer, the predicted temperature characteristics of the interface between the insulation layer and the sheath, the predicted temperature characteristics of the interface between the outer surface of the sheath and the air inside the pipe, and the predicted temperature characteristics of the interface between the outer wall of the pipe and the surrounding soil, arranged in dimensional order.

[0020] The parameter estimation vector at the current moment is a four-dimensional parameter estimation vector corresponding to the dimension of the regression vector;

[0021] The converged parameter estimation vector at the current moment is decomposed in dimensional order and subjected to boundary condition correction coefficient amplitude limiting processing to obtain multiple boundary condition correction coefficients corresponding to the interface between the cable conductor and the insulation layer, the interface between the insulation layer and the sheath, the interface between the outer surface of the sheath and the air inside the pipe, and the interface between the outer wall of the pipe and the surrounding soil.

[0022] Furthermore, when the rate of change of the ambient medium humidity within the environmental determination time window exceeds the humidity change rate threshold, or when the change in soil temperature within the environmental determination time window exceeds the temperature change amplitude threshold, the forgetting factor in the recursive least squares method is reduced, and the recursive calculation data window is shortened.

[0023] Furthermore, scaling and adjusting the network weights corresponding to the physical interface in the multi-scale temporal convolutional network architecture using the boundary condition correction coefficients includes: establishing an interface mapping relationship between the boundary condition correction coefficients and the target channel weight matrix in the first, second, and third temporal convolutional layers; and performing element-wise multiplication of the boundary condition correction coefficients as scaling factors with the target channel weight matrix according to the interface mapping relationship.

[0024] Further, performing the convolutional kernel pruning includes: calculating the absolute value accumulation of the one-dimensional convolutional kernel weight matrix in the multi-scale temporal convolutional network architecture; when the absolute value accumulation of the target one-dimensional convolutional kernel is continuously lower than the pruning threshold within the pruning statistics window, the target one-dimensional convolutional kernel is marked as a redundant convolutional kernel; setting the memory base address pointers pointing to the input feature map and output feature map of the redundant convolutional kernel in the addressing mapping table to null, and skipping the multiply-accumulate operation of the corresponding channel of the redundant convolutional kernel in subsequent forward propagation.

[0025] Furthermore, outputting the spatial distribution parameters of the thermal field within the future prediction time domain includes: mapping the spatial distribution parameters of the thermal field output at the current prediction step size to the feature dimension and concatenating them as prediction state features to the end of the time series feature dataset; supplementing the corresponding prediction step size's real-time operating parameters and environmental parameters with the most recent valid values, planned current carrying values, or environmental prediction values ​​to the corresponding input channels; removing the oldest time data to form an updated time series feature dataset; inputting the updated time series feature dataset back into the time domain dynamic inversion model for the next prediction step size calculation; and performing exponentially weighted moving average correction on the spatial distribution parameters of the thermal field output at each prediction step size.

[0026] Furthermore, outputting the graded early warning signal includes: performing Grubbs outlier detection on the thermal field spatial distribution parameters and generating a detection judgment result; when the detection judgment result indicates the existence of an outlier and the predicted highest temperature value in the thermal field spatial distribution parameters reaches a first temperature threshold, outputting a first-level early warning signal; when the detection judgment result indicates the existence of an outlier and the predicted highest temperature value reaches a second temperature threshold higher than the first temperature threshold, outputting a second-level early warning signal.

[0027] Furthermore, when acquiring the real-time operating parameters and the environmental parameters, dual-channel redundant transmission is performed: polling the physical layer connection status of the main communication channel; when the main communication channel experiences a connection loss or packet loss rate exceeding the limit, waking up the backup communication channel; during channel switching, the unsent time-series feature dataset is cached, and after the backup communication channel establishes a connection, the cached time-series feature dataset is sent in timestamp order.

[0028] Furthermore, it also includes model warm-up and model version rollback: In the system cold start state, the pre-trained general model parameters obtained by training with historical operation data of the same type of cable in conduit are used as the initial weight matrix of the multi-scale temporal convolutional network architecture, and the third temporal convolutional layer is fine-tuned using the local regional temporal feature dataset; when the prediction residual continuously exceeds the rollback threshold within the rollback decision time window and the boundary condition correction coefficient does not converge, the parameter snapshot file of the previous stable state is loaded to overwrite the current model parameters.

[0029] The present invention also discloses a dynamic inversion system for the thermal field distribution characteristics of conduit cables, used to execute the dynamic inversion method for the thermal field distribution characteristics of conduit cables as described above, including: a data acquisition module, a model calculation module, a boundary update module, a prediction module, and an early warning module;

[0030] The data acquisition module is used to acquire the real-time operating parameters and environmental parameters of the cable in the conduit, and to perform synchronous correction and timestamp alignment on the real-time operating parameters and environmental parameters to generate a time-series feature dataset.

[0031] The model calculation module is used to input the temporal feature dataset into a multi-scale temporal convolutional network architecture, and extract the current-carrying thermal shock features, medium thermal diffusion features, and soil heat storage hysteresis features from the temporal feature dataset through the first, second, and third temporal convolutional layers with increasing receptive fields, and output the initial predicted state vector.

[0032] The boundary update module is used to acquire the temperature monitoring data of the outer surface of the sheath, calculate the prediction residual between the temperature monitoring data of the outer surface of the sheath and the predicted temperature dimension of the interface between the outer surface of the sheath and the air inside the pipe in the initial prediction state vector, identify the boundary condition correction coefficient based on the prediction residual, and use the boundary condition correction coefficient to scale and adjust the network weights of the corresponding physical interface in the multi-scale temporal convolutional network architecture.

[0033] The model calculation module is also used to perform learning rate decay and convolution kernel pruning when the absolute value of the predicted residual does not decrease for a continuous number of consecutive sampling periods, so as to obtain a time-domain dynamic inversion model.

[0034] The prediction module is used to output the spatial distribution parameters of the thermal field within the future predicted time domain using the time-domain dynamic inversion model.

