A cable fire early warning method based on physical constraint double branch network

Abstract

The application provides a cable fire warning method based on a physical constraint double-branch network, and relates to the technical field of electric digital data processing. The method comprises the following steps: obtaining a historical operation data sequence and a future prior information sequence of a cable, inputting the two independent time sequence feature extraction branches respectively, extracting historical time sequence features and future prior features, and performing adaptive fusion; based on the fused features, a thermal key parameter at multiple future time steps is synchronously predicted, including a thermal accumulation index representing the degree of cable heat accumulation; in the prediction model training process, the joule heat effect and the heat accumulation effect are physically consistent constrained through a physical constraint loss function; and the thermal state indication information of the cable is generated according to the predicted thermal accumulation index. The method fuses the historical operation law and the future prior information through a double-branch architecture, ensures that the prediction result is physically reliable through physical constraint, drives risk assessment through the thermal accumulation index, and gives effective state indication information of the cable.

Description

Technical Field

[0001] This invention belongs to the field of electrical digital data processing technology, specifically relating to a cable fire early warning method based on a physically constrained dual-branch network. Background Technology

[0002] Thermal condition assessment of cables is a crucial step in ensuring the safe operation of power systems. During cable operation, abnormal temperature rises and heat accumulation caused by poor joint contact, insulation aging, or overload will undergo a gradual process from abnormal heat generation to continuous heat accumulation. This process can last for tens of minutes or even hours. Accurately sensing this process is essential for preventing accidents.

[0003] Currently, the assessment of cable thermal condition mainly relies on two types of schemes: one is the posterior feature alarm scheme based on temperature or gas sensors. This type of scheme can only trigger a response after the abnormal state has fully manifested, and cannot achieve forward-looking perception of the thermal runaway process; the other is the temperature prediction scheme based on pure data-driven time series models. This type of scheme only learns the statistical mapping relationship between input and output from historical data, and lacks explicit constraints on physical processes such as cable heating and heat accumulation. When the training samples are insufficient or the operating conditions exceed the distribution range of the training data, it is easy to produce prediction results that do not conform to the Joule heating effect and the laws of heat conduction, resulting in a decrease in prediction reliability.

[0004] Furthermore, neither of the above two approaches provides an effective indicator for accurately quantifying the degree of heat accumulation inside the cable. They generally rely on a single temperature threshold for judgment and lack a continuous quantitative assessment method for the gradual process from abnormal heat generation to heat accumulation. This results in a lack of effective status indication information when the temperature has not yet exceeded the limit but heat is already accumulating. Summary of the Invention

[0005] In view of the above-mentioned defects or deficiencies in the prior art, a cable fire early warning method based on a physically constrained dual-branch network is provided for quantitatively assessing the thermal state of cables, including the following steps: Acquire the historical operating data sequence and future prior information sequence of the cable; the historical operating data includes load current and cable temperature; the future prior information includes planned load and weather forecast. The historical running data sequence and the future prior information sequence are respectively input into two independent time-series feature extraction branches of the prediction model to obtain historical time-series features and future prior features; the historical time-series features and future prior features are adaptively fused to obtain fused features; Based on the fusion features, the thermal key parameters for multiple future time steps are predicted simultaneously; the thermal key parameters include the thermal accumulation index, which characterizes the degree to which heat accumulates in the cable over time. The training of the prediction model is guided by a physical constraint loss function, which is based at least on the predicted values ​​of the thermal key parameters and imposes physical consistency constraints on the Joule heating effect and the thermal accumulation effect. Based on the predicted heat accumulation index, thermal status indication information of the cable is generated.

[0006] According to the technical solution provided in this application, obtaining the historical operating data sequence of the cable includes the following steps: Multi-dimensional operating parameters are collected at a fixed sampling frequency by deploying various types of sensors in and around the cable; these multi-dimensional operating parameters include load current, operating voltage, active power, cable surface temperature, ambient temperature, and ambient relative humidity. The collected multi-dimensional operating parameters are aligned by timestamps to construct a multivariate time series matrix, which serves as the historical operating data sequence; wherein the time length of the historical operating data sequence is equal to the preset historical observation window length.

[0007] According to the technical solution provided in this application, the future prior information sequence is constructed by aligning the following three types of information in time: The first type of information is the future planned load current curve provided by the power dispatching department, whose time span covers the future preset forecast period; The second type of information is the future temperature forecast sequence provided by the meteorological department, whose time resolution is consistent with the time resolution of the future planned load current curve. The third type of information is the grid dispatch plan code, which includes at least one of the following: time code for planned maintenance events, time code for load transfer events, and time code for distributed generation integration events.

[0008] According to the technical solution provided in this application, both independent temporal feature extraction branches adopt temporal convolutional networks, and the two temporal convolutional networks have the same number of dilated causal convolutional layers and residual connection structure; Specifically, the first temporal feature extraction branch performs dilated causal convolution on the historical running data sequence along the forward time direction to extract the historical temporal features; The second temporal feature extraction branch performs an expanded causal convolution on the future prior information sequence along the forward time direction to extract the future prior features.

[0009] According to the technical solution provided in this application, the simultaneous prediction of key thermal parameters for multiple future time steps based on the fusion features is achieved through a multi-task prediction head. The multi-task prediction head consists of a first fully connected layer and a second fully connected layer; the first fully connected layer is shared by all prediction tasks, and the second fully connected layer is set independently for each prediction task. The multi-task prediction head takes the fusion feature as input and synchronously outputs thermal key parameters for a preset number of time steps in the future; wherein, the thermal key parameters include: predicted values ​​of load current, predicted values ​​of cable surface temperature, predicted values ​​of heat accumulation index, and predicted values ​​of fire risk value. The predicted value of the load current is the actual expected load value after considering random load disturbances and distributed power source fluctuations, based on the planned load in the future prior information sequence.

[0010] According to the technical solution provided in this application, the physical constraint loss function includes a Joule thermal constraint loss term and a thermal accumulation constraint loss term; The Joule thermal constraint loss term is constructed in the following manner: Based on the physical relationship that Joule thermal power is proportional to the square of the load current, the predicted future load current is substituted into the Joule thermal power calculation formula to estimate the theoretical temperature rise caused by the Joule thermal effect. The deviation between the theoretical temperature rise and the predicted future cable surface temperature is calculated to obtain the Joule thermal constraint loss term. The heat accumulation constraint loss term is constructed in the following way: The theoretical heat accumulation value under the simplified thermal model is obtained by numerically integrating the difference between the predicted future cable surface temperature and the predicted future ambient temperature along the time dimension. The deviation between the theoretical heat accumulation value and the heat accumulation index is calculated to obtain the heat accumulation constraint loss term.

[0011] According to the technical solution provided in this application, the physical constraint loss function further includes a resistance-temperature coupling constraint loss term and a monotonicity constraint loss term; The resistance-temperature coupling constraint loss term is constructed in the following manner: Based on the physical law that conductor resistance changes linearly with temperature, the theoretical resistance value is calculated using the temperature coefficient of resistance correction formula according to the predicted future cable surface temperature. The equivalent resistance value calculated from the predicted future load current and operating voltage is compared with the theoretical resistance value to obtain the resistance temperature coupling constraint loss term. The monotonicity constraint loss term is constructed in the following way: Calculate the partial derivative of the predicted future cable surface temperature with respect to the predicted future load current within the time interval during which the predicted future load current is either increasing or remaining in a steady state. When the partial derivative is negative, the absolute value of the partial derivative is used as a penalty term to obtain the monotonicity constraint loss term.

[0012] According to the technical solution provided in this application, generating thermal state indication information of the cable based on the predicted heat accumulation index includes the following steps: Based on predicted key thermal parameters for multiple future time steps, a multi-dimensional risk assessment indicator system is constructed; the multi-dimensional risk assessment indicator system includes: The risk index is calculated based on the predicted future cable surface temperature exceeding the preset safety threshold by a certain amount and duration; the risk index is calculated based on the predicted future cable surface temperature changing rate within a preset time period; the risk index is calculated based on the ratio of the heat accumulation index to the cable heat accumulation critical value; and the risk index is calculated based on the predicted future load current exceeding the cable's rated current carrying capacity. The entropy weight method is adopted to dynamically determine the weight coefficient of each risk indicator based on the information entropy of each risk indicator under the current working conditions. The risk indicators are weighted and integrated according to their corresponding weight coefficients to generate a comprehensive fire risk index. Based on the threshold range of the comprehensive fire risk index, a graded early warning signal is output.

[0013] According to the technical solution provided in this application, the Joule thermal constraint loss term and the resistance temperature coupling constraint loss term share the predicted value of the load current and the predicted value of the cable surface temperature output by the multi-task prediction head as inputs. The Joule thermal constraint loss term takes the predicted value of the load current as input, calculates the theoretical temperature rise caused by the Joule thermal effect, and calculates the deviation from the predicted value of the cable surface temperature. The resistance temperature coupling constraint loss term takes the predicted value of the cable surface temperature as input, calculates the theoretical resistance value corrected by the resistance temperature coefficient, and calculates the deviation between the theoretical resistance value and the equivalent resistance value calculated by the predicted value of the load current and the operating voltage. The predicted values ​​of the load current and the cable surface temperature are simultaneously constrained by two loss terms, and the two loss terms are constrained in opposite directions: the Joule thermal constraint loss term constrains the electrothermal propagation direction from current to temperature, and the resistance-temperature coupling constraint loss term constrains the thermoelectric propagation direction from temperature to resistance and then to current, forming a joint constraint verification of the same set of predicted variables in the two propagation directions of electrothermal.

[0014] According to the technical solution provided in this application, in the multi-task prediction head, the first fully connected layer is used to extract the common physical representation shared by each prediction task from the fusion features, and the common physical representation includes electrothermal coupling features and thermal state features. Each of the second fully connected layers extracts the specific features corresponding to the prediction task from the general physical representation and outputs the corresponding thermal key parameters; Four prediction tasks are output in parallel, and there is a hierarchical relationship between the output values ​​of the four prediction tasks in terms of physical laws: According to Joule's law of heating, the predicted value of the load current physically determines the heating power of the cable, which in turn affects the predicted value of the cable surface temperature; the difference between the predicted value of the cable surface temperature and the ambient temperature over time physically determines the predicted value of the heat accumulation index; the predicted value of the heat accumulation index physically reflects the degree of heat accumulation, which in turn determines the predicted value of the fire risk value.

