A high-reliability power equipment operation temperature field prediction method

Abstract

The application provides a kind of power equipment operation temperature field high reliability prediction method, belongs to distribution network equipment intelligentization technical field, is applicable to environmental protection gas insulation ring network cabinet, circuit breaker, primary and secondary fusion complete set pole-mounted circuit breaker, primary and secondary fusion complete set ring network box, reactive power compensation device class equipment, voltage regulator, high-voltage cabinet and other equipment, for solving the problem that the internal temperature of sealed gas tank cannot be penetrated and the model lacks physical constraints in related technologies, so as to protect and improve the operation quality of power equipment. By acquiring multi-source sensor data and constructing ring network box physical topology graph adjacency matrix; using physical guided neural differential operator network, combining dynamic adaptive space-time fractional physical constraint and implicit Bayesian variational inference for real-time deduction and online self-correction, output three-dimensional temperature field distribution data. The application can realize high-precision real-time prediction and active thermal management of internal temperature field, significantly improve the safety and reliability of distribution network and maintenance-free ability.

Description

Technical Field

[0001] This application relates to the field of intelligent power distribution network equipment, and is applicable to equipment such as environmentally friendly gas-insulated ring main units, circuit breakers, integrated pole-mounted circuit breakers, integrated ring main units, reactive power compensation devices, voltage regulators, and high-voltage cabinets. In particular, it relates to a highly reliable method for predicting the operating temperature field of power equipment. Background Technology

[0002] Environmentally friendly gas-insulated ring main units, circuit breakers, integrated primary and secondary pole-mounted circuit breakers, integrated primary and secondary ring main units, reactive power compensation devices, voltage regulators, and high-voltage switchgear are core equipment in distribution networks, responsible for switching, protection, compensation, and voltage regulation. They are widely used in distribution network automation and power quality management. These devices often contain complex combinations of insulating media (such as environmentally friendly gases and epoxy resin composite insulation), and experience significant load fluctuations during actual operation. Temperature is a key physical indicator affecting their insulation performance and safe lifespan. Therefore, rapid and accurate sensing and prediction of the internal heating patterns and temperature field distribution of these devices are core technical requirements for ensuring the safe and stable operation of distribution networks.

[0003] Currently, temperature monitoring and prediction for the aforementioned equipment mainly rely on external infrared thermal imagers for surface temperature measurement, or on traditional finite element simulation methods for offline simulation of the internal temperature field. Some cutting-edge solutions incorporate machine learning models, using current, voltage, and ambient temperature as inputs to attempt to predict internal hotspot temperatures.

[0004] However, existing technologies have significant shortcomings. Thermal imagers can only acquire the surface temperature of equipment and cannot penetrate the high-temperature areas inside sealed gas boxes, and are easily affected by external weather conditions; traditional finite element simulation is slow and cannot meet the needs of real-time online prediction at the edge of the power distribution network; while simple machine learning models lack physical mechanism constraints, and their prediction accuracy drops sharply when extreme operating conditions or equipment aging causes changes in thermal properties, and they operate as a "black box," making it difficult to quantify and assess their own prediction uncertainties, let alone support highly reliable real-time self-correction and active control at the edge of the power grid. Summary of the Invention

[0005] This application provides a highly reliable prediction method for the operating temperature field of power equipment. It can deeply integrate physical topology with primary and secondary fusion sensing data. By introducing dynamic adaptive spatiotemporal fractional-order physical constraints and implicit Bayesian variational online inference, it can achieve high-precision, self-correcting real-time prediction and active safety control of the internal temperature field of power equipment, so as to ensure and improve the operating quality of power equipment.

[0006] Firstly, this application provides a highly reliable method for predicting the operating temperature field of power equipment. It acquires multi-source sensor data collected in real time by a primary and secondary fusion sensing layer, including three-phase load current and three-phase voltage outputs from capacitive voltage divider sensors and Rogowski coil current sensors, and external ambient temperature and internal insulating gas pressure outputs from temperature, humidity, and pressure sensors deployed in the gas chamber. Based on the multi-source sensor data and a pre-defined structure database, it maps the internal heat conduction entities of the power equipment as physical graph nodes and constructs an adjacency matrix of the physical topology graph according to the heat conduction path. The multi-source sensor data and the adjacency matrix are input into a physical-guided neural differential operator network. This network continuously evolves the thermodynamic state through dynamically adaptively generated spatiotemporal fractional-order physical constraints, and performs discrete integration on the temperature evolution modeled as an ordinary differential equation using a pre-defined numerical solver. During the evolution process, the network parameters are self-corrected online based on implicit Bayesian variational inference. The output is the three-dimensional temperature field distribution data inside the power equipment obtained from the evolution.

[0007] By adopting the above technical solution, a neural differential operator network with physical topological prior constraints and dynamic fractional physical laws was constructed, realizing end-to-end high-precision real-time prediction of the three-dimensional temperature field inside power equipment. At the same time, by introducing the physical mechanism into the neural network loss function and using Bayesian variational inference for parameter self-correction, the defects of traditional black box models in prediction inaccuracy under extreme working conditions and performance drift after long-term operation are effectively avoided.

[0008] Furthermore, after acquiring multi-source sensor data, the process also includes performing weighted median filtering on the multi-source sensor data to filter out electromagnetic transient pulse spikes caused by switching operations. The weighted median filtering process uses a preset sliding window to perform robust statistics on the high-frequency sampling signal, and uses a multi-sensor spatiotemporal alignment algorithm with a built-in Kalman filter to map the millisecond-level response of electrical quantities and the second-level hysteresis response of heat to the same system time base for phase compensation and timestamp alignment.

