Power distribution network equipment digital twin modeling method and system based on multi-source data fusion and 3DGS neural rendering

Abstract

The application discloses a power distribution network equipment digital twin modeling method and system based on multi-source data fusion and 3DGS neural rendering. The method comprises the following steps: constructing a state machine driven mechanism model, outputting structure constraint parameters and an optimization objective function; collecting multi-source data according to the structure constraint parameters and the optimization objective function and generating multi-modal data through preprocessing; initializing a 3DGS scene representation by using the multi-modal data and adaptively configuring initial rendering attributes according to the structure constraint parameters and a sensitivity index; encoding real-time running state parameters as conditional features, injecting a 3DGS rendering pipeline as a prior constraint of the optimization objective function; and taking an image sequence as a supervision signal and a mechanism model physical constraint as a regularization term to obtain a state-aware 3DGS digital twin model through end-to-end optimization. The application deeply fuses the mechanism model and the neural rendering, realizes the synchronous convergence of the power distribution network equipment digital twin model in visual reality and physical consistency, and dynamically responds to and visually expresses the running states such as overload, short circuit and aging.

Description

Technical Field

[0001] This invention belongs to the field of digital simulation and digital twin technology of power systems, specifically involving a digital twin modeling method and system for distribution network equipment based on multi-source data fusion and 3DGS neural rendering. Background Technology

[0002] With the rapid development of smart grid and digital twin technologies, refined modeling and real-time visualized operation and maintenance of distribution network equipment have become key aspects of the digital transformation of power systems. Traditional modeling methods for distribution network equipment mainly employ numerical simulation based on physical mechanisms or computer graphics rendering based on three-dimensional geometry. Both have their limitations: while mechanism models can accurately describe the electrical, thermal, and mechanical behavior of equipment, they suffer from high computational complexity, poor real-time performance, and difficulty in intuitively presenting the appearance and dynamic changes of the equipment; while pure geometric rendering models, although possessing high visual realism, lack physical consistency and cannot reflect the internal physical field evolution of equipment under abnormal conditions such as overload, short circuit, and insulation aging.

[0003] In recent years, 3D Gaussian Splatting (3DGS), as an emerging neural rendering technology, has attracted widespread attention for its high rendering speed and superior scene representation capabilities. However, existing 3DGS methods are mainly geared towards the reconstruction of general scenes and the synthesis of new views. The optimization of its Gaussian properties relies on the supervision signal of the image sequence, lacking the ability to perceive the physical characteristics and operating status of power distribution network equipment, making it difficult to directly apply to digital twin scenarios of power equipment with stringent physical consistency requirements. In addition, the fusion and utilization of multi-source heterogeneous data (such as telemetry data, infrared thermal images, ultrasonic signals, and point cloud data) is still insufficient in the field of power distribution networks. Most existing methods simply superimpose or process different types of data step by step, failing to achieve feature-level deep fusion and unified spatial representation, resulting in insufficient accuracy of the model in characterizing the multi-physics coupling characteristics of equipment.

[0004] Therefore, there is an urgent need for a digital twin modeling method that can organically integrate the mechanism model of power distribution network equipment with 3DGS neural rendering, so as to achieve a physical consistency expression and dynamic response capability of equipment operation status while ensuring visual realism. Summary of the Invention

[0005] To address the technical shortcomings of existing digital twin modeling of power distribution network equipment, which involves the disconnect between mechanism simulation and visual rendering and the difficulty in achieving both physical consistency and real-time visualization, this invention provides a digital twin modeling method and system for power distribution network equipment based on multi-source data fusion and 3DGS neural rendering. This method deeply integrates the state machine-driven mechanism model with the 3DGS neural rendering pipeline, initializes the scene representation through feature-level fusion of multi-source heterogeneous data, and optimizes the Gaussian parameters end-to-end using the physical constraints of the mechanism model as the regularization term, thereby achieving synchronous convergence of the digital twin model in both visual realism and physical consistency.

[0006] In a first aspect, embodiments of this application provide a method for digital twin modeling of power distribution network equipment based on multi-source data fusion and 3DGS neural rendering, the method comprising:

[0007] S1. Construct a state machine-driven mechanism model, and output structural constraint parameters and optimization objective function;

[0008] S2. Collect multi-source data based on the structural constraint parameters and optimization objective function, and generate multimodal data after preprocessing;

[0009] S3. Initialize the 3DGS scene representation using the multimodal data, and adaptively configure the initial rendering attributes according to the structural constraint parameters and sensitivity index;

[0010] S4. Encode the real-time running status parameters into conditional features, and use the optimization objective function corresponding to the current state as a priori constraint to inject them into the 3DGS rendering pipeline to achieve the dynamic response of Gaussian properties to the running status.

[0011] S5. Using image sequences as supervision signals and physical constraints of the mechanism model as regularization terms, Gaussian parameters and conditional neural modulation networks are optimized end-to-end to obtain a state-aware 3DGS digital twin model.