[0015] Compared with the prior art, the beneficial effects of this application are as follows: First, by adaptively fusing historical operational data sequences and future prior information sequences into two independent time-series feature extraction branches, the prediction model can learn from historical operational patterns and incorporate future trend information from planned loads and weather forecasts. This allows it to maintain the accuracy of predicting key future thermal parameters even in critical scenarios such as sudden load changes or abrupt weather changes.

[0016] Second, by introducing a physical constraint loss function during model training that imposes physical consistency constraints on the Joule heating effect and the heat accumulation effect based at least on the predicted values ​​of key thermal parameters, the predicted load current and cable temperature satisfy Joule's law, and the predicted heat accumulation index is consistent with the definition of temperature-time integral. This ensures that the prediction results have physical credibility and will not produce unreasonable prediction results that violate basic physical laws, even under extreme conditions where training data is sparse.

[0017] Third, by predicting the heat accumulation index and generating thermal status indication information for the cable accordingly, early status indications reflecting the trend of thermal runaway can be provided even when the cable temperature has not yet reached the dangerous threshold but heat is already accumulating. This provides technical support for early detection of thermal safety status. Attached Figure Description

[0018] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 A flowchart illustrating the steps of the cable fire early warning method based on a physically constrained dual-branch network provided in this application; Figure 2 The diagram of the physical constraint-based dual-branch temporal neural network provided in this application. Detailed Implementation

[0019] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0020] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] As mentioned in the background section, this application proposes a cable fire early warning method based on a physically constrained dual-branch network, such as... Figure 1 As shown, it includes the following steps: S1. Obtain the historical operating data sequence and future prior information sequence of the cable; wherein, the historical operating data includes load current and cable temperature; the future prior information includes planned load and weather forecast; S2. Input the historical running data sequence and the future prior information sequence into two independent time-series feature extraction branches of the prediction model to obtain historical time-series features and future prior features; perform adaptive fusion of the historical time-series features and future prior features to obtain fused features; S3. Based on the fusion features, simultaneously predict the thermal key parameters for multiple future time steps; the thermal key parameters include the thermal accumulation index, which characterizes the degree to which the cable heat accumulates over time; The training of the prediction model is guided by a physical constraint loss function, which is based at least on the predicted values ​​of the thermal key parameters and imposes physical consistency constraints on the Joule heating effect and the thermal accumulation effect. S4. Generate thermal status indication information for the cable based on the predicted heat accumulation index.

[0022] Specifically, the method is divided into three stages: data acquisition, model building and training, and online prediction and early warning.

[0023] During the data acquisition phase, historical operational data sequences are read from the historical database of the cable monitoring system. Specifically, the length of the historical observation window is determined, representing the duration of past data that the model needs to backtrack for prediction. Then, all sampling records within this time range are extracted from the database, arranged from earliest to latest sampling time. The load current and cable temperature values ​​from each record are extracted and organized into a sequence along with other operational parameters. Load current refers to the actual current flowing through the cable, and cable temperature refers to the measured surface temperature of the cable. Future prior information sequences are obtained from the power dispatching system and meteorological data interfaces. Specifically, the prediction duration is determined, i.e., the future time span to be predicted. Then, the planned load curve data within this duration is exported from the dispatching system. Planned load refers to the expected load value pre-arranged by the dispatching department based on load forecasts and power generation plans. Simultaneously, temperature forecast data for the same period is obtained from the meteorological data interface. Weather forecasts refer to the environmental temperature forecast values ​​for various future time points issued by the meteorological department. The planned load data and temperature forecast data are then matched one-to-one by time point to form a sequence. These two types of data sequences are used as inputs to the two input branches of the prediction model.

[0024] During the model building and training phases, the prediction model comprises two structurally symmetrical but functionally independent temporal feature extraction branches. Both branches employ a temporal convolutional network consisting of stacked multi-layer dilated causal convolutions and residual connections. The first branch receives the historical execution data sequence and, starting from the first time step of the sequence, performs dilated causal convolution operations layer by layer in the direction of increasing time. The output of each convolution layer serves as the input to the next layer, ultimately outputting historical temporal features. The second branch receives the future prior information sequence and similarly performs dilated causal convolution operations layer by layer in the forward time direction, ultimately outputting future prior features. Both features are high-dimensional vectors, encoding the temporal pattern of historical execution regularity and the trend features of future prior information, respectively.

[0025] After obtaining the two features, an adaptive fusion operation is performed. Adaptive fusion is achieved through a gating mechanism. The specific steps are as follows: First, the historical time-series features and the future prior features are concatenated along their feature dimensions to obtain a concatenated feature vector. Then, this concatenated feature vector is input into a fully connected layer. The output of the fully connected layer is transformed by a sigmoid activation function to obtain a gating weight vector with values ​​between 0 and 1. Each element in this vector corresponds to a dimension of the feature vector. Finally, the two features are summed element-wise according to the gating weight vector. That is, each element in the fused feature equals the gating weight multiplied by the corresponding element of the future prior feature plus 1 minus the gating weight multiplied by the corresponding element of the historical time-series feature. The value of the gating weight vector is automatically determined by the input data. When the future prior information is consistent with historical patterns, the gating weight tends to increase the contribution of the future prior feature; when the uncertainty of the future prior information is high, the gating weight tends to increase the contribution of the historical time-series feature, thus achieving an adaptive balance between the two information sources. The fused feature serves as the input to the subsequent multi-task prediction head.

[0026] The multi-task prediction head consists of two fully connected layers. The first fully connected layer takes the fused features as input and outputs a common physical representation vector shared by all prediction tasks through matrix multiplication and nonlinear activation function transformation. The second fully connected layer contains four parallel independent fully connected layers, each performing matrix multiplication transformation on the common physical representation vector again, and each outputting a sequence of prediction variables. The four prediction variables are the predicted load current, the predicted cable surface temperature, the predicted heat accumulation index, and the predicted fire risk value. Each variable contains predicted values ​​for several equally spaced time steps from the current moment to a preset future time period. Among them, the heat accumulation index is a key physical quantity defined in this invention. Its physical meaning is the cumulative heat value obtained by integrating the difference between the cable temperature and the ambient temperature over time. It represents the net heat absorbed and accumulated by the cable over a period of time. The larger the value, the more severe the heat accumulation and the higher the fire risk.

[0027] The training of the prediction model is completed offline. Before training, a training dataset needs to be prepared. Each sample in the training dataset contains a segment of historical running data, a corresponding segment of future prior information, and a corresponding segment of future actual observation data. During training, the historical running data and future prior information are input into the model to obtain the predicted values ​​of the four prediction variables, and then the total loss function is calculated. The total loss function consists of two weighted sums. The first part is the data prediction loss, which uses the mean square error function to calculate the deviations between the predicted values ​​and the actual observation values ​​of the four prediction variables and then sums them weighted. The second part is the physical constraint loss function. The Joule heating effect constraint loss term is calculated by taking the predicted load current value output by the model, squaring it, multiplying it by the cable conductor resistance value to obtain the Joule heating power, then using the Joule heating power to estimate the theoretical temperature rise value based on the cable heat capacity parameters, and calculating the mean square error between the theoretical temperature rise value and the predicted cable surface temperature value output by the model to obtain this loss term. The loss term constrained by the heat accumulation effect is calculated as follows: the difference between the predicted cable surface temperature output by the model and the predicted ambient temperature at the corresponding time step in the future temperature forecast sequence is taken, and the difference is accumulated over time steps to obtain the theoretical heat accumulation value. The mean square error is then calculated between the theoretical heat accumulation value and the predicted heat accumulation index directly output by the model to obtain the loss term. After the total loss function is calculated, all learnable parameters of the model are updated using the gradient backpropagation algorithm, and this process is repeated until the loss function converges. Guided by the physical constraint loss function, the model is forced to comply with Joule's law of heating and the definition of heat accumulation while learning the statistical laws of the data, ensuring the physical reliability of the prediction results.

[0028] After model training, it is deployed and put into use. The operational steps of the online prediction and early warning phase are as follows: The latest historical operational data sequence is obtained in real time from the monitoring system, and the latest future prior information sequence is obtained from the scheduling system and meteorological interface. These are input into the trained prediction model, which outputs predicted values ​​of the heat accumulation index for multiple future time steps. The future fire risk value is calculated based on the predicted heat accumulation index value. Specifically, the predicted heat accumulation index value is compared with the heat accumulation critical value determined according to the cable type, and the ratio is taken as the basic risk score. Simultaneously, the basic risk score is corrected based on whether the temperature prediction value exceeds a preset safety threshold, ultimately yielding a fire risk value between 0 and 1. A yellow warning is issued when the fire risk value is in the first preset range, indicating a possible abnormal temperature rise within the next 30 to 60 minutes; an orange warning is issued when it is in the second preset range, indicating a significant heat risk within the next 15 to 30 minutes; and a red warning is issued when it is in the third preset range, indicating an extremely high fire risk within the next 5 to 15 minutes. Because the heat accumulation index can capture dangerous precursor states where heat has been continuously accumulating even before the temperature exceeds the limit, this method can provide early warning tens of minutes before a fire actually occurs.

[0029] The technical principle of this solution lies in its dual-branch architecture, which organically integrates historical operational patterns and future a priori trends at the feature level. It embeds the laws of electrothermal physics into the model training process through a physical constraint loss function and uses the heat accumulation index as a leading criterion for fire risk, thus constructing a complete causal reasoning chain from operational condition perception to thermal state prediction and risk warning. The technical benefits include prediction results that combine data fitting accuracy with physical consistency, and a significantly earlier warning window compared to traditional smoke alarms and temperature threshold solutions.