[0009] By adopting the above technical solution, electromagnetic pulse interference in the power system is effectively filtered out, and sampling delay and phase deviation caused by the different physical locations and physical response mechanisms of electrical and thermal sensors are eliminated, providing clean and synchronous input data for subsequent physical graph topology construction and neural network deduction.

[0010] Furthermore, the internal insulating cylinder, arc-extinguishing chamber, instrument transformer, stainless steel gas box inner wall and internal insulating gas area of ​​the power equipment are defined as nodes of the physical graph, and the adjacency matrix of the physical topology graph is constructed based on whether there is a direct physical contact surface or convective heat transfer path between each physical graph node.

[0011] By adopting the above technical solution, the complex physical structure relationship inside the power equipment is accurately mapped into a graph data structure, enabling the subsequent neural network to perform heat flow diffusion inference based on the real heat conduction path, which significantly improves the temperature field prediction model's inherent understanding of the equipment's physical topology.

[0012] Furthermore, the multi-source sensor data is input into the state-aware network built into the physical-guided neural differential operator network. The state-aware network outputs the time fractional order and the spatial fractional order in real time. The time fractional order and the spatial fractional order are then substituted into the thermodynamic partial differential equation based on the time fractional order derivative and the spatial fractional order derivative to form the dynamically adaptively generated spatiotemporal fractional order physical constraint.

[0013] By adopting the above technical solution, the constraint of fixed fractional order in traditional physical models is broken, and the order of physical constraints is dynamically adjusted according to the actual thermodynamic state of the equipment (such as the rate of temperature change). This enables the accurate characterization of the abnormal thermal diffusion effect of composite insulating media during thermal expansion and contraction or aging, giving the model a more accurate deduction capability that fits the physical essence.

[0014] Furthermore, the physical-guided neural differential operator network utilizes a graph attention mechanism to aggregate the temperature features of neighboring nodes for each physical graph node in the physical topology graph and calculates the attention weights. Then, a graph transformer network is used to perform nonlinear transformation on the aggregated temperature features and attention weights to simulate the thermal diffusion process of physical heat on the physical topology graph.

[0015] By adopting the above technical solution, using the graph attention mechanism to perform weighted aggregation of physical nodes, and then performing nonlinear transformation through the graph transformer network, the algorithm can more accurately simulate the actual conduction and exchange process of physical heat in complex topological structures with multiple media and multiple paths.

[0016] Furthermore, the change of the three-dimensional temperature field inside the power equipment over time is modeled as an ordinary differential equation, and the ordinary differential equation is discretely integrated using a preset numerical solver to output the temperature state of each physical graph node at the next moment. The preset numerical solver is a high-order Runge-Kutta method solver based on time discrete step size.

[0017] By adopting the above technical solution, the continuous thermophysical evolution process is transformed into a neural network-driven ordinary differential equation, and the high-order Runge-Kutta method is used for discrete integration, which ensures the continuity, stability and high accuracy of the temperature field in the time dimension.

[0018] Furthermore, the online self-calibration of network parameters based on implicit Bayesian variational inference specifically includes constructing an overall loss function for the physical-guided neural differential operator network. The overall loss function includes a data fitting residual term, a dynamic fractional physical residual term, and a KL divergence regularization term representing the difference between the posterior and prior probability distributions of the network parameters. The overall loss function is then used to perform online posterior updates of the network parameters of the physical-guided neural differential operator network.

[0019] By adopting the above technical solution, multiphysics constraints and statistical learning theory are organically integrated. While the network is performing a round of forward inference, the posterior probability update of the network parameters is automatically completed, giving the algorithm a strong "inference and self-healing" capability. It can maintain high accuracy throughout its entire life cycle without relying on an external independent feedback module.

[0020] Furthermore, after outputting the three-dimensional temperature field distribution data inside the power equipment, the method also includes locking the highest temperature hot zone node based on the three-dimensional temperature field distribution data and calculating the thermodynamic safety margin. The thermodynamic safety margin represents the percentage difference between the current highest temperature hot zone node and the preset insulation medium temperature resistance threshold. It is combined with the preset insulation medium thermal aging activation energy to calculate and accumulate dynamic thermal aging loss. The accumulated dynamic thermal aging loss is obtained by time-domain integration of the instantaneous thermal aging rate within a predetermined time interval.

[0021] By adopting the above technical solutions, the temperature field data derived from AI can be directly converted into physical indicators such as thermodynamic safety margin and cumulative thermal aging loss that can be evaluated, thereby realizing a quantitative assessment of the operating status and insulation life of power equipment and providing an intuitive and clear basis for equipment operation and maintenance.

[0022] Furthermore, it also includes generating hierarchical decision instructions based on the thermodynamic safety margin and temperature change trend. The hierarchical decision instructions include cloud-based macro-level operation and maintenance early warning information and edge active control instructions. When the thermodynamic safety margin is lower than a first threshold, cloud-based macro-level operation and maintenance early warning information is generated and uploaded to the power distribution master station system through the communication module. When the thermodynamic safety margin is lower than a second threshold, active control instructions are generated and sent to the power equipment operating mechanism to perform load reduction or tripping actions.

[0023] By adopting the above technical solutions, the temperature field prediction results are linked with the macro-operation and maintenance management of G06Q level and the edge active protection of G05B level, and a complete closed-loop system of "sensing-prediction-decision-control" is constructed. The system outputs graded operation instructions for different levels of overheating risk, which greatly improves the comprehensive safety protection capability of the power grid.