[0035] The early warning module is used to perform outlier detection and temperature threshold determination on the spatial distribution parameters of the thermal field, and output graded early warning signals.

[0036] Meanwhile, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the dynamic inversion method for the thermal field distribution characteristics of conduit cables as described above.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] This invention incorporates real-time operating parameters, environmental parameters, and sheath outer surface temperature monitoring data into a single dynamic inversion chain, enabling the dynamic inversion of the thermal field distribution characteristics of conduit cables to no longer rely on fixed boundary conditions or fixed mapping relationships. Real-time operating parameters and environmental parameters, after synchronous correction and timestamp alignment, generate a time-series feature dataset, reducing time mismatches caused by different sampling links and improving the consistency between input data and actual operating states. A multi-scale temporal convolutional network architecture extracts current-carrying thermal shock features, medium thermal diffusion features, and soil heat storage hysteresis features through first, second, and third temporal convolutional layers, respectively, reducing prediction bias caused by the aliasing of thermal changes at different time scales. After the sheath outer surface temperature monitoring data and the initial prediction state vector form a prediction residual, the prediction residual is used to identify boundary condition correction coefficients. These coefficients are further scaled according to the interface mapping relationship to adjust the target channel weight matrix, enabling the network computation process to have online correction capabilities oriented towards the physical interface. Learning rate decay and convolutional kernel pruning suppress parameter oscillations and reduce computational load when the prediction residual does not decrease. Recursive prediction, exponentially weighted moving average correction, Grubbs outlier detection, and temperature threshold determination further enhance the continuity of the output of thermal field spatial distribution parameters and the reliability of the output of graded early warning signals. Attached Figure Description

[0039] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0040] Figure 1 is a flowchart of the method of the present invention.

[0041] Figure 2 is a flowchart of the time-series feature dataset generation process of the present invention.

[0042] Figure 3 is a flowchart of the multi-scale temporal convolutional network architecture of the present invention.

[0043] Figure 4 is a flowchart of the boundary condition correction coefficient identification process of the present invention.

[0044] Figure 5 is a flowchart of the network weight scaling adjustment process of the present invention.

[0045] Figure 6 is a flowchart of the learning rate decay and convolution kernel pruning process of this invention.

[0046] Figure 7 is a flowchart of the recursive prediction process of the present invention.

[0047] Figure 8 is a flowchart of the graded early warning signal output of the present invention.

[0048] Figure 9 is a flowchart of the virtual device module of the present invention. Detailed Implementation

[0049] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0050] The following is in conjunction with the appendix Figures 1-9 The embodiments of the present invention will be described in detail below.

[0051] Example 1: This example provides a dynamic inversion method for the thermal field distribution characteristics of cables laid in conduits. This method is executed by a dynamic inversion system for the thermal field distribution characteristics of cables laid in conduits. The system includes a data acquisition module, a model calculation module, a boundary update module, a prediction module, and an early warning module. The data acquisition module collects real-time operating parameters and environmental parameters along the cable laying path. The model calculation module performs multi-scale feature extraction on the time-series feature dataset. The boundary update module generates boundary condition correction coefficients based on the sheath outer surface temperature monitoring data. The prediction module outputs the spatial distribution parameters of the thermal field within the future prediction time domain. The early warning module outputs graded early warning signals based on the spatial distribution parameters of the thermal field.

[0052] The data acquisition module obtains real-time operating and environmental parameters of the cable in the conduit. Real-time operating parameters include cable conductor current and voltage, while environmental parameters include air temperature inside the conduit, soil temperature, and ambient humidity. Cable conductor current characterizes the Joule heat input intensity of the cable in the conduit; conductor voltage helps determine the cable's operating status; air temperature inside the conduit characterizes the heat transfer between the outer surface of the sheath and the air inside the conduit; and soil temperature and ambient humidity characterize the heat dissipation between the outer wall of the conduit and the surrounding soil. These parameters enter the data acquisition module from different sensor links. Without time synchronization, the cable conductor current in one sampling period might correspond to the soil temperature or ambient humidity in another sampling period. Subsequent inversion results would mistakenly interpret temperature deviations caused by time mismatches as changes in the heat transfer boundary. Therefore, the data acquisition module adds absolute timestamps to the real-time operating and environmental parameters and compares the absolute timestamps of different parameters. Data segments with timestamp deviations exceeding the synchronization deviation threshold are discarded, and data that meets the synchronization requirements are concatenated according to the sampling period to form a time-series feature dataset.

[0053] The temporal feature dataset is arranged according to historical time windows. Within a single sampling period, the temporal feature dataset includes at least the cable conductor current, conductor voltage, air temperature inside the conduit, soil temperature, and ambient humidity. Historical time windows are used to preserve hysteresis information regarding the thermal process of the cable in the conduit. Since the heat transfer from the cable conductor to the surrounding soil requires passing through the cable conductor, insulation layer, sheath, air inside the conduit, the outer wall of the conduit, and the surrounding soil, the influence of external environmental parameters on the predicted temperature of the interface between the outer surface of the sheath and the air inside the conduit, and the internal interface temperature, is not immediate but exhibits a certain time lag. The temporal feature dataset stores real-time operating parameters and environmental parameters in continuous sampling periods, enabling the subsequent multi-scale temporal convolutional network architecture to identify short-term current carrying capacity changes and long-term heat dissipation changes from the time series.

[0054] In this embodiment, before generating the time-series feature dataset, the data acquisition module performs low-pass filtering on the cable conductor current and conductor voltage, and removes abnormal fluctuations in the air temperature, soil temperature, and surrounding medium humidity inside the pipeline. Abnormal fluctuation removal is based on the magnitude of parameter changes in adjacent sampling periods. When the magnitude of change of a certain environmental parameter in a single sampling period exceeds the range achievable by the corresponding physical object, and adjacent sensors or adjacent sampling periods do not show the same trend, the data acquisition module marks the environmental parameter as an abnormal sampling value and performs interpolation correction using adjacent valid data. This processing is used to prevent instantaneous sensor interference from directly entering the time-series feature dataset.