[0030] In a preferred embodiment, obtaining the historical operating data sequence of the cable includes the following steps: Multi-dimensional operating parameters are collected at a fixed sampling frequency by deploying various types of sensors in and around the cable; these multi-dimensional operating parameters include load current, operating voltage, active power, cable surface temperature, ambient temperature, and ambient relative humidity. The collected multi-dimensional operating parameters are aligned by timestamps to construct a multivariate time series matrix, which serves as the historical operating data sequence; wherein the time length of the historical operating data sequence is equal to the preset historical observation window length.

[0031] Specifically, in implementation, the first step is to deploy a sensor network in the cable and its installation environment. This sensor network includes the following types: current transformers at both ends or intermediate joints of the cable for measuring load current; voltage transformers at both ends of the cable for measuring operating voltage; power transmitters in the distribution cabinet for measuring active power; distributed fiber optic temperature measurement systems laid along the cable surface or thermocouples installed at critical nodes for measuring cable surface temperature; ambient temperature sensors installed in areas within the cable laying channel not directly affected by the cable's own heating for measuring ambient temperature; and humidity sensors installed in the same area for measuring relative humidity. All sensors are connected to a unified data acquisition terminal and set to the same fixed sampling frequency, such as 1 Hz, i.e., synchronously acquiring data once per second.

[0032] During each sampling period, the data acquisition terminal synchronously sends acquisition commands to each sensor, and each sensor returns its current measurement value. The acquisition terminal combines the load current value, operating voltage value, active power value, cable surface temperature value, ambient temperature value, and ambient relative humidity value collected at the same time into a six-dimensional vector, and appends the current sampling timestamp to this vector. Each sampling cycle generates a six-dimensional data vector with a timestamp.

[0033] When constructing the historical operational data sequence, the first step is to determine the preset historical observation window length, which is measured by the number of sampling points. For example, a window length of 1800 sampling points corresponds to 30 minutes of historical data. Then, a six-dimensional data vector containing the most recent window length is retrieved from the historical database in reverse chronological order by timestamp. This vector is then sorted in ascending chronological order by timestamp, resulting in a matrix with the number of rows equal to the window length and the number of columns equal to 6. This matrix is ​​the multivariate time series matrix, serving as the historical operational data sequence. Each row of the matrix corresponds to a sampling time, and each column corresponds to a parameter dimension. Before inputting the data into the prediction model, each column of the matrix needs to be normalized. Normalization can be achieved by subtracting the mean from each column's data and then dividing by its standard deviation, ensuring that the data across all dimensions have similar numerical ranges, which is beneficial for the stability of model training.

[0034] The technical principle of this solution lies in unifying the three heterogeneous physical quantities of cable operation status—electrical, thermal, and environmental quantities—into a standardized input format through synchronous acquisition and matrix organization of multiple types of sensors. The technical effect is that the selection of six-dimensional parameters covers the core physical quantities required for electrothermal coupling analysis, and the model can simultaneously learn from historical data the patterns of load changes, thermal state evolution, and the influence of environmental conditions.

[0035] In a preferred embodiment, the sequence of future prior information is constructed by aligning the following three types of information in time: The first type of information is the future planned load current curve provided by the power dispatching department, whose time span covers the future preset forecast period; The second type of information is the future temperature forecast sequence provided by the meteorological department, whose time resolution is consistent with the time resolution of the future planned load current curve. The third type of information is the grid dispatch plan code, which includes at least one of the following: time code for planned maintenance events, time code for load transfer events, and time code for distributed generation integration events.

[0036] Specifically, in practice, the construction of the future prior information sequence requires the sequential acquisition of three types of information and time alignment operations.

[0037] The first type of information acquisition steps involve calling the planned load query interface from the power dispatch automation system, inputting the line number of the monitored cable and the query time range. The start point of the query time range is the current time, and the end point is the current time plus the predicted duration. The dispatch system returns the planned load current value sequence for each time period within that time range. The power dispatch department refers to the dispatching agency responsible for the operation and load management of the power grid in the area to which the cable belongs. The future planned load current curve is a load arrangement pre-prepared by the dispatch department based on day-ahead load forecasts, generation plans, and safety constraints, providing the expected load current value for the line at fixed time intervals for each period.

[0038] The steps for obtaining the second type of information are as follows: From the meteorological data service platform or meteorological bureau data interface, call the temperature forecast query interface, input the geographical coordinates of the cable laying area and the query time range, which must be completely consistent with the time range of the first type of information. The meteorological platform returns the environmental temperature forecast numerical sequence for each time point within that time range. It is necessary to ensure that the time resolution of the obtained temperature forecast sequence is the same as the time resolution of the planned load current curve of the first type of information. If the resolutions are inconsistent, interpolation processing is required to ensure that the two strictly correspond at each time point.

[0039] The steps for obtaining the third type of information are as follows: Query the dispatch plan events related to the monitored cable from the power grid dispatch plan management system or power outage management system. The query time range is the same as for the first type of information. For each queried plan event, extract its event type and planned execution time. Planned maintenance events refer to planned power outage maintenance operations scheduled for the cable or its connected equipment. The planned execution time is converted into a corresponding time step index within the predicted duration and represented by a preset first coding value. Load transfer events refer to operations that switch the load that should be borne by this cable to other lines. The planned execution time is converted into a time step index and represented by a preset second coding value. Distributed power source integration events refer to the planned commissioning or decommissioning of distributed photovoltaic, energy storage, or other power sources connected to the cable. The planned execution time is converted into a time step index and represented by a preset third coding value. Time steps without events are filled with zero-value codes. Arrange the coding values ​​of each time step in chronological order to obtain the dispatch plan coding sequence.

[0040] After acquiring the three types of information, a time alignment operation is performed to construct a sequence of future prior information. Specifically, using the time axis of the planned load current curve in the first type of information as a reference, each time point of the temperature forecast sequence in the second type of information is matched and aligned with the reference time axis, and each time step of the scheduling plan coding sequence in the third type of information is also matched and aligned with the reference time axis. At each time point, the planned load current value, temperature forecast value, and scheduling plan coding value are organized into a three-dimensional vector. All three-dimensional vectors at all time points are arranged in chronological order to form the sequence of future prior information. The number of rows in this sequence equals the number of time steps within the prediction duration, and the number of columns is 3.

[0041] The technical principle of this solution lies in organizing three types of known future information from different sources into a sequence along a unified timeline. This allows the prediction model to simultaneously reference three types of prior knowledge—load change trends, ambient temperature change trends, and structural operational mode changes—when inferring future states. The technical effect is that when cables face abnormal operating conditions such as planned overload, sudden weather changes, or adjustments to operating modes, the model can anticipate trend changes and predict thermal risks based on prior information, solving the problem of severe prediction lag in pure historical data extrapolation schemes during sudden changes in operating conditions.

[0042] In a preferred embodiment, both independent temporal feature extraction branches employ temporal convolutional networks, and the two temporal convolutional networks have the same number of dilated causal convolutional layers and residual connection structures; Specifically, the first temporal feature extraction branch performs dilated causal convolution on the historical running data sequence along the forward time direction to extract the historical temporal features; The second temporal feature extraction branch performs an expanded causal convolution on the future prior information sequence along the forward time direction to extract the future prior features.

[0043] In implementation, both independent temporal feature extraction branches are constructed using temporal convolutional networks. A temporal convolutional network is a convolutional neural network structure specifically designed for processing sequential data; its core components include dilated causal convolutional layers and residual connection modules. For example... Figure 2 As shown, the two branches of the dual-branch temporal neural network maintain complete consistency in the number of network layers, the number of convolutional kernels per layer, the size of the convolutional kernels, and the dilation coefficient settings. They also use the same number of dilated causal convolutional layers and residual connection structures to ensure that the features extracted by the two branches match in dimension, which facilitates subsequent fusion operations.

[0044] The first temporal feature extraction branch operates by taking the historical data sequence as its input. This sequence is a multivariate time series matrix, where each row corresponds to a historical sampling time and each column corresponds to a running parameter dimension. The first temporal feature extraction branch processes the sequence layer by layer along the forward time direction, meaning it proceeds from the first time step to the last, following the sampling time from earliest to latest. In each layer of dilated causal convolution, the convolution kernel skips several intermediate time steps and performs convolution operations on the input sequence according to a preset dilation coefficient. The dilation coefficient increases exponentially with the number of layers, allowing shallower layers to capture short-term temporal dependencies and deeper layers to capture long-term temporal dependencies. The characteristics of causal convolution ensure that when calculating the output at any time step, the convolution kernel only sees information from that time step and previous time steps, without using information from future times, conforming to the temporal causality law in the physical world where cause precedes effect. After each convolutional operation, the output is transformed by a nonlinear activation function and added to the input of that layer through a residual connection. The residual connection directly adds the input of the convolutional layer to its output to alleviate the vanishing gradient problem during deep network training. After layer-by-layer processing by all dilated causal convolutional layers and residual connection modules, the first branch outputs historical time-series features. This feature is a high-dimensional vector that encodes the evolution of historical operating parameters such as load current and cable temperature over time and the correlation patterns between them.

[0045] The second time-series feature extraction branch operates by using the future prior information sequence as its input. This future prior information sequence is also a time-ordered sequence matrix, where each row corresponds to a future time step, and each column corresponds to information in three dimensions: planned load current, temperature forecast, and scheduling plan encoding. The second time-series feature extraction branch also processes this sequence layer by layer along the forward time direction using dilated causal convolution and residual connections, with the processing flow and network structure identical to the first branch. Although the input sequence of the second branch corresponds to future times in physical time, the dilated causal convolution still follows the rule of using only the current position and previous position information during sequence processing. Since the value of each time step in the future prior information sequence is a known planned value and forecast value at the current time, rather than an unknown quantity that the model needs to extrapolate from history, using causal convolution to process this sequence will not cause information leakage or logical contradictions. After processing through all layers, the second branch outputs future prior features, which are high-dimensional vectors with the same dimensions as the historical time-series features, encoding information such as future planned load change trends, temperature forecast change trends, and the impact of scheduling events on the operating mode.