[0024] Furthermore, the multi-source sensor data is collected and output in real time by the capacitive voltage divider sensor and Rogowski coil current sensor integrated inside the pole of the pole-mounted circuit breaker. The real-time forward inference and online self-correction calculation of the physical-guided neural differential operator network are deployed in the distribution network terminal unit integrated with a neural network acceleration unit. The dynamically adaptively generated spatiotemporal fractional physical constraints and implicit Bayesian variational inference are both executed in the distribution network terminal unit.

[0025] By adopting the above technical solution, the sensor is integrated inside the circuit breaker pole, and real-time inference and online self-healing calculation are completed on the edge terminal with built-in NPU. The hardware advantages of the primary and secondary fusion architecture are fully utilized, and high-precision thermal field perception can be completed independently at the device end, avoiding the delay and signal instability caused by frequent cloud communication.

[0026] In summary, this application has at least the following beneficial effects:

[0027] A highly reliable method for predicting the operating temperature field of power equipment is provided. By deeply integrating physical topology, fractional physical constraints and Bayesian online inference, a high-fidelity, self-correcting real-time prediction of the three-dimensional temperature field inside power equipment is achieved, which significantly improves the safety and reliability of power distribution network equipment operation.

[0028] By designing dynamic adaptive fractional-order physical constraints and graph attention mechanisms, we can accurately characterize the non-integer-order thermal diffusion phenomenon of composite insulating media on complex topological structures, enabling the prediction model to have strong physical interpretability and multi-condition generalization ability.

[0029] By constructing an overall loss function that includes KL divergence and physical residuals and implementing online posterior updates, edge terminal algorithms are endowed with intelligent self-healing capabilities that require no manual maintenance throughout the entire lifecycle of the device, significantly reducing operation and maintenance costs.

[0030] It should be understood that the description in the Summary Section is not intended to limit the key or essential features of the embodiments of this application, nor is it intended to restrict the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0031] The above and other features, advantages, and aspects of the embodiments of this application will become more apparent from the accompanying drawings and the following detailed description. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0032] Figure 1 A schematic diagram of an exemplary operating environment in which embodiments of this application can be implemented is shown.

[0033] Figure 2A flowchart of a high-reliability prediction method for the operating temperature field of power equipment is shown in an embodiment of this application. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0035] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0036] This application provides a highly reliable method for predicting the operating temperature field of power equipment. It achieves high-precision real-time prediction and online self-correction of the internal three-dimensional temperature field by fusing physical topology maps and multi-source sensing data, utilizing a physical-guided neural differential operator network, and combining dynamic adaptive spatiotemporal fractional-order physical constraints with implicit Bayesian variational online inference. This effectively improves the safety, reliability, and maintenance-free capability of power distribution network equipment, thereby ensuring and improving the operating quality of power equipment.

[0037] Figure 1 A schematic diagram of an exemplary operating environment in which embodiments of this application can be implemented is shown.

[0038] Reference Figure 1 The operating environment includes the physical body of the primary and secondary integrated pole-mounted ring network box (the power equipment here is described using the primary and secondary integrated pole-mounted ring network box as an example; the same applies to environmentally friendly gas-insulated ring network cabinets, circuit breakers, primary and secondary integrated pole-mounted circuit breakers, reactive power compensation devices, voltage regulators, high-voltage cabinets, and other similar equipment), as well as a collection of four core hardware subsystems necessary for realizing the method of this invention: sensing and acquisition, edge computing, action execution, and communication interaction.

[0039] As the physical foundation of this method, the physical body of the primary and secondary integrated column-mounted ring main body mainly includes a stainless steel gas box, a vacuum interrupter chamber set inside the gas box, a solid-sealed pole, an insulating cylinder, and an electric spring operating mechanism. Among them, the stainless steel gas box is made of a material with a thickness of at least 2.5 mm, has a fully sealed characteristic, and the protection level reaches IP67. The solid-sealed pole and the insulating cylinder are both integrally cast with insulating material and conductive circuit to support the conductive and insulating structure of the circuit breaker.

[0040] The sensing and acquisition subsystem mainly consists of a capacitive voltage divider sensor and a Rogowski coil current sensor embedded inside the solid-sealed pole of the pole-mounted circuit breaker, as well as a temperature and humidity sensor and a pressure sensor deployed inside and outside the stainless steel gas chamber of the ring main unit, respectively, for real-time acquisition of multi-source sensor data. Specifically, the voltage measurement range of the capacitive voltage divider sensor can be set to 3 kV to 12 kV, preferably 10 kV; the rated current measurement range of the Rogowski coil current sensor is set to 200 amps to 1200 amps, preferably 600 amps; both types of sensors are directly cured with epoxy resin into the solid-sealed pole, and their signal response time is typically between 1 ms and 5 ms, preferably 3 ms, to ensure rapid capture of high-frequency transient surges in the power system. The temperature and humidity sensors deployed outside the ring main unit employ outdoor probes with UV-resistant coatings and anti-condensation designs. Their temperature measurement range covers -40°C to 85°C, and their humidity measurement range is 5% to 95% relative humidity. The pressure sensors deployed inside the gas chamber are absolute pressure sensors, typically with a measurement range of 50 kPa to 150 kPa, preferably 100 kPa, and their temperature resistance must meet the highest operating temperature requirements inside the gas chamber. All sensors in this subsystem connect to the intelligent terminal block at the bottom of the ring main unit via shielded signal cables that extend through a sealed flange and exit the gas chamber, enabling centralized signal aggregation and output.

