A time series data analysis method for metering box component aging

Abstract

The present application relates to the technical field of electrical aging, in particular to a time sequence data analysis method for metering box component aging; time sequence data of electrical load, local temperature, surface leakage current and external weather of multiple space nodes in the metering box is obtained; based on the physical topology position of the space node, the electrical load time sequence data and the local temperature time sequence data are constructed into a multi-node three-dimensional space-time tensor; the multi-node three-dimensional space-time tensor is input into a graph attention network, a dynamic heat transfer weight matrix is identified, and the mixed heat of the space node is decoupled into an independent heating sequence and a heat cross-talk sequence; the independent heating sequence and the heat cross-talk sequence are used as a composite heat source excitation to reconstruct the solid temperature sequence of the insulating surface of each space node; based on the dynamic dew point, the surface condensation time sequence is obtained in combination with the solid temperature sequence; the surface condensation time sequence and the surface leakage current time sequence are fused and mapped to identify the insulation aging cumulative index, and a dynamic peak shaving and load shedding instruction is sent.

Description

Technical Field

[0001] This invention relates to the field of electrical aging technology, specifically to a time-series data analysis method for the aging of metering box components. Background Technology

[0002] As a key device for electricity metering and distribution control, the metering box is installed outdoors for extended periods. It houses various components, including terminals, insulating partitions, and metering elements. During long-term operation, each space within the box continuously bears the thermal stress caused by the load current. Simultaneously, the diurnal temperature variation and changes in external air pressure drive humid air to periodically enter the box, repeatedly condensing and evaporating on the insulating surfaces of the components, resulting in continuous moisture erosion.

[0003] The long-term alternating coupling of thermal stress and wet stress causes the insulation material to gradually deteriorate and age, eventually inducing leakage, flashover, and even power outage accidents, seriously threatening the accuracy of metering and the safety of power supply.

[0004] To monitor the aging status of components, existing technologies typically involve placing temperature and leakage current measuring points inside the metering chamber and performing threshold comparisons or trend analyses on the collected time-series data. However, the internal space of the metering chamber is compact and the nodes are densely packed. The measured temperature rise of any node is a mixture of its own heating and thermal crosstalk from nearby heavily loaded nodes, and single-point temperature data cannot reflect the true thermal state of that node. At the same time, the condensation process on the insulating surface is hidden and brief. Conventional analysis cannot grasp the surface solid temperature or the dynamic dew point, making it difficult to determine the true duration of water film precipitation and residence, resulting in wet stress remaining in the monitoring blind zone for a long time.

[0005] Existing time-series data analysis methods cannot decouple the independent heating and thermal crosstalk of each spatial node from the aliased monitoring data, nor can they quantify the degree of aging accumulation under the dual stress coupling of condensation and leakage activities. This leads to distorted insulation aging assessment, inability to locate the source of degradation, and difficulty in early intervention.

[0006] To address this, a time-series data analysis method for metering box component aging is proposed. Summary of the Invention

[0007] The purpose of this invention is to provide a time-series data analysis method for metering box component aging, which integrates and maps the surface condensation duration sequence with the surface leakage current time-series data to identify the insulation aging cumulative index and send dynamic peak shaving and load reduction commands.

[0008] To achieve the above objectives, the present invention provides the following technical solution: A time-series data analysis method for metering box component aging includes: Acquire electrical load time-series data, local temperature time-series data, surface leakage current time-series data, and external meteorological time-series data from multiple spatial nodes inside the metering box; based on the physical topological location of the spatial nodes, perform spatial mapping and temporal alignment of the electrical load time-series data and the local temperature time-series data to construct a multi-node three-dimensional spatiotemporal tensor; The multi-node three-dimensional spatiotemporal tensor is input into a graph attention network with thermophysical prior constraints to identify the dynamic heat transfer weight matrix and decouple the mixed heat of spatial nodes into independent heat generation sequences and thermal crosstalk sequences. Independent heating sequences and thermal crosstalk sequences are used as composite heat source excitations and input into the physical thermal inertia extrapolation model to reconstruct the solid temperature sequence of the insulating surface of each spatial node; dynamic dew point is calculated based on external meteorological time series data, and then the surface condensation duration sequence is obtained by combining the solid temperature sequence. The surface condensation duration sequence and surface leakage current time series data are fused and mapped to identify the insulation aging cumulative index of each spatial node under dual stress coupling. When the cumulative insulation aging index exceeds the safety threshold, the overloaded heating node is traced based on the thermal crosstalk sequence, and a dynamic peak-shaving and load-reduction command is sent.

[0009] Preferably, the steps for obtaining time-series data of multiple spatial nodes inside the metering box specifically include: extracting the effective value of the operating current and the value of the active power within a continuous time period of the metering box to form the electrical load time-series data; capturing the temperature data inside the metering box to form the local temperature time-series data; obtaining the creepage fluctuation signal of the insulating components as the surface leakage current time-series data; and obtaining the temperature and humidity change data of the physical environment in which the box is located as the external meteorological time-series data.

[0010] Preferably, the steps for constructing a multi-node three-dimensional spatiotemporal tensor specifically include: Extract the asynchronous timestamp deviation caused by communication delay, use the dynamic time warping algorithm to find the optimal matching path between the peak of the electrical load time series data and the peak of the local temperature time series data, elastically stretch and compress the time axis, and resample and output the time dimension synchronized alignment feature sequence. Establish a two-dimensional spatial coordinate axis consistent with the internal hardware layout of the metering box, and map the alignment feature sequence into the corresponding two-dimensional spatial matrix according to the horizontal and vertical relative positions of each spatial node in the box. Along the advancing direction of the time sliding window, multiple consecutive two-dimensional spatial matrices containing time attributes and spatial coordinate attributes are stacked and combined to generate the multi-node three-dimensional spatiotemporal tensor.

[0011] Preferably, the recognition process in the graph attention network by inputting the three-dimensional spatiotemporal tensor includes: In the graph attention network, each spatial node is defined as a graph node, and the physical distance between nodes and the airflow path are defined as graph edges. A physical penalty function containing energy conservation rules and thermodynamic constraints is constructed to restrict the direction of heat conduction between adjacent nodes to only be from the high temperature region to the low temperature region. Under the constraint of the physical penalty function, self-attention operation is performed to extract the hidden state features of nodes, and a dynamic heat transfer weight matrix that conforms to the real physical heat transfer logic is output. The dynamic heat transfer weight matrix is ​​used to perform algebraic separation calculation on the overall temperature rise of the spatial nodes, and outputs the independent heating sequence generated by the current flowing through the node itself, as well as the thermal crosstalk sequence generated by the superposition of radiation and conduction from the surrounding heavily loaded nodes.

[0012] Preferably, the process of reconstructing the solid temperature sequence of the insulating surface of each spatial node includes: A physical thermal inertia extrapolation model is constructed using a neural network of ordinary differential equations; the specific heat capacity and mass parameters of the insulating material are input into a multilayer perceptron network to establish a temperature change rate derivative prediction layer that characterizes the temperature rise hysteresis of solids. The independent heating sequence and the thermal crosstalk sequence are superimposed and combined into a composite heat source excitation, and the composite heat source excitation and the node air temperature are synchronously input into the temperature change rate derivative prediction layer. The ordinary differential equation solver is called to perform continuous integration calculation on the time step according to the temperature change rate gradient output by the network, filter out the high-frequency fluctuation noise in the node air temperature, and invert the output to output the real state trajectory of the insulating surface with physical smooth response and phase hysteresis characteristics, which is denoted as the solid temperature sequence.

[0013] Preferably, the process of obtaining the surface condensation duration sequence includes: Extract temperature and humidity data from the external meteorological time series data, calculate the dynamic critical condensation boundary curve under the interaction of cold air outside the box and hot air inside the box, and output the dynamic dew point; The reconstructed solid temperature sequence and the dynamic dew point are cross-referenced in the time domain. The intersection critical time point where the solid temperature value decreases and is lower than the dynamic dew point value is extracted as the starting point of liquid water film precipitation, and the intersection critical time point where the solid temperature value rises and is higher than the dynamic dew point value is extracted as the end point of water film evaporation. By accumulating the time difference between the precipitation start point and the evaporation end point, a surface condensation duration sequence is output to characterize the duration of continuous moisture immersion on the surface of the insulating material.

[0014] Preferably, the step of identifying the cumulative insulation aging index of each spatial node under dual stress coupling specifically includes: A multimodal cross-attention deep network is constructed as the core mapping analysis model; the surface condensation duration sequence is input into the time-domain gating layer for activation function operation to generate a mask matrix representing the probability of water adhesion; the surface leakage current time series data is converted into feature query vector and feature key value vector, and the mask matrix is ​​used as the guiding parameter of the attention mechanism to perform fusion calculation, and the leakage current peak feature tensor that occurs during water retention is extracted and amplified; The leakage current peak feature tensor is input into a deep Weibull regression layer, and the scale and shape parameters in the Weibull distribution model are output by nonlinear mapping. The cumulative degradation value of the survival function is calculated based on the scale and shape parameters, and the output is the insulation aging cumulative index.