[0046] After the two branches output historical time-series features and future prior features, since the two features have the same dimension, subsequent adaptive fusion operations can be performed directly.

[0047] The technical principle of this solution lies in employing two symmetrical temporal convolutional networks to process historical operational data and future prior information respectively. The dilated causal convolution mechanism enables the network to capture long-range temporal dependencies with a relatively small number of parameters, while the residual connection mechanism ensures the training stability of the deep network. The technical benefits are that the symmetrical structure of the two branches ensures the consistency of the output feature dimensions, facilitating subsequent fusion. Simultaneously, the independent parameters of each branch allow them to focus on learning two different types of information—historical patterns and future trends—avoiding mutual interference when a single network processes two types of heterogeneous information simultaneously.

[0048] In a preferred embodiment, the simultaneous prediction of key thermal parameters for multiple future time steps based on the fused features is achieved through a multi-task prediction head. The multi-task prediction head consists of a first fully connected layer and a second fully connected layer; the first fully connected layer is shared by all prediction tasks, and the second fully connected layer is set independently for each prediction task. The multi-task prediction head takes the fused features as input and synchronously outputs thermal key parameters for a preset number of time steps in the future; wherein, the thermal key parameters include: predicted values ​​of load current, predicted values ​​of cable surface temperature, predicted values ​​of heat accumulation index, and predicted values ​​of fire risk value. The predicted value of the load current is the actual expected load value after considering random load disturbances and distributed power source fluctuations, based on the planned load in the future prior information sequence.

[0049] In implementation, the multi-task prediction head receives the fused features obtained through adaptive fusion as input. The multi-task prediction head consists of two stacked fully connected network layers, with the first and second fully connected layers connected by a non-linear activation function.

[0050] The first fully connected layer operates by taking the fused feature vector as input, performing matrix multiplication with this vector using a learnable weight matrix, and then adding a learnable bias vector to obtain an intermediate representation vector. This intermediate representation vector, after being transformed by a nonlinear activation function, serves as the shared input to the second fully connected layer. The parameters of the first fully connected layer are shared by all prediction tasks. This means that the four prediction tasks use the same weight matrix and bias vector in the first layer, allowing the first fully connected layer to learn a general physical representation related to electrothermal coupling and thermal state evolution, rather than features specific to any particular task.

[0051] The second fully connected layer consists of four independent sublayers, each corresponding to a prediction task. Each of the four sublayers has its own independent weight matrix and bias vector, and they do not share parameters. Each sublayer operates by taking the intermediate representation vector output from the first fully connected layer as input, performing matrix multiplication through its own weight matrix, and adding its own bias vector to output a sequence of predicted values ​​for the corresponding predicted variable. The four sublayers output predicted values ​​for load current, cable surface temperature, heat accumulation index, and fire risk, respectively. Each predicted variable contains a preset number of predicted values ​​for future time steps, equal to the prediction duration divided by the sampling time interval. The outputs of the four sublayers maintain a consistent number of time steps, and the predicted values ​​for the four predicted variables correspond one-to-one at each time step.

[0052] The predicted load current refers to the actual expected load value after considering random load disturbances and distributed power source fluctuations, based on the planned load current in the future prior information sequence. The planned load current is an idealized arrangement value compiled by the dispatching department based on load forecasts. However, in actual operation, the load fluctuates randomly, and the output of distributed photovoltaic and other power sources is also uncertain due to weather conditions. Therefore, the model needs to predict a load current value closer to reality based on the planned load. The predicted cable surface temperature refers to the temperature value of the cable surface predicted by the model at each future time step under the predicted load current and ambient temperature conditions. The predicted heat accumulation index is the predicted physical quantity directly output by the model, characterizing the degree of heat accumulation in the cable. The predicted fire risk value is the comprehensive score predicted value directly output by the model, characterizing the probability of fire occurring at each future time step.

[0053] The four prediction tasks are carried out synchronously during training, sharing the general physical representation extracted from the first fully connected layer. Knowledge transfer is achieved through multi-task joint training. That is, the electrothermal coupling law learned by the load current prediction task can assist the cable surface temperature prediction task, and the thermal state evolution law learned by the temperature prediction task can assist the prediction of heat accumulation index and fire risk value.

[0054] The technical principle of this solution lies in its structural design, which uses a shared first fully connected layer to extract general physical representations and an independent second fully connected layer to extract task-specific features. This allows multiple related prediction tasks to retain their individual flexibility while sharing underlying physical knowledge. The technical benefits are that the electrothermal coupling and thermal state features extracted by the shared layer can be reused by multiple tasks, improving the prediction accuracy and generalization ability of each task. Meanwhile, the independent second layer ensures that each task is optimized specifically for its respective prediction target.

[0055] In a preferred embodiment, the physical constraint loss function includes a Joule thermal constraint loss term and a thermal accumulation constraint loss term; The Joule thermal constraint loss term is constructed in the following manner: Based on the physical relationship that Joule thermal power is proportional to the square of the load current, the predicted future load current is substituted into the Joule thermal power calculation formula to estimate the theoretical temperature rise caused by the Joule thermal effect. The deviation between the theoretical temperature rise and the predicted future cable surface temperature is calculated to obtain the Joule thermal constraint loss term. The heat accumulation constraint loss term is constructed in the following way: The theoretical heat accumulation value under the simplified thermal model is obtained by numerically integrating the difference between the predicted future cable surface temperature and the predicted future ambient temperature along the time dimension. The deviation between the theoretical heat accumulation value and the heat accumulation index is calculated to obtain the heat accumulation constraint loss term.

[0056] In practice, the physical constraint loss function is an important component of the total loss function during the training of the prediction model. It consists of two parts: the Joule thermal constraint loss term and the thermal accumulation constraint loss term. The two parts are calculated independently and then summed with weights.

[0057] The Joule thermal constraint loss term is constructed in the following three steps. First, the predicted load current values ​​for each future time step output by the multi-task prediction head are taken. The square of the predicted load current value for each time step is multiplied by the resistance per unit length of the cable conductor to obtain the predicted Joule thermal power value for each future time step. The physical relationship that Joule thermal power is proportional to the square of the current originates from Joule's law and is the most fundamental quantitative relationship of electrothermal coupling. Second, based on the cable's thermal capacity parameters and the time interval length of each time step, the predicted Joule thermal power value for each time step is converted into the theoretical temperature rise increment caused by the Joule thermal effect within that time step. The theoretical temperature rise increment of each time step is accumulated over time to obtain the cumulative theoretical temperature rise caused by the Joule thermal effect at each time step. Third, the deviation between the cumulative theoretical temperature rise of each time step and the predicted cable surface temperature value for the corresponding time step output by the multi-task prediction head is calculated. The sum of the squares of the differences between the two values ​​for each time step is divided by the rated power of the cable for normalization, resulting in the Joule thermal constraint loss term. The purpose of normalization by dividing by the cable's rated power is to eliminate the influence of different cable capacities on the magnitude of the loss term, making the loss term comparable between cables of different specifications. The physical meaning of this loss term is that the predicted load current should generate a corresponding temperature rise through the Joule heating effect. If the predicted current is large but the predicted temperature is low, it indicates that the prediction violates Joule's law, and the value of this loss term will increase, driving the model to correct this inconsistency during subsequent training.

[0058] The construction of the heat accumulation constraint loss term involves the following three steps: First, the predicted cable surface temperature at each future time step is obtained from the multi-task prediction head output. Simultaneously, the predicted ambient temperature at each corresponding time step is obtained from the future temperature forecast sequence. The predicted cable surface temperature at each time step is subtracted from the predicted ambient temperature at that time step to obtain the temperature difference at each time step. Second, the temperature difference at each time step is numerically integrated along the time dimension. Numerical integration involves multiplying the temperature difference at each time step by the time step length and then accumulating it step by step from the first time step to obtain the theoretical heat accumulation value under the simplified thermal model at each time step. The simplified thermal model ignores the dynamic term of cable heat dissipation and uses only the time integral of the temperature difference as an approximate measure of heat accumulation. This simplification has engineering-acceptable accuracy within the short-term prediction window. Third, the deviation between the theoretical heat accumulation value at each time step and the predicted heat accumulation index at the corresponding time step output by the multi-task prediction head is calculated. The deviation calculation also uses the mean square error form to obtain the heat accumulation constraint loss term. The physical meaning of this loss term is that the heat accumulation index is defined as the accumulation of the temperature difference between the temperature and the ambient temperature. Therefore, the predicted value of the heat accumulation index directly output by the model should be consistent with the theoretical value calculated by the temperature difference integral. The larger the deviation, the less the model's prediction of heat accumulation conforms to its physical definition.

[0059] In the total loss function of model training, the Joule heat constraint loss term and the heat accumulation constraint loss term are multiplied by preset weight coefficients and then added to the data prediction loss to form the total loss function. The model minimizes both the data prediction loss and the physical constraint loss through gradient backpropagation, so that the predicted value output by the trained model not only fits the statistical regularity in the training data, but also satisfies the physical consistency required by Joule's heat law and the definition of heat accumulation.

[0060] The technical principle behind this solution lies in explicitly embedding Joule's law of heat and the definition of heat accumulation—two core physical laws—in the form of a loss function into the model training process. This allows the neural network to learn not only statistical mapping relationships from data but also predictive patterns that conform to objective laws from physical equations. The technical effect is that the physical constraint loss function provides the model with regularization guidance independent of training sample labels. In scenarios where training data is sparse or operating conditions exceed the training distribution range, physical constraints can regulate the model's prediction space from a physical mechanism perspective, avoiding unreasonable prediction results that violate the laws of energy conservation and the definition of heat accumulation.