[0041] The core of the edge computing and action execution subsystem is an intelligent distribution network terminal unit that integrates a neural network acceleration unit. This terminal unit receives multi-source data from the sensing and acquisition subsystem through a high-speed data bus and completes the real-time forward inference of the physical-guided neural differential operator network, online Bayesian variational inference self-correction, and dynamic adaptive fractional physical constraint calculation process on its built-in computing platform. The intelligent distribution network terminal unit adopts an industrial-grade embedded architecture. Its main control chip is an ARM-based microprocessor, coupled with a dedicated neural network acceleration unit, providing an integer operation capability ranging from one trillion to eight trillion operations per second, preferably four trillion operations per second to meet the high-frequency real-time computing requirements of complex graph neural networks and neural differential equations. Its built-in random access memory has a storage capacity of two gigabytes to eight gigabytes, preferably four gigabytes. The algorithm inference cycle within the edge computing terminal unit can be dynamically set according to actual line load changes. For example, under normal steady-state load conditions, it can be set to execute a complete inference and self-correction cycle once every 100 milliseconds. When the load fluctuates drastically or switching operations occur, an interrupt is triggered to actively shorten the inference cycle to 50 milliseconds or even 10 milliseconds to ensure real-time tracking of transient thermal effects. Simultaneously, the terminal unit is connected to the ring main unit's own electric spring operating mechanism via a control bus to directly control the operating mechanism to perform corresponding load reduction or switching actions based on the calculated hierarchical decision commands. Specifically, the control bus adopts the CAN industrial fieldbus with high anti-interference capability, and its baud rate can be set from 250 kilobits per second to 500 kilobits per second; the electric spring operating mechanism is equipped with a set of closing springs and opening springs driven by the energy storage motor, and is equipped with an auxiliary contact feedback mechanism. The intelligent distribution network terminal unit sends start or stop commands to the energy storage motor controller of the operating mechanism through the digital output port, and sends action commands to the opening and closing coils through relay contact signals, thereby realizing hardware-level closed-loop electrical control.

[0042] The communication and interaction subsystem relies on 5G or fiber optic communication modules configured on terminal units to achieve data interoperability between the edge computing side and the distribution master station side. The distribution master station system receives macro-level operation and maintenance early warning information uploaded by the terminal units based on algorithm judgment results, which is used to realize equipment life-cycle status assessment and overall scheduling. Among them, the fiber optic communication module supports the single-mode dual-fiber transmission standard, and its transmission distance can cover the communication between substations and distribution network nodes from one kilometer to forty kilometers, with a preferred transmission specification of ten kilometers to meet the networking requirements of urban distribution ring networks. The 5G communication module is compatible with the frequency bands of different operators and supports network slicing to ensure the priority of power control signals. The distribution master station system includes a front-end communication server, a data acquisition and monitoring control server, and a human-machine interaction terminal, which is interconnected with the above communication modules through firewalls and power-specific network security isolation devices. The edge-side terminal units encapsulate the calculated alarm or early warning data, including thermodynamic safety margin, cumulative thermal aging loss, and current hot spot temperature, into data frames conforming to power standard communication protocols, and send them to the distribution master station periodically or triggered by this interaction subsystem, forming a complete macro-monitoring data stream.

[0043] In summary, the above subsystems are physically connected in series via signal lines, buses, and communication links to form a closed loop, which together constitute the macroscopic operating environment that supports the method of the present invention to operate independently and in real time at the edge and has online self-correction capabilities.

[0044] Figure 2 A flowchart illustrating a high-reliability prediction method for the operating temperature field of power equipment, according to an embodiment of this application, is shown. (Refer to...) Figure 2 The method specifically includes the following steps:

[0045] S1: Acquire multi-source sensor data collected in real time by the primary and secondary fusion sensing layer, including the three-phase load current and three-phase voltage output by the capacitive voltage divider sensor and the Rogowski coil current sensor, and the external ambient temperature and internal insulating gas pressure output by the temperature, humidity and pressure sensors deployed in the gas box.

[0046] The method in this step specifically includes: after acquiring multi-source sensor data, performing weighted median filtering on the aforementioned data to filter out electromagnetic transient pulse spikes caused by switching operations. The weighted median filtering process uses a preset sliding window to perform robust statistics on the high-frequency sampled signals, and utilizes a multi-sensor spatiotemporal alignment algorithm with a built-in Kalman filter to map the millisecond-level response of electrical quantities and the second-level hysteresis response of heat to the same system time base for phase compensation and timestamp alignment. During the weighted median filtering process, the data point sequence within the sliding window is assigned different weight values ​​based on its distance from the center of the window. These weight values ​​decrease symmetrically with increasing distance and are referred to as weights. ; Utilize this weight The temperature output value that minimizes the weighted absolute deviation is the filtered clean measurement value. Weighted median filtering can preferentially suppress the amplitude of high-frequency pulse spike signals while protecting the gradual changes in the rising and falling edges from being smoothed out, thus achieving the dual purpose of filtering out electromagnetic pulse interference and retaining effective physical signals. The multi-source sensor data is acquired and output in real time by capacitive voltage divider sensors and Rogowski coil current sensors integrated inside the pole-mounted circuit breaker poles. The real-time forward inference and online self-correction calculation of the physical-guided neural differential operator network are deployed in the distribution network terminal unit integrated with a neural network acceleration unit. Furthermore, the dynamically adaptively generated spatiotemporal fractional-order physical constraints and implicit Bayesian variational inference are both executed in the distribution network terminal unit. In the multi-sensor spatiotemporal alignment module, the coefficient matrix used to describe the physical dynamic characteristics of the system is called the system state transition matrix. The system contains a multi-dimensional set of parameters, including current, voltage, and temperature, which constitute the state vector. Random disturbances present during the actual measurement process are due to process noise. Characterization; at discrete sampling times, the matrix describing the mapping relationship between the state vector and the sensor observations is called the observation matrix. The measured error is caused by measurement noise. Characterization; to perform discretization calculations, the above state transition relationships are transformed into a discrete-form state transition matrix. And through the Kalman gain matrix The predicted state values ​​are successively weighted and corrected to obtain the phase alignment results of electrical quantities and thermal quantities that are accurately synchronized at each discrete moment. After the above filtering and alignment processing, the original multi-source heterogeneous signals are transformed into compact, denoised, and phase-consistent standardized input feature vectors, providing high-quality physical data input for physical graph topology construction and deep neural network deduction in subsequent steps.