[0015] Preferably, the process of sending dynamic peak shaving and load reduction commands includes: A global monitoring and comparison matrix is ​​constructed. When the cumulative insulation aging index of a spatial node exceeds the safety threshold, the historical thermal crosstalk sequence corresponding to the spatial node is retrieved. By searching the spatial coordinates of the source node with the largest heat contribution value and the longest duration in the historical thermal crosstalk sequence, the source device that generates the influence of nearby thermal radiation is reversely locked as the heavy-load heat-generating node. A dynamic peak-shaving and load-reducing command is generated to reduce the upper limit of the operating power of the heavy-load heat-generating node.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention, based on the physical topological location of spatial nodes, spatially maps and aligns electrical load time-series data with local temperature time-series data to construct a multi-node three-dimensional spatiotemporal tensor, enabling time-series analysis to possess complete temporal and spatial coordinate attributes. Subsequently, this tensor is input into a graph attention network with thermophysical prior constraints. By leveraging energy conservation rules and thermodynamic constraints, a physical penalty function is constructed to limit the direction of heat conduction to only from high-temperature regions to low-temperature regions. This identifies a dynamic heat transfer weight matrix that conforms to the actual heat transfer logic, and based on this, the mixed temperature rise algebra of each node is separated into independent heating sequences and thermal crosstalk sequences. This significantly improves the physical reliability of time-series data analysis and the accuracy of node-level evaluation.

[0017] 2. This invention addresses the problem of persistent monitoring blind spots for wet stress and the inability to quantify the degree of aging accumulation. It superimposes independent heating sequences and thermal crosstalk sequences into a composite heat source excitation, inputs it into a physical thermal inertia model, and reconstructs a solid temperature sequence on the insulating surface with characteristics of temperature rise lag and phase hysteresis. Then, based on external meteorological time-series data, it calculates the dynamic dew point and cumulatively outputs a surface condensation duration sequence, transforming the hidden and transient condensation process into a measurable time-series indicator. Furthermore, it generates a moisture adhesion mask based on the condensation duration, guiding an attention mechanism to capture and amplify the leakage current peak characteristics during moisture retention. Through deep Weibull regression, it outputs an insulation aging accumulation index, achieving unified quantification of the dual stress coupling effects of condensation wetting and leakage current activity.

[0018] 3. This invention retrieves the spatial coordinates of the source node with the largest and longest-lasting heat contribution, reversely pinpointing the heavily loaded heat-generating node that continuously exerts the influence of neighboring thermal radiation, and generates a dynamic peak-shaving and load-reduction command to reduce its operating power limit. This design integrates aging analysis results with thermal crosstalk tracing capabilities, allowing mitigation measures to directly target the source of damage rather than the affected node, avoiding power loss caused by blindly reducing load or shutting down the entire metering box. Simultaneously, by reducing the power of heavily loaded nodes in advance, it reduces thermal stress input from the source, delaying the insulation aging process, and achieving closed-loop protection from state perception and source location to proactive intervention, significantly improving the safety of metering box operation and the targeted nature of maintenance decisions. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating a time-series data analysis method for the aging of metering box components according to the present invention. Figure 2 This is a schematic diagram of the process for decoupling hybrid heat according to the present invention; Figure 3 This is a flowchart illustrating the process of obtaining the cumulative insulation aging index according to the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1: This invention proposes a time-series data analysis method for the aging of metering box components. The flow of the method is as follows: Figure 1 As shown, the process of decoupling mixed heat is as follows: Figure 2 As shown, the insulation aging cumulative index is obtained as follows: Figure 3 As shown, it specifically includes: Acquire electrical load time-series data, local temperature time-series data, surface leakage current time-series data, and external meteorological time-series data from multiple spatial nodes inside the metering box; based on the physical topological location of the spatial nodes, perform spatial mapping and temporal alignment of the electrical load time-series data and the local temperature time-series data to construct a multi-node three-dimensional spatiotemporal tensor; The multi-node three-dimensional spatiotemporal tensor is input into a graph attention network with thermophysical prior constraints to identify the dynamic heat transfer weight matrix and decouple the mixed heat of spatial nodes into independent heat generation sequences and thermal crosstalk sequences. Independent heating sequences and thermal crosstalk sequences are used as composite heat source excitations and input into the physical thermal inertia extrapolation model to reconstruct the solid temperature sequence of the insulating surface of each spatial node; dynamic dew point is calculated based on external meteorological time series data, and then the surface condensation duration sequence is obtained by combining the solid temperature sequence. The surface condensation duration sequence and surface leakage current time series data are fused and mapped to identify the insulation aging cumulative index of each spatial node under dual stress coupling. When the cumulative insulation aging index exceeds the safety threshold, the overloaded heating node is traced based on the thermal crosstalk sequence, and a dynamic peak-shaving and load-reduction command is sent.

[0022] The specific steps for obtaining time-series data of multiple spatial nodes inside the metering box include: The effective value of the operating current and the active power value of the metering box within a continuous time period are extracted to form the electrical load time series data; the temperature data inside the metering box is captured to form the local temperature time series data; the creepage fluctuation signal of the insulating components is obtained as the surface leakage current time series data; and the temperature and humidity change data of the physical environment in which the box is located are obtained as the external meteorological time series data.

[0023] The acquisition of electrical load time-series data is accomplished using current transformers and energy metering units installed in the incoming circuit of the metering box. The current transformers continuously sample the operating current according to a preset sampling period, and the effective value of the operating current is obtained through RMS calculation; the active power value is obtained from the synchronously sampled voltage and current values ​​through power calculation. Arranging the above effective current values ​​and active power values ​​in chronological order constitutes the electrical load time-series data reflecting the power supply load intensity of the node.

[0024] The acquisition of local temperature time-series data is accomplished using temperature sensors deployed near each spatial node. The temperature sensors sense the temperature of the air medium surrounding the node and output temperature values ​​at a sampling period consistent with the electrical load. Arranging these values ​​chronologically constitutes the local temperature time-series data. It should be noted that this temperature reflects the thermal state of the air medium and is subject to fluctuations due to air convection; it is not the true temperature of the insulating solid surface. This difference is precisely what the subsequent thermal inertia extrapolation stage will address. This embodiment abandons the direct temperature measurement scheme of the attached solid surface; direct temperature measurement alters the heat capacity distribution and condensation physical field of the original insulating surface, resulting in a measured temperature that is actually the temperature of the sensor casing, not the true insulating surface temperature, leading to significant errors. This invention, based on non-contact, safe air-state measurement data, filters out air convection noise and compensates for solid thermal inertia phase shifts through a physical extrapolation model, achieving high-precision calculation of the true thermal state of a large-area solid insulating surface while ensuring electrical insulation safety.

[0025] The acquisition of surface leakage current time-series data is accomplished using a high-frequency miniature sensor attached to the surface of the insulating component. The creepage fluctuation signal refers to the microampere-level high-frequency discharge current generated when tiny conductive channels are formed on the surface of the insulating material due to the adhesion of dirt and moisture. Because this discharge current has an extremely small amplitude and a high frequency, a high-frequency miniature sensor is used to ensure the ability to capture rapidly changing, weak signals. Arranging the acquired creepage fluctuation signals in chronological order constitutes the surface leakage current time-series data.

[0026] The collection of external meteorological time series data is accomplished by using temperature and humidity sensors placed on the outside of the enclosure. The temperature and relative humidity of the environment in which the enclosure is located are collected and arranged in chronological order to form external meteorological time series data.

[0027] This invention constructs a complete and multi-dimensional underlying data foundation for the overall method by simultaneously acquiring four types of time-series data: electrical load, local temperature, surface leakage current, and external meteorological data. Specifically, electrical load data characterizes the electrical root causes of node heating, local temperature data characterizes the thermal environment surrounding the node, surface leakage current data characterizes the electrical degradation signs of insulation, and external meteorological data characterizes the environmental conditions that induce condensation. These four types of data correspond to the thermal field, electric field, and environmental field, respectively, covering the main physical factors affecting insulation aging.

[0028] Compared to traditional monitoring methods that rely solely on single-point temperature thresholds, the multi-source heterogeneous data collected in this scheme provides the necessary prerequisites for subsequent thermal field decoupling, condensation simulation, and electrochemical dual-stress fusion damage assessment. This allows the analysis of insulation aging to be upgraded from a simple judgment of a single physical quantity to a refined simulation based on the synergy of multiple physical fields. Simultaneously, the use of high-frequency miniature sensors to collect creepage fluctuation signals improves the sensitivity to capturing weak early discharge signs, facilitating the early detection of insulation degradation. The unified clock reference acquisition method also lays the foundation for the accurate alignment and fusion of subsequent multi-source data, ensuring the overall accuracy and reliability of the analysis results.