[0061] In a preferred embodiment, the physical constraint loss function further includes a resistance-temperature coupling constraint loss term and a monotonicity constraint loss term; The resistance-temperature coupling constraint loss term is constructed in the following manner: Based on the physical law that conductor resistance changes linearly with temperature, the theoretical resistance value is calculated using the temperature coefficient of resistance correction formula according to the predicted future cable surface temperature. The equivalent resistance value calculated from the predicted future load current and operating voltage is compared with the theoretical resistance value to obtain the resistance temperature coupling constraint loss term. The monotonicity constraint loss term is constructed in the following way: Calculate the partial derivative of the predicted future cable surface temperature with respect to the predicted future load current within the time interval during which the predicted future load current is either increasing or remaining in a steady state. When the partial derivative is negative, the absolute value of the partial derivative is used as a penalty term to obtain the monotonicity constraint loss term.

[0062] In implementation, the construction of the resistance-temperature coupling constraint loss term involves the following three steps. First, the predicted cable surface temperature for each future time step is obtained from the multi-task prediction head. Based on the physical law that conductor resistance changes linearly with temperature, the predicted cable surface temperature for each time step is substituted into the resistance temperature coefficient correction formula for calculation. The resistance temperature coefficient correction formula is as follows: using the conductor's resistance at a reference temperature as a baseline value, add the temperature change multiplied by the resistance temperature coefficient to obtain the resistance value at the current temperature. The resistance temperature coefficient is an inherent property parameter of the conductor material; for copper conductors, it is typically taken as 3.93 per thousandth of a degree Celsius, and for aluminum conductors, it is typically taken as 4.03 per thousandth of a degree Celsius. The theoretical resistance value for each time step is calculated using this formula. Second, the predicted load current for each time step is obtained from the multi-task prediction head, and the measured operating voltage value at the current moment is also obtained. The operating voltage value is divided by the predicted load current value to obtain the equivalent resistance value for each time step. The third step involves calculating the deviation between the equivalent resistance value at each time step and the theoretical resistance value calculated in the first step. The deviation is calculated using the mean square error (MSE) to obtain the resistance-temperature coupling constraint loss term. The physical meaning of this loss term is that the resistance of the cable conductor changes with temperature. Therefore, the predicted temperature value and the equivalent resistance value calculated from the current and voltage should satisfy a resistance-temperature coupling relationship. A larger deviation indicates a discrepancy between the temperature prediction and the resistance change, violating the thermoelectric coupling law.

[0063] The monotonicity constraint loss term is constructed in the following three steps. First, from the sequence of predicted load current values ​​for each future time step output by the multi-task prediction head, identify the time intervals where the load current shows an upward trend or remains steady. An upward trend is determined by comparing the predicted load current values ​​of adjacent time steps; if the current value of the later time step is greater than the previous time step, it is considered an upward trend. A steady state is determined by the current value change of adjacent time steps being less than a preset small threshold. The cooling phase, i.e., the time interval where the load current shows a downward trend, is not included in the monotonicity constraint calculation because when the load current decreases, the cable may still maintain a relatively high temperature due to thermal inertia. The temperature not immediately decreasing with the current decrease is a normal physical phenomenon and should not be penalized. Second, within the identified time intervals of upward trend and steady state, calculate the partial derivative of the predicted cable surface temperature with respect to the predicted load current value. The partial derivative is approximated by dividing the temperature difference between adjacent time steps by the current difference. The third step involves checking the partial derivative values ​​at each time step. When the partial derivative is negative, it indicates that the load current has increased, but the predicted temperature has decreased. This violates the physical causal relationship that increased current leads to increased heat generation, which in turn drives up the temperature. In this case, the absolute value of the partial derivative is included as a penalty in the monotonicity constraint loss term. When the partial derivative is non-negative, no penalty is applied. The monotonicity constraint loss term is the sum or average of the penalty values ​​for all penalized time steps.

[0064] In the total loss function of model training, the Joule thermal constraint loss term, the thermal accumulation constraint loss term, the resistance-temperature coupling constraint loss term, and the monotonicity constraint loss term are multiplied by their respective preset weight coefficients and then added to the data prediction loss to form the total loss function. The four physical constraints exert physical consistency guidance on the prediction model from different perspectives, jointly ensuring the physical credibility of the prediction results.

[0065] The technical principle of this solution lies in the fact that the resistance-temperature coupling constraint constrains the thermoelectric relationship based on the physical law of resistance changing with temperature, while the monotonicity constraint constrains the direction of monotonicity prediction based on the causal relationship between current and temperature. Together with the Joule thermal constraint and the heat accumulation constraint, they constitute a four-dimensional physical constraint system covering electrothermal coupling, the definition of heat accumulation, the resistance-temperature relationship, and causal monotonicity. The technical effect is that the combined effect of the four physical constraints allows the model to achieve a better balance between data fitting and physical laws. In particular, the exclusion of the cooling phase in the monotonicity constraint avoids erroneous penalties for the normal thermal inertia cooling process, ensuring the rationality of the training.

[0066] In a preferred embodiment, generating thermal status indication information of the cable based on the predicted heat accumulation index includes the following steps: Based on predicted key thermal parameters for multiple future time steps, a multi-dimensional risk assessment indicator system is constructed; the multi-dimensional risk assessment indicator system includes: The risk index is calculated based on the predicted future cable surface temperature exceeding the preset safety threshold by a certain amount and duration; the risk index is calculated based on the predicted future cable surface temperature changing rate within a preset time period; the risk index is calculated based on the ratio of the heat accumulation index to the cable heat accumulation critical value; and the risk index is calculated based on the predicted future load current exceeding the cable's rated current carrying capacity. The entropy weight method is adopted to dynamically determine the weight coefficient of each risk indicator based on the information entropy of each risk indicator under the current working conditions. The risk indicators are weighted and integrated according to their corresponding weight coefficients to generate a comprehensive fire risk index. Based on the threshold range of the comprehensive fire risk index, a graded early warning signal is output.

[0067] In practice, this step is performed after the prediction model has completed the prediction of key thermal parameters for multiple future time steps. It includes four sub-steps: risk indicator system construction, dynamic weighting using entropy weighting, multi-indicator weighted fusion, and hierarchical early warning output.

[0068] The first step is to construct a multi-dimensional risk assessment indicator system based on the predicted key thermal parameters for multiple future time steps. This multi-dimensional risk assessment indicator system refers to an assessment framework composed of multiple quantitative indicators reflecting cable fire risk from different perspectives. This solution specifically includes the following four risk indicators.

[0069] The temperature over-limit risk index is calculated by iterating through the predicted cable surface temperature values ​​for each future time step and comparing the predicted temperature value at each time step with a preset safety threshold. The preset safety threshold is set according to the temperature resistance rating of the cable insulation material; for example, it is typically set to 70 degrees Celsius for cross-linked polyethylene insulated cables. For time steps exceeding the preset safety threshold, the magnitude of the temperature prediction exceeding the threshold is calculated, and the duration of the exceedance is recorded. The magnitude and duration of the exceedance are then combined for a comprehensive calculation, for example, by multiplying the maximum exceedance magnitude by the percentage of the exceedance duration, to obtain the temperature over-limit risk index. The physical meaning of this index is that the greater the severity and the longer the duration of the temperature exceeding the safe tolerance range of the insulation material, the greater the risk of insulation degradation.

[0070] The temperature rise rate risk index is calculated as follows: The predicted cable surface temperature for each future time step is taken, the temperature difference between adjacent time steps is calculated, and the temperature difference is divided by the time step length to obtain the temperature rise rate for each time period. The maximum temperature rise rate across all time periods or the average temperature rise rate within a certain time window is taken as the temperature rise rate risk index. The physical meaning of this index is that the faster the temperature rises, the shorter the time window for maintenance personnel to take emergency measures, and the higher the urgency of thermal runaway.

[0071] The heat accumulation risk index is calculated by taking the predicted heat accumulation index values ​​for each future time step, dividing each predicted value by the corresponding heat accumulation critical value for the cable model, and obtaining the heat accumulation ratio for each time step. The heat accumulation critical value refers to the maximum accumulated heat that the cable can safely withstand, pre-calculated based on parameters such as the specific heat capacity, mass, and heat dissipation conditions of the cable insulation material. When the heat accumulation index exceeds this critical value, the cable is in a high-risk state of thermal runaway. The maximum value of the heat accumulation ratio across all time steps is taken as the heat accumulation risk index. The physical meaning of this index is that even if the cable temperature has not yet exceeded the safety threshold, if heat continues to accumulate over a long period and the heat accumulation index gradually approaches the critical value, it means that the internal thermal state of the cable is deteriorating, and there is a potential risk of thermal runaway.

[0072] The load risk index is calculated by taking the predicted load current values ​​for each future time step, dividing the predicted load current value for each time step by the cable's rated current carrying capacity, and obtaining the load ratio for each time step. The rated current carrying capacity refers to the maximum current value that the cable is allowed to carry for a long period under specified laying conditions, and is determined by the cable's conductor cross-sectional area, insulation material, and laying method. The maximum value of the load ratio among all time steps, or the percentage of cumulative time exceeding 1, is taken as the load risk index. The physical meaning of this index is that a load current exceeding the cable's design carrying capacity will lead to a sharp increase in heat generation, accelerating insulation aging and thermal runaway.

[0073] The second step involves using the entropy weighting method to dynamically determine the weight coefficients of each risk indicator based on its information entropy under the current operating conditions. The entropy weighting method is an objective weighting approach that determines the weight based on the amount of information in each indicator's data; indicators with greater information content are assigned higher weights. Specifically, an evaluation matrix is ​​first constructed. For multiple sets of samples generated from consecutive predictions within the current time window, the values ​​of four risk indicators for each set of samples are calculated, forming an evaluation matrix with samples as rows and indicators as columns. Then, each column of the evaluation matrix is ​​normalized using the minimum-maximum normalization method to map each indicator value to the interval between zero and one, eliminating the influence of different indicator dimensions. Next, the information entropy value of each indicator is calculated. This is done by first calculating the proportion of each indicator value in each sample within its respective column, and then calculating the entropy value of each column according to the information entropy formula. A smaller entropy value indicates a greater discriminative power of the indicator among different samples and contains more information. Finally, the entropy weight of each indicator is calculated based on the information entropy value. The calculation method is to subtract the entropy value of each indicator from one to obtain the information utility value, and then divide the information utility value of each indicator by the sum of the information utility values ​​of all indicators to obtain the weight coefficient of each indicator. The sum of all weight coefficients equals one.