[0047] S2: Based on the aforementioned data and the preset structure database, the heat conduction entities inside the ring network box are mapped as physical graph nodes, and an adjacency matrix of the physical topology graph is constructed according to the heat conduction path.

[0048] The specific steps of this method include: defining the internal insulating cylinder, arc-extinguishing chamber, instrument transformer, stainless steel gas box inner wall, and internal insulating gas area of ​​the ring main unit as nodes in the physical graph, and constructing an adjacency matrix of the physical topology graph based on whether there are direct physical contact surfaces or convective heat transfer paths between the physical graph nodes. The total number of defined physical graph nodes is... Its value is determined by the accuracy of the subdivision of the ring network box spatial structure based on the preset structural database, and usually depends on the number of entities and the degree of refinement of independent heat transfer zones. The set of physical graph nodes is denoted as the graph node set. This set serves as the underlying carrier for subsequent mathematical modeling and graph convolution operations, directly representing the spatial state of different hardware entities within the ring network enclosure. After completing node definition and set construction, the connection relationships between nodes need to be established through heat conduction paths. If it is determined that there is direct physical contact between two nodes, a physical heat conduction edge is assigned to them, and the weight of this edge is called the physical heat conduction edge weight. The value of this weight is determined by the equivalent thermal conductivity of the contact material. Direct physical contact area and the equivalent path length of heat along the direction of heat conduction jointly determined, that is

[0049]

[0050] Among them, equivalent thermal conductivity The parameters are calculated jointly based on the inherent thermal properties of the two contact materials in a pre-defined structural database. Their function is to determine the efficiency of heat transfer on the contact surface; the direct physical contact area... Assembly drawing parameters derived from a pre-defined structural database are used to quantify the cross-sectional area of ​​the heat exchange channel between nodes; equivalent path length. This is the converted length simulating the actual heat conduction distance, used to consider the contribution of distance to thermal resistance. If there is convective heat transfer between the insulating gas and the wall between two nodes, a convective heat transfer edge is assigned to both, and its weight is called the convective heat transfer edge weight. Its value is determined by the transient convective heat transfer coefficient. With effective heat exchange area Joint decision, that is

[0051]

[0052] Among them, the transient convective heat transfer coefficient Based on the current pressure of the insulating gas inside the gas box The heat exchange capacity of the gas inside the gas box is dynamically calculated based on the external ambient temperature and used to reflect changes in the heat exchange capacity of the gas under different external operating conditions in real time. After traversing all node pairs, the algorithm will assign all edge weights... The organization is a symmetric adjacency matrix Each non-zero element in this matrix explicitly indicates a quantifiable heat flow path between two corresponding nodes, providing a physically necessary connectivity for subsequent graph neural networks. Subsequently, based on the adjacency matrix... Generating degree matrix This degree matrix is ​​a diagonal matrix, and the values ​​on its diagonal are defined as follows:

[0053]

[0054] Indicates the relationship with the first The sum of the weights of all edges connecting each node. Finally, construct the graph Laplacian matrix based on these two matrices. Its definition

[0055]

[0056] Its function is to define a heat diffusion operator on the complete physical graph structure, ensuring that heat diffusion in subsequent deep networks, when solving thermodynamic evolution, strictly follows the internal conduction paths provided by the physical topology. This Laplace matrix... Together with the adjacency matrix The standardized input feature vector output by S1 will be fed into the next step.

[0057] S3: Input the aforementioned data and adjacency matrix into the physical-guided neural differential operator network. This network continuously evolves the thermodynamic state through dynamically adaptively generated spatiotemporal fractional physical constraints, and performs discrete integration on the temperature evolution modeled as an ordinary differential equation through a preset numerical solver. Furthermore, the network parameters are self-corrected online based on implicit Bayesian variational inference during the derivation process.