[0012] Secondly, embodiments of this application provide a digital twin modeling system for power distribution network equipment based on multi-source data fusion and 3DGS neural rendering, applied to the method described in the first aspect, the system comprising:

[0013] The mechanism model construction module is used to build state machine-driven mechanism models and output structural constraint parameters and optimization objective functions;

[0014] The multi-source data acquisition and preprocessing module is used to acquire multi-source data based on the structural constraint parameters and optimization objective function, and generate multimodal data through preprocessing.

[0015] The 3DGS scene initialization module has its input end connected to the output end of the mechanism model construction module and the output end of the multi-source data acquisition and preprocessing module. It is used to initialize the 3DGS scene representation using the multimodal data and adaptively configure the initial rendering attributes according to the structural constraint parameters and sensitivity index.

[0016] The conditional neural modulation module, whose input is connected to the output of the mechanism model construction module, is used to encode real-time running state parameters into conditional features, and inject them into the 3DGS rendering pipeline with the optimization objective function corresponding to the current state as a prior constraint to achieve the dynamic response of Gaussian properties to the running state.

[0017] The joint optimization training module has its input terminals connected to the output terminals of the 3DGS scene initialization module, the conditional neural modulation module, and the mechanism model construction module, respectively. It is used to optimize the Gaussian parameters and modulation network end-to-end using the image sequence as the supervision signal and the physical constraints of the mechanism model as the regularization term, so as to obtain a state-aware 3DGS digital twin model.

[0018] Thirdly, embodiments of this application provide an electronic device, including:

[0019] processor;

[0020] Memory used to store processor-executable instructions;

[0021] The processor is configured to implement the digital twin modeling method for power distribution network equipment based on multi-source data fusion and 3DGS neural rendering as described in the first aspect when executing the instructions.

[0022] Fourthly, embodiments of this application provide a computer-readable storage medium storing a program that instructs a device to execute the digital twin modeling method for power distribution network equipment based on multi-source data fusion and 3DGS neural rendering as described in the first aspect.

[0023] Compared with existing technologies, this invention has the following advantages: First, this invention deeply integrates the state machine-driven mechanism model of power distribution network equipment with the 3DGS neural rendering pipeline. The physical constraints output by the mechanism model serve as regularization terms to guide Gaussian parameter optimization, overcoming the lack of physical consistency in traditional pure visual rendering models. This allows the digital twin model to accurately reflect the internal physical field distribution patterns of equipment under different states such as normal operation, overload, short circuit, and insulation aging, while ensuring high-fidelity visual effects. Second, this invention uses a feature-level fusion strategy to spatially register and uniformly represent multi-source heterogeneous data (telemetry data, infrared thermography, ultrasonic signals, and point cloud data), and adaptively configures the initial rendering attributes of the 3DGS scene using sensitivity indicators. Compared with traditional simple overlay or step-by-step processing methods, this significantly improves the model's representation accuracy and initialization efficiency in multi-physics coupled scenarios. Third, this invention introduces a conditional neural modulation mechanism, encoding real-time operating state parameters as conditional features and dynamically modulating Gaussian attributes through a self-attention mechanism. This enables the digital twin model to have real-time dynamic response capabilities to abnormal states such as overload, short circuit, and aging, effectively supporting intelligent operation and maintenance and early warning decision-making for distribution network equipment. Fourth, this invention uses multi-view image sequences and infrared thermal imaging sequences as supervision signals, jointly optimizing Gaussian geometric parameters and modulation network weights. This allows the model to converge simultaneously in both visual realism and physical consistency, balancing rendering quality and simulation accuracy. Furthermore, the efficient rendering characteristics of 3DGS ensure the real-time interactive frame rate of the digital twin scene, meeting the engineering application needs of power operation and maintenance sites. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of a digital twin modeling method for power distribution network equipment based on multi-source data fusion and 3DGS neural rendering, provided as an embodiment of this application.

[0025] Figure 2 This is an architecture diagram of a digital twin modeling system for power distribution network equipment based on multi-source data fusion and 3DGS neural rendering, provided as an embodiment of this application.

[0026] Figure 3 A schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0027] 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 a part of the embodiments of this application, and not all of them.

[0028] It should be noted that in the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application.

[0029] Based on the embodiments described in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0030] Example 1

[0031] Figure 1 This is a schematic flowchart illustrating a digital twin modeling method for power distribution network equipment based on multi-source data fusion and 3DGS neural rendering, provided as an embodiment of this application. Figure 1 As shown, a digital twin modeling method for power distribution network equipment based on multi-source data fusion and 3DGS neural rendering includes:

[0032] S1. Construct a state machine-driven mechanistic model, outputting structural constraint parameters and optimization objective functions. Establish a mathematical model of state transitions describing the behavior of the equipment, extract the structural constraints of the equipment, and set optimization objectives that need to be considered simultaneously in subsequent modeling processes (such as computational accuracy and display smoothness), providing theoretical constraints and optimization directions for the entire digital twin framework.