[0055] The model computation module inputs the temporal feature dataset into a multi-scale temporal convolutional network architecture. This architecture comprises a first, second, and third temporal convolutional layer with increasing receptive fields. The first convolutional layer extracts current-carrying thermal shock features with a shorter receptive field; the second layer extracts medium thermal diffusion features with a medium receptive field; and the third layer extracts soil heat storage hysteresis features with a longer receptive field. Current-carrying thermal shock features primarily reflect the change in Joule heat input after rapid changes in cable conductor current; medium thermal diffusion features primarily reflect the thermal diffusion process between the insulation layer, sheath, and air inside the pipe; and soil heat storage hysteresis features primarily reflect the slow heat dissipation process between the outer wall of the pipe and the surrounding soil due to changes in soil temperature and the humidity of the surrounding medium.

[0056] In this embodiment, the first temporal convolutional layer performs one-dimensional convolution on the temporal feature dataset, outputting a first feature map. The first feature map preserves local variations related to short-time fluctuations in the cable conductor current. The second temporal convolutional layer performs one-dimensional convolution on the first feature map, outputting a second feature map. The second feature map preserves the medium-timescale variations formed during the diffusion of heat from the cable conductor through the insulation and sheath to the air inside the pipe. The third temporal convolutional layer performs one-dimensional convolution on the second feature map, outputting a third feature map. The third feature map preserves the long-period hysteresis variations formed during heat exchange between the outer wall of the pipe and the surrounding soil.

[0057] Because the kernel length and stride of the first, second, and third temporal convolutional layers increase sequentially, the temporal series lengths of the first, second, and third feature maps may differ. Before performing cross-layer fusion, the model computation module aligns the temporal dimensions of the first, second, and third feature maps. Using the temporal series length of the third feature map as the target fusion length, the module performs average pooling downsampling on the first and second feature maps. If any feature map is shorter than the target fusion length, linear interpolation or zero-padding is used; if any feature map is longer than the target fusion length, pooling or truncation is used to ensure that the first, second, and third feature maps have consistent temporal series lengths. After temporal alignment, the model computation module concatenates the first, second, and third feature maps along the channel dimension and inputs them into the residual connection unit for cross-layer fusion. The residual connection unit inputs the fused features into the fully connected mapping unit and outputs an initial prediction state vector.

[0058] In this embodiment, the initial predicted state vector includes the predicted temperatures of multiple physical interfaces. These multiple physical interfaces include at least the interface between the cable conductor and insulation layer, the interface between the insulation layer and the sheath, the interface between the outer surface of the sheath and the air inside the pipe, and the interface between the outer wall of the pipe and the surrounding soil. These physical interfaces correspond to the radial heat transfer path of the cable in the conduit. By setting the initial predicted state vector to the predicted temperatures of multiple physical interfaces, the model calculation module outputs not the temperature of a single measuring point, but an intermediate state that characterizes the spatial distribution of the thermal field of the cable in the conduit. This intermediate state provides the basis for subsequently calculating the predicted residuals based on the sheath outer surface temperature monitoring data and generating boundary condition correction coefficients.

[0059] In this embodiment, the typical hyperparameters of the multi-scale temporal convolutional network architecture can be set within the following ranges: the historical time window length is 30 to 240 sampling periods; the first convolutional kernel length is 3 to 7, the second convolutional kernel length is 5 to 15, and the third convolutional kernel length is 9 to 31; the first step length is 1, the second step length is 2, and the third step length is 4; the number of channels in the first, second, and third temporal convolutional layers is 16 to 128, respectively. The above hyperparameters are exemplary ranges, and those skilled in the art can adjust them according to the sampling period, the thermal inertia of the cable in the conduit, computational resources, and the prediction time domain length, but the increasing relationship of the receptive field of the first, second, and third temporal convolutional layers should be maintained.

[0060] The boundary update module acquires the temperature monitoring data of the outer surface of the sheath and compares it with the predicted temperature dimension of the interface between the outer surface of the sheath and the air inside the pipe in the initial predicted state vector to obtain the prediction residual. The temperature monitoring data of the outer surface of the sheath can be acquired by infrared temperature probes, distributed fiber optic temperature measurement devices, or contact temperature sensors placed at the temperature measurement location of the cable in the pipe. The technical meaning of the prediction residual is not ordinary training loss, but the deviation between the current heat transfer boundary state and the existing state of the multi-scale temporal convolutional network architecture. When the prediction residual is within the tolerance dead zone, the boundary update module maintains the current boundary condition correction coefficient; when the absolute value of the prediction residual exceeds the tolerance dead zone, the boundary update module starts the recursive least squares identification unit for online identification.

[0061] In this embodiment, the recursive least squares identification unit updates the current time-to-time parameter estimation vector using a regression vector formed by the time-series feature dataset and the initial predicted state vector, a measurement output formed by the predicted residuals, the parameter estimation vector from the previous time step, and the covariance matrix from the previous time step. The regression vector is a four-dimensional column vector, with its four elements, in dimensional order, representing the predicted temperature characteristics of the cable conductor-insulation interface, the insulation-sheath interface, the sheath outer surface-to-air interface, and the pipe outer wall-to-surrounding soil interface. These predicted temperature characteristics can be either normalized values ​​of the predicted temperatures at the corresponding physical interfaces or changes in the predicted temperatures relative to the previous sampling period. The current time-to-time parameter estimation vector is a four-dimensional parameter estimation vector, with its four dimensions corresponding one-to-one with the four dimensions of the regression vector.

[0062] The above recursive least squares method is updated according to the following formula:

[0063] ;

[0064] ;

[0065] ;

[0066] in: At the current sampling time, This is the regression vector at the current sampling time. for transpose, The measurement output is formed from the predicted residuals. This is the parameter estimation vector at the current time. This is the parameter estimation vector from the previous time step. Let be the covariance matrix at the current time. Let be the covariance matrix of the previous time step. For the gain vector, Forgetting factor; measurement output The normalized value of the prediction residual can be used, or the boundary correction target amount can be obtained by converting the prediction residual.