[0058] The specific steps of this method include: inputting the multi-source sensor data into the state-aware network built into the physical-guided neural differential operator network; using the state-aware network to output the time fractional order and spatial fractional order in real time; and substituting the time fractional order and spatial fractional order into the thermodynamic partial differential equations based on the time fractional derivative and spatial fractional derivative to form the dynamically adaptively generated spatiotemporal fractional order physical constraints. In this step, the standardized multi-source feature vector output by S1 is used as the initial boundary condition, and the physical graph topology information adjacency matrix constructed by S2 is... With the Laplace matrix As the input to the pre-built physical structure of the network, these are fed into the physical-guided neural differential operator network. This built-in state-aware network employs a shallow feedforward multilayer perceptron structure, designed to capture the dynamic thermodynamic boundaries of the current device. Its input layer receives key state variables processed by feature engineering, specifically including the second derivative of the current rate of change, the current pressure of the insulating gas inside the gas chamber, and the ambient temperature gradient. These core variables are first mapped to a high-dimensional hidden layer, and then processed through a weight matrix. and With state vector The linear computation and subsequent Sigmoid activation generate dynamic time fractional orders in real time. and spatial fractional order ,Right now

[0059]

[0060] as well as

[0061]

[0062] Among them, the weight matrix and These are the pre-trained parameters of the state-aware network, used to control the output layer's prediction response weights for temporal memory and spatial nonlocal diffusion characteristics; the state vector... This is the set of results from the aforementioned feature engineering. The fractional order of this dynamically outputted model replaces the fixed constant order in the traditional model and is directly applied to the thermodynamic partial differential equations to construct the following dynamically adaptive spatiotemporal fractional physical constraints:

[0063]

[0064] In this constraint equation, the left-hand side term For the time Caputo fractional derivative operator, given the fractional order of the current time step... This operator, within its integral kernel, embodies the "memory characteristic" of heat flow within the insulating material over time, clearly describing the hysteresis and memory extension effects of heat caused by the non-uniform dielectric interface; the first term on the right... It is a fractional-order Laplace operator in Riesz space. It breaks through the limitation of the traditional Fourier theorem, which only considers the heat conduction of local adjacent points, and introduces the long-range spatial jump diffusion characteristics, which can accurately simulate the heat conduction of complex multi-medium interfaces in ring network boxes. , , These represent the effective thermal conductivity, density, and specific heat capacity at constant pressure of the ring main unit's insulation material, respectively. These values ​​are derived from the initial configuration of the preset structure database. The second item on the right... In physical space Location at The time-varying Joule heat source is calculated from the three-phase load current and preset conductor resistance parameters collected in step S1. This method of dynamically adjusting the order based on real-time operating conditions completely removes the physical constraints from the scope of static formulas, enabling the model to maintain a high degree of physical self-consistency and adaptability even when facing insulation material aging or extreme environmental changes.

[0065] The physical-guided neural differential operator network utilizes a graph attention mechanism to aggregate the temperature features of neighboring nodes for each physical graph node in the physical topology graph and calculates attention weights. A graph transformer network then performs a nonlinear transformation on the aggregated temperature features and attention weights to simulate the thermal diffusion process of physical heat on the physical topology graph. In the specific implementation of the graph attention mechanism, each node... Each neighboring node within the physical neighborhood Based on the temperature field feature mapping results between them, the unnormalized attention score will be calculated. The specific calculation relationship is as follows:

[0066]

[0067] in, For the network The shared learnable linear transformation matrix in the layer is used to project the node temperature features to a high-dimensional feature space. Its function is to map the original temperature state of the physical nodes to the same feature scale so that interactive computation can be performed. This is a single-layer feedforward neural network responsible for calculating the attention response score between two high-dimensional features. After calculating the unnormalized attention score, the Softmax function is used to transform it into normalized weights.

[0068]

[0069] Normalized attention weights This determines the degree to which the temperature information of the neighboring node contributes during subsequent feature aggregation. Ultimately, through the combination of several graph attention layers, feedforward neural networks, and residual connections in the graph transformer network, an accurate physical simulation of the dynamic diffusion process of heat in the graph structure is achieved.

[0070] The time-varying three-dimensional temperature field inside the ring network enclosure is modeled as an ordinary differential equation. A pre-defined numerical solver is used to discretize and integrate this equation to output the temperature state of each physical graph node at the next time step. This pre-defined numerical solver is a high-order Runge-Kutta method solver based on time discrete step sizes. In this step, the spatial feature transformation output by the graph transformer network is considered as a continuous ordinary differential equation representing the time evolution of the temperature state across the entire physical graph node set. Its core expression is:

[0071]

[0072] in, Representatives cover all A multidimensional time-varying vector of temperature status of each physical graph node, which contains the temperature values ​​of all key heat conduction entities inside the ring network box; The output of S1 is the standardized multi-source input feature vector; This is the graph Laplacian matrix output from S2. The core function mapping of the physical-guided neural differential operator network continuously maps the input conditions, physical topology, and prior temperature state to the rate of change of the temperature field over time. To achieve high-fidelity evaluation of this differential equation, this embodiment employs a high-order fourth-order Runge-Kutta discretization solver based on time discrete step sizes for time integration. The iterative calculation formula is as follows:

[0073]

[0074] in, The preset network inference integral time step; For iterative calculations, the system function is used. Four intermediate slope terms, obtained at four different computational sampling points, reflect the rate of change of the temperature field at different locations within the prediction time interval. By multi-step coupling and weighting of these slope terms, the solver can stably and accurately deduce the temperature state of each physical graph node at the next moment, achieving real-time visualization of the thermal field of key components inside the sealed gas chamber.

[0075] The online self-calibration of network parameters based on implicit Bayesian variational inference specifically includes constructing an overall loss function for the physical-guided neural differential operator network. This overall loss function includes a data fitting residual term, a dynamic fractional physical residual term, and a KL divergence regularization term representing the difference between the posterior and prior probability distributions of the network parameters. The overall loss function is then used to perform online posterior updates of the network parameters of the physical-guided neural differential operator network. In this step, the physical-guided neural differential operator network employs an implicit Bayesian variational inference framework to ensure its self-healing capability under long-term deployment. This framework tightly couples the online posterior update process of the model parameters with a forward inference. Its core lies in constructing a multi-objective coupled overall loss function.