[0029] It should be noted that there are two different types of time misalignment between changes in electrical load and local temperature response. The first is the asynchronous time stamp deviation caused by the communication link, which stems from the fact that the electrical load sensor and the temperature sensor belong to different acquisition circuits and have inconsistent data upload delays. This deviation is an acquisition error that should be eliminated. The second is the inherent thermal response hysteresis caused by load changes, which stems from the thermal capacity inertia of solids and air. This is a real physical process characteristic and is the phase hysteresis feature that needs to be characterized in the subsequent physical thermal inertia deduction.

[0030] Preferably, the steps for constructing a multi-node three-dimensional spatiotemporal tensor specifically include: Extract the asynchronous timestamp deviation caused by communication delay, use the dynamic time warping algorithm to find the optimal matching path between the peak of the electrical load time series data and the peak of the local temperature time series data, elastically stretch and compress the time axis, and resample and output the time dimension synchronized alignment feature sequence. Establish a two-dimensional spatial coordinate axis consistent with the internal hardware layout of the metering box, and map the alignment feature sequence into the corresponding two-dimensional spatial matrix according to the horizontal and vertical relative positions of each spatial node in the box. Along the advancing direction of the time sliding window, multiple consecutive two-dimensional spatial matrices containing time attributes and spatial coordinate attributes are stacked and combined to generate the multi-node three-dimensional spatiotemporal tensor.

[0031] Because the electrical load sensor and the local temperature sensor belong to different acquisition circuits, their data experience transmission delays when uploaded via the communication network. This causes a misalignment in the timestamps of the two data streams, which should correspond to the same physical moment—a timestamp asynchronous bias. Directly mapping the misaligned data spatially will result in the load and temperature at the same moment not corresponding correctly, thus affecting the accuracy of thermal field decoupling. To eliminate this bias, a dynamic time warping algorithm is employed. This algorithm is a method for measuring the similarity of two time series and performing nonlinear alignment. Its basic principle is to match corresponding feature points in the two series by locally stretching and compressing the time axis while maintaining the temporal sequence. Specifically, an increase in electrical load usually causes a subsequent increase in local temperature, and the two have causally corresponding peaks in their waveforms. Using the peaks of the electrical load sequence and the local temperature sequence as matching features, the differences in the values ​​of the two sequences at each moment are compared point by point, and a matching cost is accumulated. The path with the minimum accumulated cost among all possible correspondences is the optimal matching path. By elastically stretching and compressing the time axis along the optimal matching path and resampling at uniform time intervals, a strictly synchronized alignment feature sequence in the time dimension can be obtained.

[0032] After obtaining the aligned feature sequence, this embodiment further performs spatial mapping. Specifically, based on the actual physical arrangement of each component inside the metering box, a two-dimensional spatial coordinate axis is established, so that the positions on the coordinate axis correspond one-to-one with the actual installation positions inside the box. Subsequently, according to the relative horizontal and vertical positions of each spatial node within the box, the aligned feature values ​​of the node are filled into the corresponding positions in the two-dimensional spatial matrix. In this way, the two-dimensional spatial matrix not only carries the feature values ​​of each node, but also preserves the relative spatial relationships between nodes through the arrangement of matrix elements, so that adjacent nodes are also in adjacent positions in the matrix. Finally, tensor stacking is performed along the time direction. As the time sliding window advances, a two-dimensional spatial matrix carrying spatial coordinate attributes is generated at each moment; the two-dimensional spatial matrices of multiple consecutive moments are stacked sequentially along the time direction, that is, combined to form a multi-node three-dimensional spatiotemporal tensor that simultaneously contains three dimensions: horizontal space, vertical space, and time. This three-dimensional spatiotemporal tensor integrates discretely distributed sensor data into a unified high-dimensional feature structure, preserving both the spatial topological relationships between nodes and the dynamic characteristics of each node's evolution over time. This provides a structurally complete input for subsequent graph attention networks to identify heat transfer relationships between nodes.

[0033] This invention integrates discrete, asynchronous, and one-dimensional time-attribute multi-channel sensor data into a multi-node three-dimensional spatiotemporal tensor that simultaneously carries spatial topology and temporal evolution information through two stages: time alignment and spatial mapping. Specifically, a dynamic time warping algorithm eliminates timestamp asynchronicity deviations caused by communication delays, ensuring the correct correspondence between electrical load and local temperature at the same physical moment and avoiding thermal field decoupling distortion caused by data misalignment. Compared to the traditional method of analyzing data from a single node at a single moment in isolation, the multi-node three-dimensional spatiotemporal tensor constructed in this scheme provides a structured, high-dimensional, unified expression for the overall modeling of heat transfer relationships. This allows subsequent algorithms to simultaneously consider spatial proximity effects and temporal cumulative effects, thereby significantly improving the accuracy and completeness of thermal field decoupling and aging analysis.

[0034] Preferably, the recognition process in the graph attention network by inputting a multi-node 3D spatiotemporal tensor into the graph includes: In the graph attention network, each spatial node is defined as a graph node, and the physical distance between nodes and the airflow path are defined as graph edges. A physical penalty function containing energy conservation rules and thermodynamic constraints is constructed to restrict the direction of heat conduction between adjacent nodes to only be from the high temperature region to the low temperature region. Under the constraint of the physical penalty function, self-attention operation is performed to extract the hidden state features of nodes, and a dynamic heat transfer weight matrix that conforms to the real physical heat transfer logic is output. The dynamic heat transfer weight matrix is ​​used to perform algebraic separation calculation on the overall temperature rise of the spatial nodes, and outputs the independent heating sequence generated by the current flowing through the node itself, as well as the thermal crosstalk sequence generated by the superposition of radiation and conduction from the surrounding heavily loaded nodes.

[0035] Graph Attention Networks (GANs) are deep learning models that utilize node features and their adjacency relationships to dynamically calculate the weights of mutual influence between nodes through an attention mechanism. Their application in this scenario is based on the fact that the components within the metering box are densely arranged in space. The temperature rise measured by any node is actually the result of its own heat generation and the heat radiation conduction from surrounding nodes. This mutual influence, determined by spatial adjacency, can be precisely expressed by a graph structure. In constructing the network structure, each spatial node is defined as a graph node, and the node's features are taken from the feature values ​​of that location in the three-dimensional spatiotemporal tensor. The physical distance between nodes and the airflow path are defined as graph edges. The closer the physical distance and the smoother the airflow between two nodes, the greater the possibility of heat transfer, and the stronger the corresponding connection. This constructed graph structure allows the network to consider the heat interaction between nodes based on the actual spatial layout during computation.

[0036] To ensure that the network's computational results conform to physical laws, a physical penalty function was constructed. This function, drawing inspiration from the design principles of physical information neural networks, penalizes computational results that violate physical laws during network training, thereby guiding the network to converge to a physically reasonable solution. Specifically, this penalty function includes two constraints. First, an energy conservation constraint requires that, within any given time window, the sum of heat flowing into a node from surrounding nodes and the heat generated by the node itself, minus the heat flowing out to surrounding nodes, must balance the change in internal energy corresponding to the node's temperature rise. Second, a thermodynamic direction constraint requires that the heat conduction direction weights between adjacent nodes can only point from high-temperature regions to low-temperature regions; that is, heat can only spontaneously transfer from high-temperature objects to low-temperature objects, which embodies the second law of thermodynamics. During the network's backpropagation, if the calculated heat conduction direction violates this law by pointing from low-temperature to high-temperature, the penalty function generates a large penalty, prompting the network to adjust its parameters to eliminate this unreasonable result.

[0037] Specifically, the energy constraint term is calculated as follows: For any spatial node and any time window, the inflow heat from the temperature rise contribution components of all surrounding source nodes flowing to that node within the window is accumulated item by item, plus the self-generated heat from the independent heating component of that node, and then the outflow heat from the temperature rise contribution components of that node flowing to surrounding nodes is subtracted to obtain the net heat of that node within the window; the change in the internal energy of that node is obtained by multiplying the mass parameter of the solid component of that node, the specific heat capacity parameter of the insulating material, and the measured temperature rise value within the window in sequence; the square of the deviation between the net heat and the change in internal energy is averaged over all spatial nodes and all time windows, and the resulting value is the penalty amount of the energy balance constraint term. The heat conduction direction constraint term is calculated as follows: For any pair of adjacent nodes and at any given time, the temperature values ​​of the two nodes are compared. When a transfer weight in the dynamic heat transfer weight matrix points from the low-temperature node to the high-temperature node and has a value greater than 0, the value of this violation weight is included in the penalty. Considering the noise in the measured temperature, when the absolute value of the temperature difference between two adjacent nodes is less than the preset isothermal tolerance, the two nodes are considered to be in an approximately isothermal state, and no direction penalty is applied. The isothermal tolerance is determined based on the measurement accuracy of the temperature sensor, and can be taken as 0.5 degrees Celsius as an example. All violation weight values ​​are summed over all pairs of adjacent nodes and at all times, and the resulting value is the penalty amount of the heat conduction direction constraint term. The overall training loss consists of two parts: a data fitting term and a physical penalty term. The physical penalty term is the sum of the energy conservation constraint term and the heat conduction direction constraint term. The internal weights of the two constraint terms can initially be set at a 1:1 ratio.