[0067] In this embodiment, the initial parameter estimation vector of the recursive least squares method It can be set as a four-dimensional unit vector, representing that the initial boundary condition correction coefficients of the four physical interfaces are all neutral proportions; the initial covariance matrix It can be set as a diagonal matrix with positive diagonal elements, used to represent the uncertainty of parameter estimation during cold start; forgetting factor. The forgetting factor can be set between 0.90 and 0.995. A larger forgetting factor is used when the environment is stable, and a smaller forgetting factor is used when the environment changes drastically. Those skilled in the art can adjust the above settings according to the sensor noise level, sampling period, and stability of the installation environment.

[0068] Once the parameter estimation vector at the current moment meets the convergence criterion, the boundary update module decomposes the current parameter estimation vector into multiple boundary condition correction coefficients in dimensional order. The convergence criterion can be set to the norm difference between the current parameter estimation vector and the parameter estimation vector at the previous moment being less than the parameter convergence threshold, and remaining less than the parameter convergence threshold for multiple consecutive sampling periods. The multiple boundary condition correction coefficients correspond to the interfaces of the cable conductor and insulation layer, the insulation layer and sheath, the outer surface of the sheath and the air inside the pipe, and the outer wall of the pipe and the surrounding soil, respectively. If an element in the parameter estimation vector exceeds the allowable range of the boundary condition correction coefficients, the boundary update module truncates the corresponding element to within the allowable range to ensure that the boundary condition correction coefficients remain a scaling factor usable for weight scaling. The allowable range of the boundary condition correction coefficients can be set from 0.5 to 1.5; when the sensor accuracy is high and the environment changes drastically, the allowable range of the boundary condition correction coefficients can be adjusted according to the historical residual distribution.

[0069] This setting addresses the issue that the deviation between the monitored temperature data of the outer surface of the sheath and the initial predicted state vector may originate from internal heat transfer lag, changes in air heat exchange within the pipe, or changes in the heat dissipation state of the surrounding soil. Adjusting the model using only a single overall correction factor would fail to differentiate the influence of different physical interfaces. This embodiment utilizes multiple boundary condition correction coefficients, each corresponding to a different physical interface, to distribute and correct the predicted residual along the radial heat transfer path of the cable in the conduit.

[0070] In this embodiment, the boundary update module records the prediction residual sequence in each sampling period. The prediction residual sequence is used to determine whether the temporal dynamic inversion model is in a state of continuous deviation. When the absolute value of the prediction residual does not decrease within a preset number of sampling periods, the model calculation module triggers learning rate decay and performs convolutional kernel pruning on the multi-scale temporal convolutional network architecture. Learning rate decay is used to reduce the backpropagation update amplitude and avoid excessive changes in network weights caused by short-term abnormal temperatures or instantaneous communication jitter. Learning rate decay can adopt a step decay method, where the current learning rate is multiplied by a decay coefficient less than 1 each time learning rate decay is triggered; the decay coefficient can be set to 0.7. Convolutional kernel pruning is used to identify one-dimensional convolutional kernels with low long-term contributions and to shield the multiply-add operations of the corresponding channels through an addressing mapping table. Learning rate decay and convolutional kernel pruning together form the online update path of the temporal dynamic inversion model.

[0071] After receiving the boundary condition correction coefficients, the model calculation module establishes an interface mapping relationship between the boundary condition correction coefficients and the target channel weight matrix in the multi-scale temporal convolutional network architecture. The boundary condition correction coefficients corresponding to the interface between the cable conductor and the insulation layer are used to adjust the target channel weight matrix in the first temporal convolutional layer, which characterizes the current-carrying thermal shock characteristics; the boundary condition correction coefficients corresponding to the interface between the insulation layer and the sheath are used to adjust the target channel weight matrix between the first and second temporal convolutional layers, which characterizes the thermal diffusion transition; the boundary condition correction coefficients corresponding to the interface between the outer surface of the sheath and the air inside the pipe are used to adjust the target channel weight matrix in the second temporal convolutional layer, which characterizes the thermal diffusion characteristics of the medium; and the boundary condition correction coefficients corresponding to the interface between the outer wall of the pipe and the surrounding soil are used to adjust the target channel weight matrix in the third temporal convolutional layer, which characterizes the soil heat storage hysteresis characteristics. The target channels can be determined using a target channel index table, which records the channel numbers of the one-dimensional convolutional kernels that need to be scaled according to the physical interface. When the multi-scale temporal convolutional network architecture uses channel grouping, the target channels are the channel groups pre-labeled with the corresponding physical interface. When the multi-scale temporal convolutional network architecture does not use channel grouping, the target channels are all the output channels within the corresponding convolutional layer used to extract the corresponding thermal features. Through this interface mapping relationship, the boundary condition correction coefficients are not treated as isolated outputs, but directly participate in the scaling adjustment of the network weights in the multi-scale temporal convolutional network architecture.

[0072] In this embodiment, the network weight scaling adjustment is completed using element-wise multiplication. The model calculation module uses the boundary condition correction coefficient as a scaling factor and multiplies it one by one with the elements in the corresponding target channel weight matrix to obtain the corrected target channel weight matrix. This processing allows changes in heat transfer boundaries to be incorporated into the model calculation process. When changes in the humidity of the surrounding medium cause a change in the heat dissipation state of the interface between the pipe outer wall and the surrounding soil, the corresponding boundary condition correction coefficient will preferentially adjust the target channel weight matrix in the third time-series convolutional layer that is related to the soil heat storage hysteresis characteristics; when a sudden change in current carrying capacity causes a significant change in the temperature of the interface between the cable conductor and the insulation layer, the corresponding boundary condition correction coefficient will preferentially adjust the target channel weight matrix in the first time-series convolutional layer that is related to the current carrying capacity thermal shock characteristics. This processing avoids uniformly distributing prediction residuals from different sources to all network weights.