[0076]

[0077] In each term of the overall loss function, The data fitting residual term represents the predicted temperature field calculated by the physical-guided neural differential operator network. The actual temperature of the peripheral nodes collected by the actual sensing equipment The mean square error between the two values ​​serves to ensure the degree of agreement between the predicted results and the actual measured values. The dynamic fractional physical residual term represents the residual between the evolution derivative of the deduced temperature field and the dynamic adaptive spatiotemporal fractional physical constraint equation. Its function is to ensure that the network deduction process always conforms to the basic physical laws of thermodynamics. Here, KL divergence regularization term represents the posterior probability distribution of the network parameters. With prior probability distribution The differences between them constrain the parameter distribution to suppress overfitting and maintain the stability of parameter updates; , and These are the adaptive dynamic weighting coefficients corresponding to the three loss terms, used to balance the weights of various constraints in real time during online training. This overall loss function calculates the gradient of the network parameters through backpropagation. During operation, it only needs a very small window of the latest high-confidence measured data to complete the online posterior update of the parameters, completely abandoning the lengthy feedback mechanism of traditional methods that rely on external independent closed-loop modules. This gives the algorithm independent, lifecycle-long parameter self-healing capabilities.

[0078] S4: Output the three-dimensional temperature field distribution data inside the ring network box obtained from the deduction.

[0079] This step specifically includes: after outputting the three-dimensional temperature field distribution data inside the ring network box, it further includes locking the highest temperature hot zone node based on the three-dimensional temperature field distribution data and calculating the thermodynamic safety margin. The thermodynamic safety margin represents the percentage difference between the current highest temperature hot zone node and the preset insulation medium temperature resistance threshold. It also combines the preset insulation medium thermal aging activation energy with the calculation of accumulated dynamic thermal aging loss. The accumulated dynamic thermal aging loss is obtained by time-domain integration of the instantaneous thermal aging rate within a predetermined time interval. This step directly reads the set of temperature state vectors of each physical graph node output in S3. First, a global search is used to locate the node with the highest temperature in the temperature field, and its corresponding highest temperature value is defined as... Then, the temperature resistance threshold of the insulating medium, which is pre-stored in the local database, is read. And calculate the thermodynamic safety margin accordingly. ,Right now

[0080]

[0081] Thermodynamic safety margin It is the percentage difference between the current highest temperature and the failure threshold of the insulation material. Its purpose is to intuitively quantify the safety margin of the equipment's current operating temperature from the physical destructive boundary. A value close to 1 indicates very safe operation, while a value close to 0 indicates that the equipment is on the verge of thermal runaway. Simultaneously, the algorithm calculates the instantaneous thermal aging rate of the insulation material in real time to address the thermal aging effect. Its calculation formula is

[0082]

[0083] in, This is the rated thermal aging rate of the insulating medium at the reference temperature, and its value is set in advance based on the factory test data of the insulating material. The preset thermal aging activation energy for insulating media is typically around 1.0 electron volts for epoxy resin insulating media. Boltzmann's constant; This represents the highest absolute thermodynamic temperature at the current moment. To obtain the degree of aging of the equipment throughout its entire lifespan, the cumulative dynamic thermal aging loss is further calculated. Its definition is

[0084]

[0085] in, This refers to the moment the equipment was put into operation or the start time of the last recorded loss. This refers to the current moment; the accumulated loss. Obtained through instantaneous integration, it serves to characterize the degree of irreversible equivalent aging of the insulating medium caused by long-term high-temperature operation, and is an irreplaceable quantitative indicator for predicting remaining life.

[0086] After calculating the safety margin and thermal aging loss, this step also includes generating tiered decision instructions based on the thermodynamic safety margin and temperature change trends. These tiered decision instructions include cloud-based macro-level operational and maintenance early warning information and edge-based proactive control instructions. This decision instruction generation process is based on the thermodynamic safety margin. Maximum temperature change rate and cumulative thermal aging loss Based on a comprehensive assessment, when the thermodynamic safety margin falls below a first threshold, a cloud-based macroscopic operation and maintenance early warning message is generated and uploaded to the power distribution master station system via a communication module. The first threshold is set as follows: to One of the values ​​between, the preferred value is This indicates that the equipment has reached a safety margin of 20% before the insulation failure threshold. At this point, the algorithm determines that the equipment has entered a non-urgent but noteworthy thermal risk zone and generates a macro-level maintenance early warning. This information is uploaded to the distribution master station system via the communication module to guide dispatchers in arranging planned load adjustments or inspection tasks. When the thermodynamic safety margin falls below a second threshold, an active control command is generated and sent to the ring main unit's operating mechanism to execute load reduction or tripping actions. The second threshold is set as follows: to One of the values ​​between, the preferred value is This indicates that the safety margin between the current highest temperature and the insulation failure threshold is extremely close to the physical limit, necessitating immediate active protection measures. This active control command is sent via the control bus to the electric spring operating mechanism of the ring main unit, driving it to directly execute load reduction or even emergency tripping, thereby cutting off the heat source before a thermal fault occurs and preventing insulation failure. In summary, this step, by seamlessly receiving the three-dimensional temperature field data from S3, sequentially completes the location of the hot zone, the quantitative calculation of safety margin and aging loss, and the dynamic decision-making of multi-level hierarchical commands. This elevates the method of this invention from a simple temperature field prediction tool to an edge self-decision-making device with forward-looking active thermal management capabilities, fully ensuring the safe operation of the ring main unit under complex conditions.