[0038] The network's computational flow is as follows: After inputting a multi-node 3D spatiotemporal tensor into the graph attention network, the network calculates attention weights for each graph node based on its feature differences with neighboring nodes to measure the influence of neighboring nodes on the node's thermal state. Under the constraint of the physical penalty function, the network extracts and updates the hidden state features of each node layer by layer through self-attention computation. After multiple rounds of computation, the network outputs a dynamic heat transfer weight matrix describing the intensity and direction of heat transfer between nodes. This matrix is ​​called dynamic because its values ​​change with time and load conditions, reflecting the real-time changes in heat transfer relationships under different operating conditions. The network's training process is as follows: Based on the temperature rise data of each node in historical operating data, an overall loss objective is constructed, including a data fitting term and the aforementioned physical penalty term. The data fitting term measures the deviation between the node temperature rise reconstructed by the network and the measured temperature rise, while the physical penalty term measures the degree to which the computational results violate energy conservation and thermodynamic direction constraints. During training, the network parameters are continuously adjusted through backpropagation to gradually reduce the overall loss, ultimately obtaining a network model that can both fit the measured data and strictly follow physical laws.

[0039] As a feasible network configuration, the graph attention network adopts a three-layer stacked structure, with four attention heads in each layer and a node hidden state feature dimension of 64. Training samples are organized in units of time sliding windows, with each window containing a 24-hour continuous three-dimensional spatiotemporal tensor slice. In the overall loss objective, the weight ratio of the data fitting term to the physical penalty term can be initially set at 10:1. During training, if the proportion of outputs violating the heat conduction direction constraint exceeds one percent, the weight of the physical penalty term is gradually increased until the violation disappears. The optimization process uses the commonly used adaptive gradient optimization method, training until the overall loss no longer decreases significantly over multiple consecutive rounds. The above configuration is only an example; those skilled in the art can adjust the number of layers and dimensions appropriately according to the scale of the metering bin nodes. After obtaining the dynamic heat transfer weight matrix, algebraic separation of heat is further performed.

[0040] For a given target node, its total temperature rise consists of two parts: the heat generated by the current flowing through the node itself and the heat radiated and conducted from surrounding heavily loaded nodes. The specific operational relationship of the algebraic separation is as follows: the total temperature rise of the target node at a given moment is considered as a linear superposition of the contributions from each heat source. The weight values ​​in the dynamic heat transfer weight matrix, pointing from the source node to the target node, represent the proportion of the temperature rise from the source node to the temperature rise of the target node. The temperature rise value of each surrounding node at that moment is multiplied by its corresponding transfer weight to obtain the temperature rise contribution component of that surrounding node to the target node. The temperature rise contribution components of all surrounding nodes are accumulated item by item and then subtracted from the total temperature rise of the target node; the remaining part is the independent heat generation component of that node. Performing the above operations time-by-time yields the independent heat generation sequence and the thermal crosstalk sequence, respectively. The rationality of the transfer weight values ​​is guaranteed by the data fitting term during the training phase, i.e., minimizing the deviation between the temperature rise of each node reconstructed by the subtraction operation and the measured temperature rise is one of the training objectives.

[0041] This invention solves the technical challenge of distinguishing heat sources in densely packed spaces due to mutual interference from heating between adjacent components by introducing a graph attention network with prior thermophysical constraints. By modeling spatial nodes as graph nodes and physical distances and airflow paths as graph edges, the network considers the heat interaction between nodes based on the actual spatial layout, aligning with the physical reality of densely packed equipment. Compared to traditional methods that rely solely on single-point temperature thresholds for alarms, this solution accurately identifies the true source of heat, providing a reliable basis for subsequent source tracing and control. It effectively avoids false alarms or frequent malfunctions at affected nodes, improving the accuracy of thermal field analysis and the reliability of system judgment.

[0042] Preferably, the process of reconstructing the solid temperature sequence of the insulating surface of each spatial node includes: A physical thermal inertia extrapolation model is constructed using a neural network of ordinary differential equations; the specific heat capacity and mass parameters of the insulating material are input into a multilayer perceptron network to establish a temperature change rate derivative prediction layer that characterizes the temperature rise hysteresis of solids. The independent heating sequence and the thermal crosstalk sequence are superimposed and combined into a composite heat source excitation, and the composite heat source excitation and the node air temperature are synchronously input into the temperature change rate derivative prediction layer. The ordinary differential equation solver is called to perform continuous integration calculation on the time step according to the temperature change rate gradient output by the network, filter out the high-frequency fluctuation noise in the node air temperature, and invert the output to output the real state trajectory of the insulating surface with physical smooth response and phase hysteresis characteristics, which is denoted as the solid temperature sequence.

[0043] The physical thermal inertia extrapolation model is constructed using a neural network of ordinary differential equations (ODEs). This type of network does not employ discrete network layers but instead models the evolution of hidden layer states as continuous-time ODEs. The rationale for using this network is that the heating process of an insulating solid physically follows the thermal resistance-capacitance law; that is, the solid's temperature response to heat input is a continuous, inertial integral process, which highly aligns with the mathematical form of ODEs. Compared to discrete models that predict point-by-point at fixed time steps, continuous-time modeling can more naturally characterize the hysteresis and smoothness of solid heating.

[0044] First, a temperature change rate derivative prediction layer is established. This layer predicts not the absolute value of the temperature at the next moment, but rather the rate of temperature change over time—the trend of heating or cooling. The reason for predicting the rate of change rather than the absolute value is that the change in solid temperature is essentially determined by the current heat balance. Predicting the rate of change better reflects the physical causal relationship of heat driving temperature evolution and facilitates obtaining a continuous and smooth temperature trajectory through integration. This prediction layer is constructed using a multilayer perceptron network. Its inputs include the specific heat capacity parameter of the insulating material and the mass parameter of the insulating solid belonging to the spatial node. The mass parameter is obtained by multiplying the density of the insulating material by the geometric volume of the solid component at that node. Specific heat capacity and mass together determine the solid's thermal capacity; the larger the thermal capacity, the slower the solid's temperature response to heat input, and the more pronounced the temperature lag. Therefore, inputting the specific heat capacity and mass parameters into the multilayer perceptron allows the prediction layer to determine the response characteristics of its temperature change rate based on the solid's inherent thermal inertia.

[0045] During the extrapolation, independent heating sequences and thermal crosstalk sequences are superimposed and combined to form a composite heat source excitation acting on the node. This excitation includes both the node's own heating and heat conducted from the surrounding environment. Subsequently, the composite heat source excitation and the node's air temperature are synchronously input into the temperature change rate derivative prediction layer. Based on the current heat source excitation, node air temperature, and the thermal inertia properties of the solid, the prediction layer outputs the temperature change rate gradient of the solid at that moment. After obtaining the temperature change rate gradient, a continuous integration operation is performed using an ordinary differential equation solver. This ordinary differential equation solver is a computational tool that performs numerical integration of state variables over time based on a given rate of change. The solver continuously integrates over the time step based on the temperature change rate gradient, gradually accumulating the temperature changes at each moment, thereby deducing the solid temperature at subsequent moments from the initial temperature. Since the integration process is essentially a smooth accumulation of rapid changes, the high-frequency fluctuation noise in the node air temperature caused by convective disturbances is naturally filtered out during the integration process. The final temperature trajectory output by inversion exhibits a physically reasonable smooth response and phase hysteresis characteristics. This trajectory is the solid temperature sequence that reflects the true thermal state of the insulating surface.

[0046] The initial temperature for integration is taken as the measured local temperature value of the node at the start of the derivation. Since the start of the derivation is usually chosen in the early morning when the load is low and the solid is in full thermal equilibrium with the surrounding air, the air temperature and the solid surface temperature are approximately equal at this time. Therefore, using the local temperature as the initial value of the solid temperature is physically reasonable. The ordinary differential equation solver can use a classical fourth-order solution method with a fixed step size or an adaptive step size method with error control. The integration step size is consistent with the data sampling period or is taken as an integer part of it.