[0074] The third step involves weighting and merging the various risk indicators according to their corresponding weighting coefficients to generate a comprehensive fire risk index. Specifically, the four risk indicator values ​​calculated at the current moment are multiplied by their respective weighting coefficients determined using the entropy weighting method. The four weighted index values ​​are then summed to obtain the comprehensive fire risk index. The comprehensive fire risk index is a dimensionless value between zero and one; a higher value indicates a higher fire risk.

[0075] The fourth step is to output tiered early warning signals based on the threshold range of the comprehensive fire risk index. The tiered early warning includes at least the following three levels: When the comprehensive fire risk index is in the first threshold range, a Level 1 warning is output. This warning indicates that an abnormal temperature rise in the cable is predicted within the next 30 to 60 minutes. In this case, it is recommended that maintenance personnel increase monitoring frequency, check the rationality of the load plan, and pay attention to temperature change trends. When the comprehensive fire risk index is in the second threshold range, a Level 2 warning is output. This warning indicates that a significant thermal risk to the cable is predicted within the next 15 to 30 minutes. In this case, it is recommended that maintenance personnel adjust the load or increase active cooling measures and arrange on-site inspections. When the comprehensive fire risk index is in the third threshold range, a Level 3 warning is output. This warning indicates that an extremely high fire risk is predicted within the next 5 to 15 minutes. In this case, it is recommended to immediately take measures to reduce the load or cut off power and activate the emergency response procedure. In addition to the current risk level, the early warning output information also includes the predicted over-temperature time (the estimated number of minutes until the temperature first exceeds the safe threshold), the hotspot location (the area with the highest temperature based on the spatial distribution of sensors), the risk evolution trend curve (the curve of the change of each risk indicator and the comprehensive risk index over time), and the recommended disposal measures.

[0076] The technical principle of this solution lies in comprehensively characterizing cable fire risks from four dimensions: temperature exceeding limits, temperature rise rate, heat accumulation, and load overload. It uses an entropy weighting method to dynamically allocate weights based on the current information content of each indicator, avoiding the subjectivity of fixed thresholds. A single comprehensive index is generated through weighted fusion, facilitating tiered judgment. The technical effect is that the complementary nature of the multi-dimensional indicators effectively distinguishes between normal temperature rises caused by load fluctuations and dangerous temperature rises caused by insulation degradation. The introduction of the heat accumulation indicator allows for early warning triggering even when the temperature has not yet exceeded limits but heat is continuously accumulating, achieving proactive early warning.

[0077] In a preferred embodiment, the Joule thermal constraint loss term and the resistance temperature coupling constraint loss term share the predicted value of the load current and the predicted value of the cable surface temperature output by the multi-task prediction head as inputs; The Joule thermal constraint loss term takes the predicted value of the load current as input, calculates the theoretical temperature rise caused by the Joule thermal effect, and calculates the deviation from the predicted value of the cable surface temperature. The resistance temperature coupling constraint loss term takes the predicted value of the cable surface temperature as input, calculates the theoretical resistance value corrected by the resistance temperature coefficient, and calculates the deviation between the theoretical resistance value and the equivalent resistance value calculated by the predicted value of the load current and the operating voltage. The predicted values ​​of the load current and the cable surface temperature are simultaneously constrained by two loss terms, and the two loss terms are constrained in opposite directions: the Joule thermal constraint loss term constrains the electrothermal propagation direction from current to temperature, and the resistance-temperature coupling constraint loss term constrains the thermoelectric propagation direction from temperature to resistance and then to current, forming a joint constraint verification of the same set of predicted variables in the two propagation directions of electrothermal.

[0078] During implementation, the Joule thermal constraint loss term and the resistance-temperature coupling constraint loss term are calculated simultaneously in each training iteration, and they share the predicted values ​​of the load current and cable surface temperature output by the multi-task prediction head as common inputs.

[0079] The calculation of the Joule thermal constraint loss term begins with the predicted load current. Specifically, the predicted load current for each future time step is squared and multiplied by the resistance per unit length of the cable conductor to obtain the predicted Joule thermal power for that time step. The Joule thermal power is proportional to the square of the current, a core quantitative relationship in Joule's law. Based on the cable's thermal capacity parameters and thermal time constant, the predicted Joule thermal power is converted into the theoretical temperature rise increment caused by the Joule heating effect within that time step. This theoretical temperature rise increment is then accumulated along the time axis from the first predicted time step to obtain the cumulative theoretical temperature rise for each time step. The deviation between the cumulative theoretical temperature rise for each time step and the predicted cable surface temperature for the same time step is calculated. This deviation is calculated by taking the mean square error between the two values, resulting in the Joule thermal constraint loss term. This loss term imposes a constraint along the direction of electrothermal propagation from current to temperature, i.e., the physical causal direction of current generating heat through the Joule heating effect, leading to a temperature rise.

[0080] The calculation of the resistance-temperature coupling constraint loss term begins with the predicted value of the cable surface temperature. Specifically, the predicted cable surface temperature for each future time step is substituted into the resistance temperature coefficient correction formula. Taking copper conductors as an example, based on a reference temperature of 20 degrees Celsius, the resistance changes by approximately 0.393 for every 1 degree Celsius change in temperature. This linear relationship is used to calculate the theoretical resistance value for each time step. Simultaneously, the predicted load current value for each time step and the measured operating voltage value at the current moment are taken. The measured operating voltage value is divided by the predicted load current value to obtain the equivalent resistance value for each time step. The measured operating voltage value is taken as the approximation for future time steps because, within a short-term prediction window of minutes, the grid voltage usually remains stable, with a much smaller variation than the load current. The deviation between the equivalent resistance value and the theoretical resistance value at each time step is calculated using the mean square error method to obtain the resistance-temperature coupling constraint loss term. The loss term imposes a constraint along the thermoelectric propagation direction from temperature to resistance to current, that is, temperature changes cause resistance changes, and resistance changes affect current under the condition that the voltage is basically stable.

[0081] After the two loss terms are calculated, they act together on the same set of predictor variables. The predicted values ​​of the load current and the cable surface temperature are simultaneously constrained by the two loss terms. The constraint propagation paths of the two loss terms are opposite. The Joule heating constraint loss term applies the constraint along the electrothermal propagation direction from current to temperature, requiring that the current prediction and temperature prediction satisfy the heating and temperature rise relationship. The resistance-temperature coupling constraint loss term applies the constraint along the thermoelectric propagation direction from temperature to resistance and then to current, requiring that the temperature prediction and the resistance change and current-voltage relationship satisfy the resistance-temperature coefficient relationship. The two constraint propagation paths are opposite, forming a simultaneous constraint verification of the same set of predictor variables in the two electrothermal propagation directions. The simultaneous constraint verification means that two sets of physical equations simultaneously impose consistency requirements on the same set of variables, so that the values ​​of the variables must simultaneously satisfy the physical laws in both directions. Any constraint in a single direction cannot completely guarantee physical consistency; only when both sets of constraints are satisfied simultaneously can the prediction results have complete physical reliability.

[0082] The technical principle of this solution lies in utilizing the bidirectional nature of cable electrothermal coupling to simultaneously apply physical equation constraints in both the current-to-temperature and temperature-to-current propagation directions, forming a simultaneous constraint verification. The technical advantage is that simultaneous constraints are more stringent than single-directional constraints, more effectively regulating the model's prediction space and ensuring that the predicted load current and cable temperature conform to physical laws in both the electrothermal and thermal propagation directions.

[0083] In a preferred embodiment, in the multi-task prediction head, the first fully connected layer is used to extract a common physical representation shared by each prediction task from the fused features, the common physical representation including electrothermal coupling features and thermal state features; Each of the second fully connected layers extracts the specific features corresponding to the prediction task from the general physical representation and outputs the corresponding thermal key parameters; Four prediction tasks are output in parallel, and there is a hierarchical relationship between the output values ​​of the four prediction tasks in terms of physical laws: According to Joule's law of heating, the predicted value of the load current physically determines the heating power of the cable, which in turn affects the predicted value of the cable surface temperature; the difference between the predicted value of the cable surface temperature and the ambient temperature over time physically determines the predicted value of the heat accumulation index; the predicted value of the heat accumulation index physically reflects the degree of heat accumulation, which in turn determines the predicted value of the fire risk value.

[0084] In practice, the multi-task prediction head performs forward computation during the inference phase as follows.

[0085] The fused feature vector is first input into the first fully connected layer. The first fully connected layer consists of a set of learnable weight matrices and bias vectors, where the dimension of the weight matrices is the fused feature dimension multiplied by a pre-defined intermediate representation dimension. The first fully connected layer performs a linear transformation and nonlinear activation on the fused feature vector, outputting an intermediate representation vector. This intermediate representation vector is the common physical representation shared by all prediction tasks. The common physical representation refers to the abstract feature encoding extracted from the fused features, related to the cable's electrothermal coupling relationship and thermal state evolution, specifically including two types of information: electrothermal coupling features and thermal state features. Electrothermal coupling features encode the interaction between load current and cable temperature, reflecting how current changes affect heating power and how temperature changes affect resistance. Thermal state features encode the current thermal state and thermal evolution trend of the cable, reflecting thermodynamic state information such as internal heat distribution and temperature change direction. The common physical representation is called common because these features are valuable for all four subsequent prediction tasks; any prediction task can benefit from these features.