[0073] The prediction module utilizes a time-domain dynamic inversion model to output the spatial distribution parameters of the thermal field within the future prediction time domain. The recursive prediction unit within the module maps the spatial distribution parameters of the thermal field output at the current prediction step to a feature dimension and appends them as prediction state features to the end of the time-series feature dataset. It then supplements the corresponding input channels with the most recent valid values, planned current-carrying values, or environmental prediction values ​​for the real-time operating parameters and environmental parameters corresponding to the prediction step. Finally, it removes the oldest data point to form an updated time-series feature dataset. The prediction module then inputs the updated time-series feature dataset back into the time-domain dynamic inversion model for the next prediction step calculation. Through continuous recursive prediction, the prediction module obtains the spatial distribution parameters of the thermal field covering the entire future prediction time domain. These spatial distribution parameters include the predicted temperatures at the interfaces of the cable conductor and insulation layer, the insulation layer and sheath, the outer surface of the sheath and the air inside the pipe, and the outer wall of the pipe and the surrounding soil at multiple prediction steps.

[0074] In this embodiment, the prediction module performs exponentially weighted moving average correction on the spatial distribution parameters of the thermal field output at each prediction step. The exponentially weighted moving average correction combines the spatial distribution parameters of the thermal field calculated at the current prediction step with the corrected spatial distribution parameters of the thermal field at corresponding future time points in the historical cache, according to weights, to obtain smoothed spatial distribution parameters of the thermal field. The smoothing coefficient of the exponentially weighted moving average correction can be set to 0.1 to 0.5; the larger the smoothing coefficient, the higher the weight of the spatial distribution parameters of the thermal field output at the current prediction step in the correction result. This process reduces glitches in multi-step recursive prediction, ensuring that the spatial distribution parameters of the thermal field received by the early warning module maintain a continuous evolution trend.

[0075] The early warning module performs Grubbs outlier detection on the spatial distribution parameters of the thermal field, obtaining the detection and judgment results. Grubbs outlier detection is used to identify abrupt changes in thermal field evolution that deviate from the overall temperature evolution trend within the future prediction time domain. The significance level of Grubbs outlier detection can be set to 0.05, or 0.01 depending on the operation and maintenance strategy. When the detection and judgment results indicate no outliers, the early warning module continues to monitor the spatial distribution parameters of the thermal field in the next sampling period. When the detection and judgment results indicate outliers, the outlier detection unit in the early warning module sends the detection and judgment results to the temperature threshold judgment unit. The temperature threshold judgment unit extracts the highest predicted temperature value from the spatial distribution parameters of the thermal field and compares the highest predicted temperature value with a first temperature threshold and a second temperature threshold. The first temperature threshold is lower than the second temperature threshold. When the detection and judgment results indicate an outlier and the highest predicted temperature value reaches the first temperature threshold, the early warning module outputs a first-level early warning signal; when the detection and judgment results indicate an outlier and the highest predicted temperature value reaches the second temperature threshold, the early warning module outputs a second-level early warning signal. The first-level warning signal is used to alert operation and maintenance personnel to the thermal condition of cables in conduits, while the second-level warning signal is used to trigger higher-level operation protection actions.

[0076] In this embodiment, the reference parameters under typical operating conditions can be set as follows: synchronization deviation threshold of 1 sampling period, tolerance dead zone of ±0.3℃, pruning statistical window of 100 sampling periods, continuous preset number of sampling periods for continuous non-decreasing judgment of 15 sampling periods, rollback threshold of 2℃, environmental judgment time window of 30 minutes, humidity change rate threshold of 5% / 10min, and temperature change amplitude threshold of 2℃ / 30min. Those skilled in the art can adjust the above parameters according to the voltage level of the cable in the conduit, sensor accuracy, sampling period, laying environment, and operation and maintenance strategy. The first temperature threshold and the second temperature threshold can be set according to the long-term allowable operating temperature and short-term allowable operating temperature of the cable insulation material.

[0077] Example 2: Based on Example 1, this example further explains the boundary condition correction, dual-channel redundant transmission, and convolution kernel pruning under drastic environmental changes.

[0078] In this embodiment, the data acquisition module continuously collects the humidity of the surrounding medium and the soil temperature. When the rate of change of the humidity of the surrounding medium within the environmental determination time window exceeds the humidity change rate threshold, or the magnitude of change of the soil temperature within the environmental determination time window exceeds the temperature change magnitude threshold, the model calculation module determines that the cable in the conduit is under a condition of drastic environmental change. Such conditions typically correspond to changes in soil moisture content after rainfall, changes in groundwater level, or changes in external heat dissipation caused by seasonal temperature differences. Under this condition, the rate of change in the heat transfer state at the interface between the outer wall of the pipe and the surrounding soil is higher than under normal conditions. If conventional forgetting factors and conventional recursive calculation data windows are still used, the boundary condition correction coefficient will lag behind the actual changes in heat transfer state.

[0079] Furthermore, under conditions of drastic environmental change, the model calculation module reduces the forgetting factor in the recursive least squares method and shortens the recursive calculation data window. After reducing the forgetting factor, newer prediction residuals have a higher weight in parameter identification; and after shortening the recursive calculation data window, the influence of data from earlier sampling periods that no longer represent the current soil condition on the current boundary condition correction coefficient is reduced. This processing allows the boundary condition correction coefficient corresponding to the interface between the pipe outer wall and the surrounding soil to more quickly follow changes in the humidity and soil temperature of the surrounding medium.

[0080] In this embodiment, the boundary condition correction under drastic environmental changes does not alter the main workflow of the multi-scale temporal convolutional network architecture. Instead, it prioritizes adjusting the target channel weight matrix related to the soil heat storage hysteresis characteristics. The model calculation module writes the boundary condition correction coefficients corresponding to the interface between the pipe outer wall and the surrounding soil into the target channel weight matrix corresponding to the third temporal convolutional layer, thereby enhancing the response of the third temporal convolutional layer to changes in the humidity of the surrounding medium and soil temperature. The overall dynamic inversion chain in Embodiment 1 is retained, while rapid correction is performed on the interface most prone to causing changes in external heat dissipation.