[0047] The model training process is as follows: In the offline calibration phase, the typical load conditions of the metering box are reproduced on the test bench. The synchronously acquired composite heat source excitation and node air temperature are used as model inputs, and the real temperature of the insulation surface is obtained using non-contact temperature measurement methods such as infrared thermal imaging as training labels. Since the calibration is completed in a controlled test environment, non-contact temperature measurement does not disturb the original heat capacity distribution and condensation physical field of the insulation surface, and can cover a large surface area. The deviation between the solid temperature output by the network inversion and the non-contact measured temperature is used as the training target. The parameters of the multilayer sensor are adjusted through backpropagation so that the temperature change rate output by the network, after integration, can approximate the real solid temperature evolution. After calibration, the model is deployed in the online operation phase, and no contact surface temperature measurement device needs to be installed on site.

[0048] This invention reconstructs the solid temperature sequence of an insulating surface using a neural network of constant differential equations, solving the technical problem of traditional monitoring methods that rely solely on air temperature and neglect the hysteresis effect of solid thermal inertia. By employing continuous-time modeling and using the rate of temperature change as the prediction object, the model can naturally characterize the hysteresis and smoothness of solid temperature rise based on the physical causal relationship of heat balance, which is more consistent with physical laws than discrete point-by-point prediction. Specific heat capacity and mass parameters are input into the prediction layer as key parameters characterizing solid thermal inertia, enabling the model to determine the degree of sluggishness in temperature response based on the material's own properties, thus improving the physical rationality of solid temperature inversion. This provides a reliable temperature basis for accurately determining whether condensation has occurred on the solid surface, thereby significantly improving the accuracy of condensation prediction and aging analysis.

[0049] Preferably, the process of obtaining the surface condensation duration sequence includes: Extract temperature and humidity data from the external meteorological time series data, calculate the dynamic critical condensation boundary curve under the interaction of cold air outside the box and hot air inside the box, and output the dynamic dew point; The reconstructed solid temperature sequence and the dynamic dew point are cross-referenced in the time domain. The intersection critical time point where the solid temperature value decreases and is lower than the dynamic dew point value is extracted as the starting point of liquid water film precipitation, and the intersection critical time point where the solid temperature value rises and is higher than the dynamic dew point value is extracted as the end point of water film evaporation. By accumulating the time span difference between the precipitation start point and the evaporation end point, a surface condensation duration sequence is output to characterize the duration of continuous moisture immersion on the surface of the insulating material.

[0050] First, the dynamic dew point is calculated based on external meteorological time-series data. The dew point refers to the temperature at which air, under conditions of constant water vapor content, cools to the point where water vapor reaches saturation and begins to condense. When the surface temperature of an object drops below the dew point, water vapor in the air in contact with it will condense into liquid water on the object's surface. This embodiment extracts the absolute temperature and relative humidity values ​​from external meteorological data and calculates the dew point based on the correspondence between air temperature and humidity and saturated water vapor. Due to the continuous interaction between cold air outside the enclosure and hot air inside, the ambient temperature and humidity are dynamically changing. Therefore, the calculated dew point also dynamically changes, forming a critical condensation boundary curve that varies with time, called the dynamic dew point. This dynamic dew point characterizes the critical limit at which condensation will occur on a solid surface once the temperature drops below this level under current environmental conditions.

[0051] After obtaining the dynamic dew point, a time-domain cross-comparison analysis is performed. This analysis involves placing two curves with the same time axis in the same coordinate system and determining the change in physical state by finding their intersection. The solid temperature sequence and the dynamic dew point curves are compared along the same time axis. When the solid temperature decreases over time and intersects the dynamic dew point curve, and remains below the dynamic dew point thereafter, it indicates that the solid surface temperature has dropped below the critical dew point, and water vapor in the air begins to condense on the solid surface. Therefore, this critical intersection point is determined as the starting point for the precipitation of the liquid water film. When the solid temperature rises over time and intersects the dynamic dew point curve again, and remains above the dynamic dew point thereafter, it indicates that the solid surface temperature has risen above the critical dew point, and the precipitated liquid water film begins to evaporate and dissipate. Therefore, this critical intersection point is determined as the end point for the evaporation of the water film.

[0052] After obtaining the initiation point and evaporation endpoint of the water film precipitation, the condensation duration is further calculated. Specifically, for each pair of adjacent precipitation initiation points and evaporation endpoints, the time span difference between them is calculated. This difference characterizes the time during which the solid surface is continuously wetted by the liquid water film in this condensation process. By accumulating and arranging the time spans of each condensation process during operation in chronological order, a surface condensation duration sequence is constructed to characterize the continuous water wetting time of the insulating material surface.

[0053] This invention achieves algorithmic measurement of invisible trace amounts of condensation on solid surfaces inside electrical equipment through dynamic dew point calculation and time-domain cross-comparison analysis. It solves the technical problem of traditional environmental monitoring that relies solely on high air temperature and humidity to determine hazard, failing to calculate actual condensation on solid surfaces. Compared to the crude judgment of directly equating air humidity with condensation risk, this solution, through precise comparison of solid temperature and dew point, can distinguish situations where the air is humid but the solid surface temperature is high due to thermal inertia, and condensation has not actually occurred. This avoids misjudgment, ensuring that condensation identification is based on the true thermal state of the solid surface, providing accurate data on moisture retention time for subsequent electrochemical dual-stress coupled aging damage assessment.

[0054] Preferably, the step of identifying the cumulative insulation aging index of each spatial node under dual stress coupling specifically includes: A multimodal cross-attention deep network is constructed as the core mapping analysis model; The surface condensation duration sequence is input into a time-domain gating layer for activation function operation to generate a mask matrix representing the probability of moisture adhesion. The surface leakage current time series data is converted into feature query vector and feature key value vector, and the mask matrix is ​​used as the guiding parameter of the attention mechanism to perform fusion calculation, extracting and amplifying the leakage current peak feature tensor that occurs during moisture retention. The leakage current peak feature tensor is input into a deep Weibull regression layer, and the scale and shape parameters in the Weibull distribution model are output by nonlinear mapping. The cumulative degradation value of the survival function is calculated based on the scale and shape parameters, and the output is the insulation aging cumulative index.

[0055] The mapping analysis model is constructed using a multimodal cross-attention deep network. This network is capable of handling various input types, with the two modes being the surface condensation duration sequence representing environmental moisture conditions and the surface leakage current time series data representing signs of electrical degradation. The rationale for using this network is that the aging of insulating materials is the result of the combined stress of moisture and electrical discharge. Only when a liquid water film is actually present on the solid surface are the high-energy peaks of the leakage current truly destructive; leakage current fluctuations occurring during dry periods contribute less to aging. Therefore, it is necessary to use moisture conditions as a guide to filter out the truly destructive components from the leakage current.

[0056] First, a mask matrix is ​​generated from the condensation duration sequence. Specifically, the surface condensation duration sequence is input into a time-domain gating layer. This gating layer consists of a fully connected network and a saturated activation function with an output value range of zero to one, outputting a mask matrix whose values ​​reflect the probability of moisture adhesion at each time point. During periods of long condensation duration and high moisture adhesion probability, the corresponding values ​​of the mask matrix approach one, and vice versa. In the cross-fusion structure of this scheme, the condensation mode does not directly generate key values, but rather modulates the attention score time-by-time in the form of a mask matrix: the surface leakage current time series data is linearly transformed into feature query vectors and feature key value vectors, respectively. After calculating the correlation score between the query and the key value according to the attention mechanism, the score is multiplied by the corresponding value in the mask matrix, and then normalized to obtain the final attention weight.

[0057] Specifically, the surface leakage current time-series data is organized into a sliding window consistent with the condensation duration sequence, with each window containing continuous 24-hour leakage current sampling data. The leakage current sequence within each window is first subjected to a one-dimensional convolution operation to extract local waveform features, resulting in a 64-dimensional leakage current feature vector for each moment. This feature vector is then transformed into a 64-dimensional feature query vector and a 64-dimensional feature key vector through two independent linear transformations. The mask matrix corresponds to the leakage current features on the time axis, with each moment's value being a scalar between 0 and 1, representing the probability of moisture attachment at that moment. The attention score is obtained by calculating the correlation between each pair of feature query vectors and feature key vectors. Then, the attention score at each moment is multiplied by the corresponding value in the mask matrix to achieve time-by-time weighted modulation. A normalization operation is then performed to ensure the sum of the weights at each moment equals 1, yielding the final attention weight. This attention weight is then used to weight and sum the feature key vectors, resulting in a leakage current peak feature tensor focused on the moisture retention period. This feature tensor has the same 64-dimensional dimension as the single-moment feature vector. The leakage current characteristics during the drying period are modulated to have a weight close to 0, and their information is naturally suppressed rather than forcibly set to zero, thus preserving the overall continuity of the sequence while focusing on the water film participation period.