[0086] The general physical representation vector output by the first fully connected layer is then input into four second fully connected layers. Each of the four second fully connected layers has an independent weight matrix and bias vector, and they do not share parameters. The dimension of the weight matrix is ​​the general physical representation dimension multiplied by the output dimension of the respective predictor variable. Each second fully connected layer performs an independent linear transformation on the input general physical representation vector, extracting the specific features required for its respective prediction task from the general physical representation. The first specific feature extraction branch extracts features directly related to load current changes from the general physical representation, ignoring features directly related to temperature but contributing little to current prediction, and outputs a sequence of predicted load current values. The second specific feature extraction branch extracts features directly related to cable surface temperature changes from the general physical representation, and outputs a sequence of predicted cable surface temperature values. The third specific feature extraction branch extracts features directly related to the heat accumulation process from the general physical representation, and outputs a sequence of predicted heat accumulation index values. The fourth specific feature extraction branch extracts features directly related to risk assessment from the general physical representation, and outputs a sequence of predicted fire risk values. The extraction operations of the four branches are parallel, meaning that the four second fully connected layers simultaneously extract their respective required specific features from the general physical representation, without involving serial data transfer.

[0087] After the four prediction tasks are calculated, the output values ​​exhibit the following hierarchical relationship from a physical perspective. According to Joule's law of heating, the predicted load current physically determines the Joule heating power of the cable conductor. This heating power is the primary heat source for cable temperature changes; therefore, the predicted load current physically influences the predicted cable surface temperature. The difference between the predicted cable surface temperature and the ambient temperature reflects the driving force of heat dissipation from the cable to the surrounding environment. The accumulation of this difference over time is the heat accumulation index; therefore, the difference between the predicted cable surface temperature and the ambient temperature physically determines the predicted heat accumulation index. The predicted heat accumulation index reflects the net accumulation of heat within the cable. The closer the heat accumulation is to the critical value, the higher the fire risk; therefore, the predicted heat accumulation index physically reflects the fire risk. This hierarchical relationship is a causal relationship at the physical level, rather than a computational dependency at the network data flow level. At the network computing level, the four prediction values ​​are output in parallel and obtained synchronously; there is no data transfer between the second fully connected layers where the previous output is used as the input for the next. The physical hierarchy is reflected in the constraint of physical consistency between predicted values ​​through the physical constraint loss function during model training, and in the analysis of the physical interpretability of the four prediction results.

[0088] The technical principle of this solution lies in its structural design, which uses a shared first fully connected layer to extract general physical representations and an independent second fully connected layer to extract task-specific features. This achieves a balance between knowledge sharing and task specialization across multiple related prediction tasks. The shared layer captures the general laws of electrothermal coupling and thermal states, while the independent layers retain the flexibility of each task. The technical benefits are that the parallel output of the four prediction tasks ensures inference efficiency, the hierarchical relationship of the physical hierarchy ensures physical consistency among the prediction results, and the shared general physical representations ensure that each task can benefit from the physical knowledge learned by other tasks.

[0089] The following is an example of using this method to illustrate the entire execution process: Step 1: Acquisition of multi-source sensor data and construction of multivariate time series data for cable operation: (1.1) The following operating parameters, as shown in Table 1, are collected using various types of sensors deployed in the cable and surrounding environment at a fixed sampling frequency (e.g., 1 Hz): Table 1. Monitoring Parameters and Acquisition Methods for Multiple Physical Quantities of Cables

[0090] (1.2) Construction of multivariate time series: After aligning the above 6-dimensional parameters by time, construct a multivariate time series matrix:

[0091] For a historical observation window of length L, an input matrix is ​​formed and used as the input to the multi-source data TCN:

[0092] Step 2: Construction of a forward prediction model based on a physically constrained dual-branch temporal neural network: (2.1) Extraction of temporal features by two branches: Multi-source data TCN branch: Multi-source data TCN processes historical sequences forward in time, responsible for modeling historical patterns and causal dependencies. Its input is historical multivariate time series:

[0093] Future Prior Information TCN Branch: The Future Prior Information Network (TCN) processes future prior information sequences forward in time, responsible for incorporating future trends into power dispatching plans and weather forecasts. Its inputs are:

[0094] in: The future planned load current curve provided by the power dispatching department ( (H=60 corresponds to 60 minutes). : A sequence of future temperature forecasts provided by the meteorological department; : Power grid dispatch plan coding (including time coding of dispatch events such as planned maintenance, load transfer, and distributed generation integration); The structure of the future prior information TCN is symmetrical to that of the multi-source data TCN. It adopts the same dilated causal convolution and residual connection architecture to extract the temporal patterns and trend features in the future prior information along the forward time.

[0095] (2.2) Dual-branch gating feature fusion: The output features of the multi-source data TCN and the future prior information TCN are adaptively merged through a gating fusion mechanism:

[0096]

[0097] in, This represents element-wise multiplication. Here, g is the sigmoid function, and g is the adaptive gating weight vector, with values ​​ranging from 1 to 2. The gating mechanism enables the model to automatically balance the contribution ratio of historical patterns and future prior information based on the characteristics of the input data: when the reliability of future prior information is high (such as stable planned load), the gating tends to increase the weight of the future prior information branch; when the uncertainty of the future is high (such as low accuracy of weather forecasts), the gating automatically increases the weight of the historical branch of multi-source data.

[0098] (2.3) Multi-task joint prediction: Features after fusion The input multi-task prediction head (consisting of two fully connected layers, with each task sharing the first layer and having an independent second layer) synchronously outputs four key prediction variables for the next H time steps (H=60, corresponding to 60 minutes, 1 step per minute): : The predicted value of future load current (A). ; : Predicted future cable surface temperature (°C) ; : Predicted value of future heat accumulation index (°C·min). ; : The predicted value of future fire risk (dimensionless, 0~1). ; The four prediction tasks share and integrate feature representations, and knowledge transfer is achieved through multi-task joint training—load current prediction provides electrothermal coupling priors for temperature prediction, and temperature prediction provides thermal state basis for heat accumulation and risk assessment.

[0099] (2.4) Construction of physical constraint mechanism: This method embeds the thermophysical mechanism of cables into the model training process as a constraint loss function, guiding the neural network to learn a predictive mapping that conforms to physical laws. The following four physical constraint losses are designed: Firstly: Joule thermal confinement loss ; Physical principle: The main heat source in cable operation is Joule heating generated by the load current passing through the conductor resistance, and its heat power is proportional to the square of the current. The predicted future load current and the future cable temperature should satisfy the temperature rise relationship caused by Joule heating.

[0100] Joule heat power calculation formula: The corresponding physical constraint loss function:

[0101] in, The rated power of the cable is used for normalization. This loss term penalizes deviations between predicted current and predicted temperature that violate the Joule thermal coupling relationship, ensuring physical consistency of the electro-thermal prediction.

[0102] Secondly: resistance-temperature coupling constraint loss ; Physical principle: The conductor resistance changes with temperature, satisfying a linear temperature coefficient model. The predicted cable temperature and the predicted load current should satisfy a resistance-temperature coupling relationship.

[0103] Resistance temperature correction formula:

[0104] in, The DC resistance per unit length of the conductor at 20°C ( ), is the temperature coefficient of resistance of the conductor material.

[0105] The corresponding physical constraint loss function:

[0106] in, This is the equivalent resistance value indirectly derived from the load current and voltage. This loss term constrains the physical consistency between temperature prediction and resistance change, ensuring the correct modeling of the thermoelectric coupling relationship.

[0107] Thirdly: heat accumulation constraint loss ; Physical principle: The heat accumulation effect of cables is a core precursor to fire risk: when the cable temperature remains higher than the ambient temperature, heat gradually accumulates in the insulation and conductor, creating potential conditions for thermal runaway. The heat accumulation index is defined as the time integral of the temperature difference.

[0108] Definition of thermal accumulation index:

[0109] The corresponding physical constraint loss function:

[0110] in, This is the theoretical heat accumulation value calculated by numerical integration based on the predicted temperature and the predicted ambient temperature. This is a normalization constant. This loss term constrains the consistency between the predicted value of the heat accumulation exponent and its physical definition, ensuring the correct tracking of heat accumulation trends.

[0111] Fourth: Monotonicity constraint loss ; Physical principle: When other operating conditions remain approximately constant, an increase in load current will increase Joule heating power and tend to drive up cable temperature; this causal relationship is physically irreversible. Therefore, the partial derivative of the predicted future cable temperature with respect to the predicted current should be non-negative.

[0112] Physical conditions for monotonicity:

[0113] The corresponding physical constraint loss function (using penalty for violating monotonicity):

[0114] in , This loss term penalizes the model when the current increases but the predicted temperature decreases, forcing the model to learn a monotonic mapping relationship that conforms to physical causality.

[0115] (2.5) Joint loss function training: The total loss function for model training is a weighted combination of the data prediction loss and the four physical constraint losses:

[0116] Data prediction loss : The data prediction loss uses multi-task weighted mean square error (MSE) to measure the deviation between the model predictions and the actual observations.

[0117]

[0118] in, The weight coefficients for each task are set according to the dimensions of each physical quantity and the training convergence (the default value is 1.0, which can be adaptively adjusted through gradient normalization).

[0119] Step 3: Fire risk assessment and advanced classification and early warning driven by heat accumulation: This method constructs a multi-dimensional fire risk assessment system driven by thermal accumulation, completely eliminating the reliance on post-fire characteristics such as smoke, and realizing advanced risk prediction based on thermal mechanisms.

[0120] (3.1) Multidimensional risk indicator system Based on four key variables predicted by the model for the next 60 minutes, the following risk indicators are constructed: Temperature Exceeding Risk Indicators Predicting that the cable surface temperature will exceed the safety threshold within the next H steps. The percentage of time spent at (e.g., 70℃) and the extent of exceeding the limit. Calculation formula:

[0121] Temperature rise rate risk indicator Predicts the rate of future temperature change, reflecting the urgency of thermal runaway. Calculation formula:

[0122] Heat accumulation risk indicators Based on predicted thermal accumulation index This assesses the degree of heat accumulation inside the cable. This indicator is the core of early warning—a continuously rising heat accumulation index indicates a significantly increased risk of thermal runaway. Calculation formula: ,in This is the critical value for heat accumulation determined based on the cable type.