[0081] The data acquisition module implements dual-channel redundant transmission. The primary communication channel is used for regular data transmission, while the backup communication channel is used to resume data transmission in case of failure of the primary communication channel. The data acquisition module periodically polls the physical layer connection status of the primary communication channel. When a connection loss or packet loss rate exceeding the limit is detected on the primary communication channel, the data acquisition module wakes up the backup communication channel. During the switch from the primary to the backup communication channel, the data acquisition module caches any unsent time-series feature datasets in local storage. After the backup communication channel establishes a connection, the data acquisition module sends the cached time-series feature datasets in absolute timestamp order. This process ensures that the time-series feature datasets received by the model calculation module are continuous on the timeline, avoiding breaks in historical time windows caused by communication switching.

[0082] In this embodiment, when the model calculation module performs convolutional kernel pruning, it first calculates the cumulative absolute value of the weight matrix of each one-dimensional convolutional kernel. When the cumulative absolute value of the target one-dimensional convolutional kernel remains below the pruning threshold within the pruning statistics window, the model calculation module marks the target one-dimensional convolutional kernel as a redundant convolutional kernel. The model calculation module modifies the addressing mapping table, setting the memory base address pointers pointing to the input and output feature maps of the redundant convolutional kernels to null. During subsequent forward propagation, the model calculation module reads the addressing mapping table; if the corresponding memory base address pointer is null, it skips the multiply-accumulate operation of the corresponding channel of the redundant convolutional kernel.

[0083] The pruning threshold can be determined based on the distribution of the cumulative absolute values ​​of the weight matrices of each one-dimensional convolutional kernel within the pruning statistics window. For example, the cumulative absolute values ​​of all one-dimensional convolutional kernel weight matrices within the pruning statistics window can be sorted from smallest to largest, and the value located in the preset low quantile of the sorted result can be selected as the pruning threshold; alternatively, the cumulative absolute values ​​of the weight matrices of each one-dimensional convolutional kernel after the initial training can be used as the baseline value, and a preset proportion of the baseline value can be used as the pruning threshold for the corresponding one-dimensional convolutional kernel. Setting the pruning threshold in this way can avoid the imbalance in the pruning scale of convolutional kernels at different levels and channels caused by using a fixed threshold.

[0084] In this embodiment, convolutional kernel pruning is triggered by a sustained non-decreasing condition where the absolute value of the prediction residual does not decrease within a predetermined number of sampling periods, rather than being performed at fixed intervals. This avoids premature pruning while the model is still effectively converging and also reduces the number of low-contribution channels continuing to participate in computation when the model fails to reduce the prediction residual over a long period. After convolutional kernel pruning is completed, the freed computational resources are used to store longer prediction residual sequences or to perform boundary condition correction coefficient identification, thereby improving processing stability in edge computing environments.

[0085] Example 3: This example further explains the model cold start warm-up and model version rollback based on Examples 1 and 2.

[0086] In this embodiment, when the dynamic inversion system for the thermal field distribution characteristics of conduit cables is deployed for the first time or restarted after a power outage, the data acquisition module has not yet accumulated a sufficiently long time-series feature dataset. If the recursive least squares method is directly executed at this time, the regression vector length will be insufficient, and the boundary condition correction coefficients will easily fluctuate. To solve the above problems, the model calculation module performs model preheating. Model preheating includes loading pre-trained general model parameters and local area fine-tuning. The pre-trained general model parameters are obtained from historical operating data of similar conduit cables. During training, the measured temperature of each physical interface, the offline simulation temperature, or the temperature calculated by the verified thermal path model are used as supervision labels. The mean squared error loss function and gradient descent optimizer are used to achieve convergence, and these convergences are used as the initial weight matrix of the multi-scale temporal convolutional network architecture. Local area fine-tuning uses the local area time-series feature dataset collected after system startup.

[0087] During local fine-tuning, the model calculation module keeps the weight parameters of the first and second temporal convolutional layers fixed, updating only the weight parameters of the third temporal convolutional layer. This setting is based on the fact that the current-carrying thermal shock characteristics and the medium thermal diffusion characteristics are mainly determined by the cable structure, insulation layer structure, sheath structure, and pipe structure, and the differences are small among similar types of conduit cables; the soil heat storage hysteresis characteristics are related to the local soil temperature, the humidity of the surrounding medium, and the external heat dissipation state of the pipe. Prioritizing the fine-tuning of the third temporal convolutional layer during the cold start phase helps to reduce the initial deviation caused by local environmental differences.

[0088] In this embodiment, the model computation module includes a model version management unit. Whenever the learning rate decays, convolutional kernels are pruned, boundary condition correction coefficients are updated, or model warm-up is complete, the model version management unit records the network weights, bias parameters, boundary condition correction coefficients, triggering reasons, and absolute timestamps of the current multi-scale temporal convolutional network architecture, forming a parameter snapshot file. The parameter snapshot file is written to non-volatile storage. The parameter snapshot file is used to restore the previous stable state when the current model parameters enter an unstable state.

[0089] When the prediction residuals consistently exceed the rollback threshold within the rollback decision time window, and the boundary condition correction coefficients fail to meet the convergence criteria, the model version management unit determines that the current model parameters are unsuitable for continued use in predicting the spatial distribution parameters of the thermal field. The model version management unit reads the parameter snapshot file from the previous stable state, loads the network weights, bias parameters, and boundary condition correction coefficients from the previous stable state into the model computation module, overwriting the current model parameters. After overwriting, the model computation module re-executes forward propagation based on the current time-series feature dataset, and the boundary update module recalculates the prediction residuals. This rollback process avoids the long-term impact of abnormal sampling values, communication jitter, or locally unstable updates on subsequent predictions of the spatial distribution parameters of the thermal field.

[0090] To facilitate a better understanding of the present invention by those skilled in the art, the present invention will be further illustrated below with reference to a specific application scenario.