[0058] In the cross-attention mechanism, during periods with larger mask matrix values, the corresponding leakage current features receive higher attention weights, thus being truncated and amplified; during periods with smaller mask matrix values, the corresponding leakage current features are suppressed. In this way, the computation focuses on the high-energy leakage current peaks occurring during water retention, summarizing them into a high-energy leakage current peak feature tensor; only the truly destructive leakage current features involved in the water film are retained, while harmless leakage current fluctuations during the dry period are suppressed.

[0059] After obtaining the high-energy leakage current peak feature tensor, it is input into a deep Weibull regression layer. The deep Weibull regression layer is characterized by not directly outputting a single aging prediction value, but rather outputting parameters of a Weibull distribution model used to describe the material's lifetime distribution. The Weibull distribution is a probability distribution widely used to describe the lifetime and failure patterns of materials; its shape is determined by both scale and shape parameters. The scale parameter reflects the characteristic magnitude of lifetime, while the shape parameter reflects the trend of failure rate over time. The basis for using the Weibull distribution for modeling is that the degradation and failure of insulating materials statistically conform to the characteristics of this distribution, and characterizing aging with distribution parameters can more comprehensively reflect the uncertainty of material lifetime. The deep Weibull regression layer uses nonlinear mapping to deduce the corresponding scale and shape parameters from the input high-energy leakage current peak feature tensor. Finally, the aging accumulation index is calculated based on the obtained scale and shape parameters. Specifically, the survival function is determined based on the scale and shape parameters of the Weibull distribution. This function describes the probability that the insulating material will remain intact after a given cumulative operating time. Starting from the equipment commissioning time and ending at the current time, the cumulative decrease of the survival function from one value to the current value within this interval is calculated. This cumulative decrease is used as the normalized cumulative degradation value characterizing the degree of insulation failure risk. This value is the insulation aging cumulative index, which takes a value between zero and one. The larger the value, the deeper the insulation degradation.

[0060] The model training process is as follows: Using historically collected condensation duration sequences, leakage current sequences, and the actual degradation levels of insulation materials detected during corresponding periods as samples, the deviation between the network's output aging cumulative index and the actual degradation level is used as the training objective. The parameters of the time-domain gating layer, cross-attention mechanism, and deep Weibull regression layer are jointly adjusted through backpropagation, enabling the network to accurately calculate the aging level of the insulation from moisture and leakage current data. During the training of the deep Weibull regression layer, the negative log-likelihood function of Weibull survival analysis is used as the model's loss function to handle right-censored data resulting from equipment not yet being completely damaged.

[0061] This invention achieves coupled damage assessment based on the dual stresses of moisture and electrical leakage through a multimodal cross-attention network and a deep Weibull regression layer. This solves the technical problems of traditional methods that separate environmental moisture and electrical leakage, resulting in coarse aging quantification. Compared to simple judgments based solely on leakage current amplitude or a single physical quantity, this scheme recreates the carbonization physical process of insulating materials under the combined effects of humidity and high electric field strength. It only includes aging equivalents within the time window where a water film is present, thus significantly improving the accuracy and reliability of insulation aging prediction and providing a dependable quantitative basis.

[0062] Preferably, the process of sending dynamic peak shaving and load reduction commands includes: A global monitoring and comparison matrix is ​​constructed. When the cumulative insulation aging index of a spatial node exceeds the safety threshold, the historical thermal crosstalk sequence corresponding to the spatial node is retrieved. By searching the spatial coordinates of the source node with the largest heat contribution value and the longest duration in the historical thermal crosstalk sequence, the source device that generates the influence of nearby thermal radiation is reversely locked as the heavy-load heat-generating node. A dynamic peak-shaving and load-reducing command is generated to reduce the upper limit of the operating power of the heavy-load heat-generating node.

[0063] First, a global monitoring and comparison matrix is ​​constructed. This matrix aggregates the cumulative insulation aging index of each spatial node and compares the cumulative aging index of each node with a preset material failure critical benchmark, which serves as the safety threshold. When the cumulative aging index of any node exceeds this critical benchmark, the node is identified as a vulnerable spatial node facing the risk of insulation failure. For the vulnerable spatial nodes, reverse tracing is performed. This reverse tracing refers to not directly addressing the vulnerable nodes exhibiting aging symptoms, but rather tracing back along the connection relationships of the graph network to find the source node exerting the greatest thermal impact. The material failure critical benchmark is determined as follows: during the offline calibration phase, dielectric performance testing is performed on several decommissioned insulation components. The cumulative aging index statistics corresponding to components judged to be nearing failure are used as the benchmark calibration basis. For example, the benchmark can be set to 0.7, meaning that when the cumulative aging index of a node exceeds 0.7, it is determined to face the risk of insulation failure. This benchmark can be adjusted within the range of 0.6 to 0.8 depending on the importance of the equipment and the operation and maintenance strategy.

[0064] Specifically, the historical thermal crosstalk sequence corresponding to the affected spatial node is retrieved. This sequence records the changes in the heat contribution of surrounding nodes to the affected node over time. The source node with the largest heat contribution value and the longest duration is retrieved from this sequence, and its corresponding spatial coordinates point to the source device that actually causes the influence of neighboring thermal radiation. Thus, this source device is identified as a heavily loaded heat-generating node. The basis for this source tracing method is that the overheating and aging of the affected node often does not originate from itself, but is caused by the continuous thermal radiation from neighboring heavily loaded nodes. Therefore, only by identifying and addressing the true heat source can the cause of overheating be fundamentally eliminated. After identifying the heavily loaded heat-generating node, a dynamic peak-shaving and load-reducing command is generated. This command is used to reduce the upper limit of the operating power of the heavily loaded heat-generating node and is sent to the intelligent load control terminal corresponding to the node. The intelligent load control terminal adjusts the load limit value of the circuit according to the command, and reduces the upper limit of the operating power by cutting off non-safety load branches under the jurisdiction of the node in stages or lowering the current limiting setting. When the node power exceeds the upper limit after voltage drop, the intelligent load control terminal will trigger the corresponding branch switch to perform load reduction. After the power limit is reduced, the heating of the heavily loaded heat-generating node will decrease accordingly, and its thermal radiation crosstalk to the affected node will also decrease, thereby weakening the temperature conditions for condensation to form due to the convergence of hot and cold, and inhibiting the further development of condensation and aging.

[0065] To avoid frequent actions under critical conditions, an anti-oscillation recovery dead zone temperature range is set in the control logic loop. This anti-oscillation recovery dead zone borrows the principle of a hysteresis comparator in industrial control. Its function is to introduce a temperature buffer in the determination of power lockout release, preventing repeated shunt switch operation caused by small fluctuations in node temperature at the threshold edge. Specifically, after load reduction, the lockout is not released immediately upon a slight temperature drop. Instead, the node is only considered to have sufficiently dissipated heat when the node temperature is continuously decreasing and remains below the lower limit of the dead zone for more than the set heat release integration period. As an example, the width of the anti-oscillation recovery dead zone temperature range can be five to ten degrees Celsius, meaning the upper limit of the dead zone is the node temperature at the time of load reduction triggering, and the lower limit is five to ten degrees Celsius lower. The heat release integration period can be thirty to sixty minutes, with the specific value determined based on the heat capacity of the node's solid components; a larger heat capacity results in a longer integration period. This avoids frequent shunt switch operation near the critical state, ensuring a smooth control process.

[0066] This invention constructs a closed-loop control strategy to address overheating through reverse tracing and dead-zone hysteresis control, solving the technical problem of traditional methods that blindly cut off power to affected nodes without eliminating the cause of overheating. By using a global monitoring and comparison matrix to uniformly monitor and determine the aging accumulation index of each node, timely detection of insulation failure risks is achieved. Compared to passive methods that only alarm or cut off power to affected nodes, this invention proactively eliminates the cause of overheating, ensuring the thermal safety of the equipment.

[0067] Example 2: This invention proposes a time-series data analysis method for the aging of metering box components, including: Acquire electrical load time-series data, local temperature time-series data, surface leakage current time-series data, and external meteorological time-series data from multiple spatial nodes inside the metering box; based on the physical topological location of the spatial nodes, perform spatial mapping and temporal alignment of the electrical load time-series data and the local temperature time-series data to construct a multi-node three-dimensional spatiotemporal tensor; The multi-node three-dimensional spatiotemporal tensor is input into a graph attention network with thermophysical prior constraints to identify the dynamic heat transfer weight matrix and decouple the mixed heat of spatial nodes into independent heat generation sequences and thermal crosstalk sequences. Independent heating sequences and thermal crosstalk sequences are used as composite heat source excitations and input into the physical thermal inertia extrapolation model to reconstruct the solid temperature sequence of the insulating surface of each spatial node; dynamic dew point is calculated based on external meteorological time series data, and then the surface condensation duration sequence is obtained by combining the solid temperature sequence. The surface condensation duration sequence and surface leakage current time series data are fused and mapped to identify the insulation aging cumulative index of each spatial node under dual stress coupling. When the cumulative insulation aging index exceeds the safety threshold, the overloaded heating node is traced based on the thermal crosstalk sequence, and a dynamic peak-shaving and load-reduction command is sent.