[0087] The method disclosed in this application can analyze and predict the internal operating temperature of power equipment such as ring main units, so as to accurately determine the three-dimensional temperature field distribution data inside the power equipment, so as to facilitate active thermal management of the power equipment, thereby helping to ensure and improve the operating quality of the power equipment.

[0088] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to the embodiments of this application, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0089] 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 disclosure 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 foregoing disclosed 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 highly reliable method for predicting the operating temperature field of power equipment, characterized in that, include: The system acquires multi-source sensor data collected in real time by the primary and secondary fusion sensing layers, including the three-phase load current and three-phase voltage output by the capacitive voltage divider sensor and the Rogowski coil current sensor, as well as the external ambient temperature and internal insulating gas pressure output by the temperature, humidity and pressure sensors deployed in the gas box. Based on multi-source sensor data and a pre-structured database, the internal heat conduction entities of power equipment are mapped as physical graph nodes, and an adjacency matrix of the physical topology graph is constructed according to the heat conduction path. Multi-source sensor data and adjacency matrix are input into a physical-guided neural differential operator network. This network continuously evolves the thermodynamic state through dynamically adaptively generated spatiotemporal fractional physical constraints. It also performs discrete integration of the temperature evolution modeled as an ordinary differential equation through a preset numerical solver. Furthermore, the network parameters are self-corrected online based on implicit Bayesian variational inference during the derivation process. The output is based on the calculated three-dimensional temperature field distribution data inside the power equipment.

2. The method according to claim 1, characterized in that, After acquiring multi-source sensor data, the process also includes performing weighted median filtering on the multi-source sensor data to filter out electromagnetic transient pulse spikes caused by switching operations. The weighted median filtering process uses a preset sliding window to perform robust statistics on high-frequency sampling signals, and uses a multi-sensor spatiotemporal alignment algorithm with a built-in Kalman filter to map the millisecond-level response of electrical quantities and the second-level hysteresis response of heat to the same system time base for phase compensation and timestamp alignment.

3. The method according to claim 1, characterized in that, The internal insulating cylinder, arc-extinguishing chamber, instrument transformer, stainless steel gas box inner wall and internal insulating gas area of ​​the power equipment are defined as nodes of the physical graph, and the adjacency matrix of the physical topology graph is constructed based on whether there is a direct physical contact surface or convective heat transfer path between each physical graph node.

4. The method according to claim 1, characterized in that, The multi-source sensor data is input into the state-aware network built into the physical-guided neural differential operator network. The state-aware network outputs the time fractional order and spatial fractional order in real time. The time fractional order and spatial fractional order are then substituted into the thermodynamic partial differential equation based on the time fractional order derivative and the spatial fractional order derivative to form the dynamically adaptively generated spatiotemporal fractional order physical constraint.

5. The method according to claim 1, characterized in that, The physical-guided neural differential operator network uses a graph attention mechanism to aggregate the temperature features of neighboring nodes for each physical graph node in the physical topology graph and calculates the attention weights. Then, a graph transformer network is used to perform nonlinear transformation on the aggregated temperature features and attention weights to simulate the thermal diffusion process of physical heat on the physical topology graph.

6. The method according to claim 1, characterized in that, The change of the three-dimensional temperature field inside the power equipment over time is modeled as an ordinary differential equation, and the ordinary differential equation is discretely integrated using a preset numerical solver to output the temperature state of each physical graph node at the next moment. The preset numerical solver is a high-order Runge-Kutta solver based on the time discrete step size.

7. The method according to claim 1, characterized in that, The online self-calibration of network parameters based on implicit Bayesian variational inference includes constructing an overall loss function for the physical-guided neural differential operator network. The overall loss function includes a data fitting residual term, a dynamic fractional physical residual term, and a KL divergence regularization term representing the difference between the posterior and prior probability distributions of the network parameters. The overall loss function is then used to perform online posterior updates of the network parameters of the physical-guided neural differential operator network.

8. The method according to claim 1, characterized in that, After outputting the three-dimensional temperature field distribution data inside the power equipment, the method further includes locking the highest temperature hot zone node based on the three-dimensional temperature field distribution data and calculating the thermodynamic safety margin. The thermodynamic safety margin represents the percentage difference between the current highest temperature hot zone node and the preset insulation medium temperature resistance threshold. It is also combined with the preset insulation medium thermal aging activation energy calculation and cumulative dynamic thermal aging loss. The cumulative dynamic thermal aging loss is obtained by time-domain integration of the instantaneous thermal aging rate within a predetermined time interval.

9. The method according to claim 8, characterized in that, It also includes generating tiered decision instructions based on the thermodynamic safety margin and temperature change trend. The tiered decision instructions include cloud-based macro-level operation and maintenance early warning information and edge active control instructions. When the thermodynamic safety margin is lower than a first threshold, cloud-based macro-level operation and maintenance early warning information is generated and uploaded to the power distribution master station system through the communication module. When the thermodynamic safety margin is lower than a second threshold, active control instructions are generated and sent to the power equipment operating mechanism to perform load reduction or tripping actions.

10. The method according to claim 1, characterized in that, The multi-source sensor data is acquired and output in real time by the capacitive voltage divider sensor and Rogowski coil current sensor integrated inside the pole of the pole-mounted circuit breaker. The real-time forward inference and online self-correction calculation of the physical-guided neural differential operator network are deployed in the distribution network terminal unit integrated with a neural network acceleration unit. The dynamically adaptively generated spatiotemporal fractional physical constraints and implicit Bayesian variational inference are both executed in the distribution network terminal unit.

Images