[0091] The proposed method, the fixed boundary condition transient thermal circuit method, the single-scale temporal convolution method, and the multi-scale temporal convolution method without boundary condition correction were compared and recalculated using the same operating record of the cable in the conduit. The recalculation used the same monitoring data on cable conductor current, conductor voltage, air temperature inside the conduit, soil temperature, ambient medium humidity, and outer surface temperature of the sheath. The future prediction time domain length was set to 60 minutes. The fixed boundary condition transient thermal circuit method does not use prediction residuals to correct boundary conditions; the single-scale temporal convolution method does not distinguish between current-carrying thermal shock characteristics, medium thermal diffusion characteristics, and soil heat storage hysteresis characteristics; the multi-scale temporal convolution method without boundary condition correction retains multi-scale feature extraction but does not convert prediction residuals into boundary condition correction coefficients.

[0092] Table 1 is a comparison table of processing results.

[0093] Table 1:

[0094]

[0095] As shown in Table 1, under normal current-carrying conditions, all processing methods can output temperature prediction results. However, the method of this invention has a lower mean absolute deviation under normal current-carrying conditions because the real-time operating parameters and environmental parameters are synchronously corrected and aligned with the timestamp, and the multi-scale temporal convolutional network architecture can distinguish thermal changes at different time scales. Under short-term overload conditions, the fixed boundary condition transient thermal path method cannot reflect the current-carrying thermal shock in a timely manner, the single-scale temporal convolutional method easily mixes short-term current-carrying changes with slow environmental changes, and the multi-scale temporal convolutional method without boundary condition correction can extract multi-scale features, but cannot correct the heat transfer boundary through prediction residuals. The method of this invention extracts the current-carrying thermal shock features through the first temporal convolutional layer and uses the prediction residuals to generate boundary condition correction coefficients, so that the spatial distribution parameters of the thermal field after short-term overload can be updated with the feedback of the outer surface temperature monitoring data of the sheath.

[0096] Under conditions of rising ambient humidity, the deviations of the fixed boundary condition transient thermal path method and the multi-scale temporal convolution method without boundary condition correction mainly stem from the failure to promptly correct the heat dissipation state at the interface between the pipe outer wall and the surrounding soil. The single-scale temporal convolution method also suffers from insufficient extraction of soil heat storage hysteresis features. This invention addresses this issue by reducing the forgetting factor in the recursive least squares method and shortening the recursive calculation data window when the ambient humidity change rate exceeds a threshold, enabling faster updates to the boundary condition correction coefficients at the interface between the pipe outer wall and the surrounding soil. The updated boundary condition correction coefficients further scale the target channel weight matrix corresponding to the soil heat storage hysteresis features in the third temporal convolution layer, ensuring that the spatial distribution parameters of the thermal field are updated according to changes in external heat dissipation conditions.

[0097] In summary, the technical effect of this invention is not achieved by a single neural network prediction, but rather by a continuous chain of technologies: synchronous correction and timestamp alignment of time-series feature datasets, multi-scale time-series feature extraction, formation of prediction residuals from sheath outer surface temperature monitoring data, identification of boundary condition correction coefficients from prediction residuals, scaling of network weights corresponding to physical interfaces using boundary condition correction coefficients, online updating of the time-domain dynamic inversion model, prediction of thermal field spatial distribution parameters, and output of graded early warning signals. The results of each step in this chain serve as inputs or triggers for subsequent steps, ensuring clear data sources and reproducible processes.

[0098] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0099] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for dynamic inversion of the thermal field distribution characteristics of cables in conduits, characterized in that, include: The real-time operating parameters and environmental parameters of the cable in the conduit are obtained, and the real-time operating parameters and environmental parameters are synchronously corrected and timestamp aligned to generate a time-series feature dataset. The temporal feature dataset is input into a multi-scale temporal convolutional network architecture. The current-carrying thermal shock features, medium thermal diffusion features, and soil heat storage hysteresis features in the temporal feature dataset are extracted through the first, second, and third temporal convolutional layers with increasing receptive fields, and the initial predicted state vector is output. Acquire the temperature monitoring data of the outer surface of the sheath, calculate the prediction residual between the temperature monitoring data of the outer surface of the sheath and the predicted temperature dimension of the interface between the outer surface of the sheath and the air inside the pipe in the initial prediction state vector, and identify the boundary condition correction coefficient based on the prediction residual. The network weights corresponding to the physical interface in the multi-scale temporal convolutional network architecture are scaled and adjusted using the boundary condition correction coefficients. When the absolute value of the prediction residual does not decrease for a continuous number of consecutive sampling periods, the learning rate decay and convolutional kernel pruning are performed to obtain the temporal dynamic inversion model. The time-domain dynamic inversion model is used to output the spatial distribution parameters of the thermal field within the future predicted time domain. Outlier detection and temperature threshold determination are performed on the spatial distribution parameters of the thermal field, and a graded early warning signal is output.

2. The method for dynamic inversion of the thermal field distribution characteristics of conduit cables according to claim 1, characterized in that, Generating the time-series feature dataset includes: collecting cable conductor current and conductor voltage to form the real-time operating parameters; collecting air temperature, soil temperature and ambient medium humidity in the pipeline to form the environmental parameters; adding absolute timestamps to the real-time operating parameters and the environmental parameters, removing data segments whose timestamp deviation exceeds the synchronization deviation threshold, and splicing the aligned parameters according to the sampling period to form the time-series feature dataset covering the historical time window.

3. The method for dynamic inversion of the thermal field distribution characteristics of conduit cables according to claim 1, characterized in that, The first temporal convolutional layer outputs a first feature map with a first convolutional kernel length and a first step length. The second temporal convolutional layer outputs a second feature map with a second convolutional kernel length greater than the first convolutional kernel length and a second step length greater than the first step length. The third temporal convolutional layer outputs a third feature map with a third convolutional kernel length greater than the second convolutional kernel length and a third step length greater than the second step length. The third feature map is mapped through a fully connected layer to generate the initial prediction state vector containing the predicted temperatures of multiple physical interfaces.