[0068] Furthermore, in the process of obtaining the surface condensation duration sequence, the vertical height attribute of each spatial node in the multi-node three-dimensional spatiotemporal tensor in the two-dimensional spatial coordinate axis is extracted, and a humidity vertical attenuation mapping function is constructed with the bottom of the box as the high humidity condensation tendency zone and the top of the box as the heat accumulation low humidity tendency zone. Based on the vertical height attribute, the humidity vertical attenuation mapping function is used to perform gradient correction on the absolute humidity of the box base obtained based on external meteorological time series data to obtain the true local humidity at each height level. The true local humidity is substituted into the saturated water vapor conversion relationship to calculate the exclusive differential dynamic dew point curve of the spatial node at different vertical heights. Finally, the reconstructed solid temperature sequence and the corresponding exclusive dynamic dew point curve are cross-compared in the time domain to output the surface condensation duration sequence.

[0069] In actual operation, metering chambers commonly exhibit a "breathing effect," meaning that humid external air is largely drawn in through the inlet / outlet openings at the bottom of the chamber. After entering through these openings, the newly arrived external air, being cooler and denser, remains trapped at the bottom of the chamber. Simultaneously, the continuous heating from the heavy-duty nodes at the top of the chamber increases the temperature of the upper air, correspondingly increasing its saturation capacity. Furthermore, the thermally rising airflow hinders the upward diffusion of the humid air from the bottom. As a result, an absolute humidity distribution pattern gradually decreases from bottom to top within the chamber. In implementation, the longitudinal coordinates of each spatial node in the multi-node three-dimensional spatiotemporal tensor are first extracted to determine its vertical height.

[0070] Subsequently, a humidity vertical attenuation mapping function is established. Its construction method is as follows: During the offline calibration phase, three temporary temperature and humidity measuring points (top, middle, and bottom) are arranged vertically inside the chamber. The temperature and humidity distribution inside the chamber and the corresponding external meteorological data are continuously recorded for at least one complete 24-hour cycle. The ratio of the absolute humidity at the bottom measuring point to the absolute humidity outside is used as the bottom anchoring coefficient, and the ratio of the absolute humidity at the top measuring point to the bottom measuring point is used as the top attenuation coefficient. The temperature difference between the top and bottom points characterizes the intensity of heat convection from bottom to top inside the chamber. The attenuation coefficients obtained from actual measurements under different heat convection intensities are fitted to obtain a nonlinear mapping function with vertical height and top-to-bottom temperature difference as input and humidity conversion coefficient as output. After calibration, the temporary measuring points are removed. During online operation, the true local humidity at each height level can be calculated using this mapping function based solely on the vertical height attributes of each node and the top-to-bottom temperature difference given by the reconstructed temperature field.

[0071] After obtaining the calculated local humidity at each altitude level, it is substituted into the saturated water vapor conversion relationship to deduce the unique dynamic dew point curve for each spatial node. Finally, the solid temperature sequence reconstructed by the physical thermal inertia model is cross-compared with the corrected unique dynamic dew point curve to obtain water film precipitation and evaporation time nodes that are more in line with physical reality, thus outputting a high-precision surface condensation duration sequence.

[0072] This invention overcomes the limitations of traditional data analysis methods that treat the interior of electrical enclosures as a single, uniform humidity environment. By introducing a vertical humidity gradient correction mechanism, it recreates the true microclimate physical field of the metering box, where condensation easily occurs at the bottom due to the entry and retention of cold, humid air, while the top remains relatively dry. This eliminates the technical pain points of underestimating condensation in the cold bottom area caused by using uniform environmental dew point calculations, and misjudging condensation in the hot top area due to high temperatures, thus improving the physical authenticity and accuracy of moisture retention time sequence extraction.

[0073] It should be noted that the metering chamber is not an airtight structure. Its bottom inlet and outlet pores continuously exchange air with the external environment due to diurnal temperature variations and air pressure fluctuations. The absolute water vapor content of the air inside the chamber tends to be consistent with the external environment on an hourly timescale. Therefore, the dew point calculated using temperature and humidity data from external meteorological time series can be used as an engineering approximation of the dew point inside the chamber. For chambers with good sealing or significant humidity stratification, the local humidity at each height level can be corrected using the following humidity vertical attenuation mapping method before calculation to further improve the accuracy of the dew point.

[0074] Furthermore, the temperature rise contribution of any spatial node within a unit time step is multiplied sequentially by the mass parameter of the solid component and the specific heat capacity parameter of the insulating material at that node to obtain the heat value corresponding to the temperature rise contribution. Therefore, there is a definite one-to-one correspondence between the temperature rise contribution and the heat. In this invention, the independent heating sequence and thermal crosstalk sequence output by algebraic separation both refer to sequences with temperature rise contribution as the dimension. Both can be converted into the corresponding heat sequence through the above conversion relationship. The composite heat source excitation mentioned in this invention refers to the equivalent heat source driving quantity characterized by the superposition of the independent heating sequence and the thermal crosstalk sequence. When it is input into the physical thermal inertia deduction model, the model internally completes the conversion from temperature rise contribution to heat driving based on the synchronously input specific heat capacity parameter and mass parameter.

[0075] Furthermore, in the fusion mapping stage, the temperature gradient characteristics under the combined effect of independent heating sequence and thermal crosstalk sequence are extracted. The amount of thermal expansion and contraction micro-slippage and the time of slippage occurrence are calculated by combining the thermal expansion coefficient of the metal material. The time domain comparison is performed between the time of slippage occurrence and the water film duration recorded in the surface condensation duration sequence. When the time of slippage occurs completely falls within the water film duration range, it is determined that the water is drawn into the newly exposed gap by capillary action. Then, the deep gap electrochemical corrosion penalty weight is added to the calculation model of the insulation aging cumulative index to accelerate the output of aging damage assessment results.

[0076] Inside the metering box, the main incoming busbar and the terminals of each branch switch are typically fastened together by different metals with significantly different coefficients of thermal expansion. Under the intense and highly uneven thermal impact of tidal power loads, the contact surfaces experience frequent thermal expansion and contraction, leading to relative sliding friction. If this physical-level relative slippage occurs during the humid period before the condensation film has completely evaporated, capillary action will instantly draw liquid moisture from the surface into the depths of the metal microcracks and fastening gaps exposed by the slippage, causing deep crevice corrosion. In practice, the reconstructed solid temperature sequence is used to extract the derivative step characteristics of rapid temperature increases or decreases over time. Combined with the known linear expansion coefficients of the metals, the theoretical microscopic slippage displacement and occurrence time of the contact surface are calculated. Subsequently, a time-domain coordinate comparison is performed. If the precise time point of the slippage is found to fall entirely within the undried region covered by the surface condensation duration sequence, an accelerated penalty for deep crevice corrosion is applied on top of the original dual-stress fusion mapping, amplifying the aging cumulative damage assessment value at this point.

[0077] This invention captures the intersection of microscopic mechanical deformation of metals and phase change of environmental moisture in the time dimension, extending the reach of macroscopic time-series data analysis to the hidden microscopic damage mechanism inside electrical terminals. It provides early warning support for weak links in metal contacts that are prone to instantaneous poor contact or even short circuit and burnout, avoiding the underestimation of deep hidden corrosion by traditional surface monitoring methods and enhancing the comprehensiveness of system disaster prevention.

[0078] Furthermore, after identifying the cumulative insulation aging index, the target damaged spatial node with the highest cumulative insulation aging index is extracted; the peak-valley temperature difference fluctuation amplitude of the solid temperature sequence corresponding to the target damaged spatial node within a continuous sliding time window, and the surface leakage current characteristics generated during the existence of surface condensation are obtained; a dual mapping benchmark is established, which characterizes the metal thermal fatigue resistance drift of the power metering sampling element with the peak-valley temperature difference fluctuation amplitude and the parasitic shunting of the sampling circuit with the surface leakage current characteristics; the dual mapping benchmark is combined for coupling calculation to output the predicted value of the power metering error drift of the target damaged spatial node, and a statutory verification warning command is generated when the predicted value exceeds the limit.