[0123] Load risk indicators Assess the extent to which future load current exceeds the cable's rated current-carrying capacity. Calculation formula:

[0124] (3.2) Adaptive multi-index fusion based on entropy weight method The entropy weight method is used to achieve adaptive fusion of multi-dimensional risk indicators. The weights of each indicator are dynamically allocated according to the information entropy of each indicator under the current working conditions, thus avoiding the subjectivity of fixed thresholds.

[0125] Fusion steps: (3.2.1): Constructing the evaluation matrix. For the N groups of predicted samples within the current time window, calculate the 4-dimensional risk index for each group of samples and construct the evaluation matrix. Evaluation matrix.

[0126] (3.2.2): Indicator normalization. Min-Max normalization is used to map each indicator to the [0,1] interval.

[0127] (3.2.3): Calculate the information entropy of each indicator. ,in The weight of the j-th indicator in the i-th sample after normalization.

[0128] (3.2.4): Calculate the entropy weight of each indicator. The lower the information entropy of an indicator (the greater its distinguishability), the greater its weight.

[0129] (3.2.5): Weighted fusion. Comprehensive fire risk index. .

[0130] (3.3) Advanced Tiered Early Warning Strategy: The system is based on the comprehensive fire risk index. To achieve a three-level advanced early warning system: Yellow alert ( ): An abnormal temperature rise is predicted to occur within the next 30 to 60 minutes. It is recommended to strengthen monitoring and check the rationality of the load plan.

[0131] Orange alert ( (This indicates a significant thermal risk is predicted within the next 15-30 minutes. It is recommended to adjust the load or increase active cooling measures and arrange for inspections.)

[0132] Red Alert ( (This indicates that there is an extremely high risk of fire within the next 5-15 minutes. It is recommended to immediately reduce the load or cut off the power and activate the emergency response.)

[0133] The early warning output information includes: the current risk level, the predicted time of temperature exceedance (the estimated number of minutes until the temperature exceeds the limit), the hotspot location (based on the spatial distribution of sensors), the risk evolution trend curve, and the recommended response measures.

[0134] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A cable fire early warning method based on a physically constrained dual-branch network, used for quantitative assessment of the thermal state of cables, characterized in that, Includes the following steps: Acquire the historical operating data sequence and future prior information sequence of the cable; the historical operating data includes load current, cable surface temperature and operating voltage; the future prior information includes planned load and weather forecast. The historical running data sequence and the future prior information sequence are respectively input into two independent time-series feature extraction branches of the prediction model to obtain historical time-series features and future prior features; the historical time-series features and future prior features are adaptively fused to obtain fused features; Based on the fusion features, the thermal key parameters for multiple future time steps are predicted simultaneously; the thermal key parameters include the thermal accumulation index, which characterizes the degree to which heat accumulates in the cable over time. The training of the prediction model is guided by a physical constraint loss function, which is based at least on the predicted values ​​of the thermal key parameters and imposes physical consistency constraints on the Joule heating effect and the thermal accumulation effect. Based on the predicted heat accumulation index, generate thermal status indication information for the cable; Based on the fusion features, the thermal key parameters for multiple future time steps are predicted simultaneously, which is achieved through a multi-task prediction head. The multi-task prediction head consists of a first fully connected layer and a second fully connected layer; the first fully connected layer is shared by all prediction tasks, and the second fully connected layer is set independently for each prediction task. The multi-task prediction head takes the fused features as input and synchronously outputs thermal key parameters for a preset number of time steps in the future; wherein, the thermal key parameters include: predicted values ​​of load current, predicted values ​​of cable surface temperature, predicted values ​​of heat accumulation index, and predicted values ​​of fire risk value. The predicted value of the load current is the actual expected value of the load after considering random load disturbances and distributed power source fluctuations, based on the planned load in the future prior information sequence. The physical constraint loss function includes a Joule thermal constraint loss term, a thermal accumulation constraint loss term, a resistance-temperature coupling constraint loss term, and a monotonicity constraint loss term; the monotonicity constraint loss term is constructed in the following manner: Calculate the partial derivative of the predicted future cable surface temperature with respect to the predicted future load current within the time interval during which the predicted future load current is either increasing or remaining in a steady state. When the partial derivative is negative, the absolute value of the partial derivative is used as a penalty term to obtain the monotonicity constraint loss term. The Joule thermal constraint loss term and the resistance temperature coupling constraint loss term share the predicted values ​​of the load current and cable surface temperature output by the multi-task prediction head as inputs. The Joule thermal constraint loss term takes the predicted value of the load current as input, calculates the theoretical temperature rise caused by the Joule thermal effect, and calculates the deviation from the predicted value of the cable surface temperature. The resistance temperature coupling constraint loss term takes the predicted value of the cable surface temperature as input, calculates the theoretical resistance value corrected by the resistance temperature coefficient, and calculates the deviation between the theoretical resistance value and the equivalent resistance value calculated by the predicted value of the load current and the operating voltage. The predicted values ​​of the load current and the cable surface temperature are simultaneously constrained by two loss terms, and the two loss terms are constrained in opposite directions: the Joule thermal constraint loss term constrains the electrothermal propagation direction from current to temperature, and the resistance-temperature coupling constraint loss term constrains the thermoelectric propagation direction from temperature to resistance and then to current, forming a joint constraint verification of the same set of predicted variables in the two propagation directions of electrothermal.

2. The cable fire early warning method based on a physically constrained dual-branch network according to claim 1, characterized in that, The acquisition of the historical operating data sequence of the cable includes the following steps: Multi-dimensional operating parameters are collected at a fixed sampling frequency by deploying various types of sensors on the cable and its surrounding environment. These multi-dimensional operating parameters include load current, operating voltage, active power, cable surface temperature, ambient temperature, and ambient relative humidity. The collected multi-dimensional operating parameters are aligned by timestamps to construct a multivariate time series matrix, which serves as the historical operating data sequence; wherein the time length of the historical operating data sequence is equal to the preset historical observation window length.

3. The cable fire early warning method based on a physically constrained dual-branch network according to claim 1, characterized in that, The sequence of future prior information is constructed by aligning the following three types of information in time: The first type of information is the future planned load current curve provided by the power dispatching department, whose time span covers the future preset forecast period; The second type of information is the future temperature forecast sequence provided by the meteorological department, whose time resolution is consistent with the time resolution of the future planned load current curve. The third type of information is the grid dispatch plan code, which includes at least one of the following: time code for planned maintenance events, time code for load transfer events, and time code for distributed generation integration events.

4. The cable fire early warning method based on a physically constrained dual-branch network according to claim 1, characterized in that, Both of the independent temporal feature extraction branches employ temporal convolutional networks, and the two temporal convolutional networks have the same number of dilated causal convolutional layers and residual connection structures; Specifically, the first temporal feature extraction branch performs dilated causal convolution on the historical running data sequence along the forward time direction to extract the historical temporal features; The second temporal feature extraction branch performs an expanded causal convolution on the future prior information sequence along the forward time direction to extract the future prior features.

5. The cable fire early warning method based on a physically constrained dual-branch network according to claim 1, characterized in that, The Joule thermal constraint loss term is constructed in the following manner: Based on the physical relationship that Joule thermal power is proportional to the square of the load current, the predicted future load current is substituted into the Joule thermal power calculation formula to estimate the theoretical temperature rise caused by the Joule thermal effect. The deviation between the theoretical temperature rise and the predicted future cable surface temperature is calculated to obtain the Joule thermal constraint loss term. The heat accumulation constraint loss term is constructed in the following way: The theoretical heat accumulation value under the simplified thermal model is obtained by numerically integrating the difference between the predicted future cable surface temperature and the predicted future ambient temperature along the time dimension. The deviation between the theoretical heat accumulation value and the heat accumulation index is calculated to obtain the heat accumulation constraint loss term.

6. The cable fire early warning method based on a physically constrained dual-branch network according to claim 5, characterized in that, The resistance-temperature coupling constraint loss term is constructed in the following manner: Based on the physical law that conductor resistance changes linearly with temperature, the theoretical resistance value is calculated using the temperature coefficient of resistance correction formula according to the predicted future cable surface temperature. The equivalent resistance value calculated from the predicted future load current and operating voltage is compared with the theoretical resistance value to obtain the resistance temperature coupling constraint loss term.

7. The cable fire early warning method based on a physically constrained dual-branch network according to claim 1, characterized in that, The step of generating thermal status indication information for the cable based on the predicted heat accumulation index includes the following steps: Based on predicted key thermal parameters for multiple future time steps, a multi-dimensional risk assessment indicator system is constructed; the multi-dimensional risk assessment indicator system includes: The risk index is calculated based on the predicted future cable surface temperature exceeding the preset safety threshold by a certain amount and duration; the risk index is calculated based on the predicted future cable surface temperature changing rate within a preset time period; the risk index is calculated based on the ratio of the heat accumulation index to the cable heat accumulation critical value; and the risk index is calculated based on the predicted future load current exceeding the cable's rated current carrying capacity. The entropy weight method is adopted to dynamically determine the weight coefficient of each risk indicator based on the information entropy of each risk indicator under the current working conditions. The risk indicators are weighted and integrated according to their corresponding weight coefficients to generate a comprehensive fire risk index. Based on the threshold range of the comprehensive fire risk index, a graded early warning signal is output.

8. The cable fire early warning method based on a physically constrained dual-branch network according to claim 1, characterized in that, In the multi-task prediction head, the first fully connected layer is used to extract a common physical representation shared by each prediction task from the fused features. The common physical representation includes electrothermal coupling features and thermal state features. Each of the second fully connected layers extracts the specific features corresponding to the prediction task from the general physical representation and outputs the corresponding thermal key parameters; Four prediction tasks are output in parallel, and there is a hierarchical relationship between the output values ​​of the four prediction tasks in terms of physical laws: According to Joule's law of heating, the predicted value of the load current physically determines the heating power of the cable, which in turn affects the predicted value of the cable surface temperature; the difference between the predicted value of the cable surface temperature and the ambient temperature over time physically determines the predicted value of the heat accumulation index; the predicted value of the heat accumulation index physically reflects the degree of heat accumulation, which in turn determines the predicted value of the fire risk value.

Images