4. The method for dynamic inversion of the thermal field distribution characteristics of conduit cables according to claim 3, characterized in that, The first feature map, the second feature map, and the third feature map are aligned in the time dimension to make their time series lengths consistent. The time-aligned first feature map, second feature map, and third feature map are then concatenated in the channel dimension and input into the residual connection unit for cross-layer fusion. The fused features are used to generate the initial prediction state vector.

5. The method for dynamic inversion of the thermal field distribution characteristics of conduit cables according to claim 1, characterized in that, The boundary condition correction coefficients obtained based on the predicted residuals include: When the absolute value of the prediction residual exceeds the tolerance dead zone, the regression vector formed by the time series feature dataset and the initial prediction state vector, the measurement output formed by the prediction residual, the parameter estimation vector of the previous time step, and the covariance matrix of the previous time step are extracted, and the recursive least squares method is used to update the parameter estimation vector of the current time step. The regression vector includes the predicted temperature characteristics of the interface between the cable conductor and the insulation layer, the predicted temperature characteristics of the interface between the insulation layer and the sheath, the predicted temperature characteristics of the interface between the outer surface of the sheath and the air inside the pipe, and the predicted temperature characteristics of the interface between the outer wall of the pipe and the surrounding soil, arranged in dimensional order. The parameter estimation vector at the current moment is a four-dimensional parameter estimation vector corresponding to the dimension of the regression vector; The converged parameter estimation vector at the current moment is decomposed in dimensional order and subjected to boundary condition correction coefficient amplitude limiting processing to obtain multiple boundary condition correction coefficients corresponding to the interface between the cable conductor and the insulation layer, the interface between the insulation layer and the sheath, the interface between the outer surface of the sheath and the air inside the pipe, and the interface between the outer wall of the pipe and the surrounding soil.

6. The method for dynamic inversion of the thermal field distribution characteristics of conduit cables according to claim 5, characterized in that, When the rate of change of the ambient medium humidity within the environmental determination time window exceeds the humidity change rate threshold, or when the magnitude of change of the soil temperature within the environmental determination time window exceeds the temperature change magnitude threshold, the forgetting factor in the recursive least squares method is reduced, and the recursive calculation data window is shortened.

7. The method for dynamic inversion of the thermal field distribution characteristics of conduit cables according to claim 1, characterized in that, The scaling adjustment of network weights corresponding to physical interfaces in the multi-scale temporal convolutional network architecture using the boundary condition correction coefficients includes: establishing an interface mapping relationship between the boundary condition correction coefficients and the target channel weight matrices in the first, second, and third temporal convolutional layers; and performing element-wise multiplication of the boundary condition correction coefficients as scaling factors with the target channel weight matrices according to the interface mapping relationship.

8. The method for dynamic inversion of the thermal field distribution characteristics of conduit cables according to claim 1, characterized in that, Performing the convolutional kernel pruning includes: calculating the cumulative absolute value of the weight matrix of the one-dimensional convolutional kernel in the multi-scale temporal convolutional network architecture; when the cumulative absolute value of the target one-dimensional convolutional kernel is continuously lower than the pruning threshold within the pruning statistics window, the target one-dimensional convolutional kernel is marked as a redundant convolutional kernel; setting the memory base address pointers pointing to the input feature map and output feature map of the redundant convolutional kernel in the addressing mapping table to null, and skipping the multiply-accumulate operation of the corresponding channel of the redundant convolutional kernel in subsequent forward propagation.

9. The method for dynamic inversion of the thermal field distribution characteristics of conduit cables according to claim 1, characterized in that, The process of outputting the spatial distribution parameters of the thermal field within the predicted time domain includes: mapping the spatial distribution parameters of the thermal field output at the current prediction step size to the feature dimension and then concatenating them as predicted state features to the end of the time-series feature dataset; supplementing the corresponding prediction step size's real-time operating parameters and environmental parameters with the most recent valid values, planned current carrying values, or environmental prediction values ​​to the corresponding input channels; removing the oldest time data to form an updated time-series feature dataset; inputting the updated time-series feature dataset back into the time-domain dynamic inversion model for the next prediction step size calculation; and performing exponentially weighted moving average correction on the spatial distribution parameters of the thermal field output at each prediction step size.

10. A dynamic inversion system for the thermal field distribution characteristics of cables in conduits, used to implement the dynamic inversion method for the thermal field distribution characteristics of cables in conduits as described in any one of claims 1 to 9, characterized in that, It includes a data acquisition module, a model calculation module, a boundary update module, a prediction module, and an early warning module; The data acquisition module is used to acquire the real-time operating parameters and environmental parameters of the cable in the conduit, and to perform synchronous correction and timestamp alignment on the real-time operating parameters and environmental parameters to generate a time-series feature dataset. The model calculation module is used to input the temporal feature dataset into a multi-scale temporal convolutional network architecture, and extract the current-carrying thermal shock features, medium thermal diffusion features, and soil heat storage hysteresis features from the temporal feature dataset through the first, second, and third temporal convolutional layers with increasing receptive fields, and output the initial predicted state vector. The boundary update module is used to acquire the temperature monitoring data of the outer surface of the sheath, calculate the prediction residual between the temperature monitoring data of the outer surface of the sheath and the predicted temperature dimension of the interface between the outer surface of the sheath and the air inside the pipe in the initial prediction state vector, identify the boundary condition correction coefficient based on the prediction residual, and use the boundary condition correction coefficient to scale and adjust the network weights of the corresponding physical interface in the multi-scale temporal convolutional network architecture. The model calculation module is also used to perform learning rate decay and convolution kernel pruning when the absolute value of the predicted residual does not decrease for a continuous number of consecutive sampling periods, so as to obtain a time-domain dynamic inversion model. The prediction module is used to output the spatial distribution parameters of the thermal field within the future predicted time domain using the time-domain dynamic inversion model. The early warning module is used to perform outlier detection and temperature threshold determination on the spatial distribution parameters of the thermal field, and output graded early warning signals.

Images