[0079] The core function of metering equipment is accurate billing. When a damaged spatial node is subjected to prolonged high-temperature radiation from surrounding heavily loaded nodes and frequent condensation-induced moisture erosion, the alloy shunt used for current sampling inside will undergo accelerated thermal fatigue aging due to intense heating and cooling, causing a physical drift in its inherent resistance value. Simultaneously, the weak leakage current on the insulation surface under the water film will form a parasitic leakage path in parallel with the original current in the sampling circuit. This dual damage directly leads to severe billing inaccuracies. After locating the target damaged spatial node whose insulation aging cumulative index exceeds the limit, the peak-valley temperature difference amplitude of the actual solid temperature sequence experienced by the node, as well as the peak characteristics of the surface leakage current time series data during condensation, are retrieved. These characteristic data are input into a pre-constructed metal lattice thermal fatigue and parasitic shunt coupling error calculation matrix to deduce the predicted value of the electricity metering error drift for the metering station in the future. When the predicted value of the electricity metering error drift exceeds the maximum allowable error range, in addition to the conventional insulation aging warning, a legally mandated verification warning instruction for the metering station is further output, prompting the metering station to be removed for verification in advance.

[0080] The coupling error calculation matrix is ​​constructed as follows: During the calibration phase, different amplitudes and numbers of cold and hot cycle tests are applied to current sampling alloy shunts of the same model, and the correspondence between the peak-valley temperature difference amplitude and the resistance drift rate is recorded; simultaneously, on a standard metrology calibration bench, simulated leakage current branches of different magnitudes are connected in parallel to the sampling circuit, and the correspondence between the leakage current peak amplitude and the metrology deviation is recorded; the above two sets of calibration relationships are organized into a two-dimensional lookup table with the peak-valley temperature difference amplitude and leakage current peak characteristics as the query entry and the metrology error component as the output, and the two types of error components are synthesized into the overall metrology error prediction value by linear superposition. The maximum allowable error range is determined according to the statutory tolerance of the accuracy class to which the metrology device belongs.

[0081] This invention transforms the abstract physical temperature change and electrical insulation dual stress index into an early warning defense against losses, enabling aging monitoring results to be directly linked to meter calibration services, thus enhancing the application value of this invention.

[0082] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A time-series data analysis method for the aging of metering box components, characterized in that, include: Acquire electrical load time-series data, local temperature time-series data, surface leakage current time-series data, and external meteorological time-series data for multiple spatial nodes inside the metering box; Based on the physical topological location of spatial nodes, the electrical load time series data and local temperature time series data are spatially mapped and time-aligned to construct a multi-node three-dimensional spatiotemporal tensor. The multi-node three-dimensional spatiotemporal tensor is input into a graph attention network with thermophysical prior constraints to identify the dynamic heat transfer weight matrix and decouple the mixed heat of spatial nodes into independent heat generation sequences and thermal crosstalk sequences. The recognition process in a graph attention network, which inputs a multi-node 3D spatiotemporal tensor into a graph, includes: In the graph attention network, each spatial node is defined as a graph node, and the physical distance between nodes and the airflow path are defined as graph edges; Construct a physical penalty function that includes energy conservation rules and thermodynamic constraints; under the constraints of the physical penalty function, perform self-attention operation to extract the hidden state features of nodes, and output a dynamic heat transfer weight matrix that conforms to the real physical heat transfer logic; The dynamic heat transfer weight matrix is ​​used to perform algebraic separation calculation on the overall temperature rise of the spatial nodes, and outputs the independent heating sequence generated by the current flowing through the node itself, as well as the thermal crosstalk sequence generated by the superposition of radiation and conduction from the surrounding heavily loaded nodes. Independent heating sequences and thermal crosstalk sequences are used as composite heat source excitations and input into the physical thermal inertia extrapolation model to reconstruct the solid temperature sequence of the insulating surface of each spatial node; dynamic dew point is calculated based on external meteorological time series data, and then the surface condensation duration sequence is obtained by combining the solid temperature sequence. The surface condensation duration sequence and surface leakage current time series data are fused and mapped to identify the insulation aging cumulative index of each spatial node under dual stress coupling. The specific steps for identifying the cumulative insulation aging index of each spatial node under dual stress coupling include: A multimodal cross-attention deep network is constructed as the core mapping analysis model; the surface condensation duration sequence is input into the time-domain gating layer for activation function operation to generate a mask matrix representing the probability of water adhesion; the surface leakage current time series data is converted into feature query vector and feature key value vector, and the mask matrix is ​​used as the guiding parameter of the attention mechanism to perform fusion calculation, and the leakage current peak feature tensor that occurs during water retention is extracted and amplified; The leakage current peak feature tensor is input into a deep Weibull regression layer, and the scale and shape parameters in the Weibull distribution model are output by nonlinear mapping. The cumulative degradation value of the survival function is calculated based on the scale and shape parameters, and the output is the insulation aging cumulative index. When the cumulative insulation aging index exceeds the safety threshold, the overloaded heating node is traced based on the thermal crosstalk sequence, and a dynamic peak-shaving and load-reduction command is sent.

2. The time-series data analysis method for aging of metering box components according to claim 1, characterized in that: The steps for obtaining time-series data of multiple spatial nodes inside the metering box specifically include: extracting the effective value of the operating current and the active power value of the metering box within a continuous time period to form the electrical load time-series data; capturing the temperature data inside the metering box to form the local temperature time-series data; obtaining the creepage fluctuation signal of the insulating components as the surface leakage current time-series data; and obtaining the temperature and humidity change data of the physical environment in which the box is located as the external meteorological time-series data.

3. The time-series data analysis method for aging of metering box components according to claim 1, characterized in that: The specific steps for constructing a multi-node three-dimensional spatiotemporal tensor include: extracting the asynchronous timestamp deviation caused by communication delay, using a dynamic time warping algorithm to find the optimal matching path between the peak of the electrical load time series data and the peak of the local temperature time series data, elastically stretching and compressing the time axis, and resampling to output an aligned feature sequence that synchronizes the time dimension. Establish a two-dimensional spatial coordinate axis consistent with the internal hardware layout of the metering box, and map the alignment feature sequence into the corresponding two-dimensional spatial matrix according to the horizontal and vertical relative positions of each spatial node in the box. Along the advancing direction of the time sliding window, multiple consecutive two-dimensional spatial matrices containing time attributes and spatial coordinate attributes are stacked and combined to generate the multi-node three-dimensional spatiotemporal tensor.

4. The time-series data analysis method for aging of metering box components according to claim 1, characterized in that: The process of reconstructing the solid temperature sequence of the insulating surface of each spatial node includes: A physical thermal inertia extrapolation model is constructed using a neural network of ordinary differential equations; the specific heat capacity and mass parameters of the insulating material are input into a multilayer perceptron network to establish a temperature change rate derivative prediction layer that characterizes the temperature rise hysteresis of solids. The independent heating sequence and the thermal crosstalk sequence are superimposed and combined into a composite heat source excitation, and the composite heat source excitation and the node air temperature are synchronously input into the temperature change rate derivative prediction layer. The ordinary differential equation solver is called to perform continuous integration calculation on the time step according to the temperature change rate gradient output by the network, filter out the high-frequency fluctuation noise in the node air temperature, and invert the output to output the real state trajectory of the insulating surface with physical smooth response and phase hysteresis characteristics, which is denoted as the solid temperature sequence.

5. The time-series data analysis method for aging of metering box components according to claim 1, characterized in that: The process of obtaining the surface condensation duration sequence includes: Extract temperature and humidity data from the external meteorological time series data, calculate the dynamic critical condensation boundary curve under the interaction of cold air outside the box and hot air inside the box, and output the dynamic dew point; The reconstructed solid temperature sequence and the dynamic dew point are cross-referenced in the time domain. The intersection critical time point where the solid temperature value decreases and is lower than the dynamic dew point value is extracted as the starting point of liquid water film precipitation, and the intersection critical time point where the solid temperature value rises and is higher than the dynamic dew point value is extracted as the end point of water film evaporation. By accumulating the time difference between the precipitation start point and the evaporation end point, a surface condensation duration sequence is output to characterize the duration of continuous moisture immersion on the surface of the insulating material.

6. The time-series data analysis method for aging of metering box components according to claim 1, characterized in that: The process of sending dynamic peak shaving and load reduction commands includes: constructing a global monitoring and comparison matrix; when the insulation aging cumulative index of a spatial node exceeds a safety threshold, retrieving the historical thermal crosstalk sequence corresponding to the spatial node; by retrieving the spatial coordinates of the source node with the largest heat contribution value and the longest duration in the historical thermal crosstalk sequence, inversely locking the source device that generates the influence of nearby thermal radiation as the heavy-load heat-generating node; and generating a dynamic peak shaving and load reduction command for reducing the upper limit of the operating power of the heavy-load heat-generating node.

Images