[0033] Specifically, in this embodiment, step S1 includes:

[0034] S1.1 Based on the physical characteristics and control logic of distribution network equipment, a finite set of states including four states—normal operation, overload, short circuit, and insulation aging—is defined, along with electrical quantity threshold events that trigger state transitions, thus constructing a state machine-driven mechanism model. In this embodiment, a state machine-driven mechanism model is first constructed based on the physical operating characteristics and control and protection logic of the distribution network equipment. A state machine is a mathematical abstraction tool used to describe the switching of a system between different operating modes according to predetermined rules. Specifically, this embodiment defines a finite set of states including four states: normal operation, overload, short circuit, and insulation aging. Normal operation refers to all electrical parameters of the equipment being within the rated operating range; overload refers to a continuous overcurrent condition where the operating current exceeds the rated value but does not trigger protection action; short circuit refers to an extreme high-current fault condition caused by abnormal electrical connections between components; and insulation aging refers to the gradual process of irreversible degradation of insulation materials due to long-term electrical, thermal, and mechanical stress. Simultaneously, electrical quantity threshold events that trigger transitions between states are defined; for example, an overload state is determined when the current amplitude exceeds 1.5 times the rated value and the delay exceeds three seconds. By combining and configuring the above state sets and transition rules, a state machine mechanism model that can reflect the dynamic behavior of distribution network equipment is formed, providing theoretical support for subsequent structural parameter extraction and optimization target setting.

[0035] S1.2 Extract the structural constraint parameters of the equipment's geometric dimensions, phase-to-phase distance, and heat dissipation path based on the aforementioned mechanism model.

[0036] In this embodiment, structural constraint parameters of the power distribution network equipment are extracted from the established state machine-driven mechanism model. These structural constraint parameters include three types of physical quantities: equipment geometry, phase-to-phase distance, and heat dissipation path. Equipment geometry refers to the length, width, and height of the equipment in three-dimensional space, as well as the relative spatial positions of its components. Phase-to-phase distance refers to the spatial interval between different phase conductors; this parameter directly affects the distribution of the electric field intensity inside the equipment and the electrodynamic amplitude during short-circuit faults. Heat dissipation path refers to the physical channels through which heat is transferred from the heat source inside the equipment to the external environment via conduction, convection, and radiation, and its thermal resistance characteristics. These structural constraint parameters are determined based on equipment design drawings and physical principles, reflecting the actual spatial layout and heat transfer characteristics of the equipment. They will subsequently serve as hard constraints limiting the relative displacement range of Gaussian elements in the three-dimensional scene, ensuring that the rendering results conform to the actual physical layout of the equipment at the spatial geometric level.

[0037] S1.3. With the goals of minimizing the device temperature field prediction error and maximizing the 3D rendering frame rate, a multi-objective optimization function is constructed, and the structural constraint parameters and optimization objective function are used as prior inputs for the subsequent rendering process.

[0038] In this embodiment, a multi-objective optimization function is constructed, comprehensively considering two core performance indicators of digital twin models in engineering applications. The first indicator is minimizing the temperature field prediction error, which requires minimizing the deviation between the temperature distribution output by the model and the theoretical temperature distribution calculated by the mechanistic model, to ensure the physical credibility of the twin model at the thermal level. The second indicator is maximizing the 3D rendering frame rate, which requires maximizing the number of frames generated by the model per unit time to meet the response speed requirements of real-time interaction and inspection operations for power operation and maintenance personnel. There is an inherent trade-off between the two indicators. This embodiment constructs a multi-objective optimization function that includes the above two sub-objectives, and uses its output value as a priori constraint for determining the optimization direction and adjustment range in the subsequent rendering process. At the same time, the structural constraint parameters extracted in step S1.2 are also input into the subsequent rendering pipeline as hard constraints on spatial geometric layout.

[0039] In this embodiment, the multi-objective optimization function is specifically represented in the following weighted summation form:

[0040] ,

[0041] in, It represents the overall output value of the multi-objective optimization function, which serves as a priori constraint scalar for the optimization direction and adjustment magnitude during subsequent rendering. This represents the temperature field prediction error; the smaller the value, the more accurate the model's thermal prediction. This represents the 3D rendering frame rate; a higher value indicates better real-time rendering performance. These are the weighting coefficients for the temperature field prediction error term and the rendering frame rate term, respectively. The mean square error between the predicted temperature and the temperature calculated by the mechanistic model at all Gaussian nodes is calculated by summing up the mean square error. This is obtained by real-time statistics of the number of frames output per unit time in the rendering pipeline. Because... and There are differences in units and numerical ranges. Before substituting them into the multi-objective optimization function, the following are respectively... and Perform max-min normalization to map the two to the same order of magnitude before weighted summation.

[0042] S2. Collect multi-source data based on the structural constraint parameters and optimization objective function, and generate multimodal data through preprocessing. According to the structural characteristics of the equipment and modeling requirements, acquire various types of data from different sources, such as historical operation records, infrared images, ultrasonic detection signals, and 3D point clouds. After denoising, time synchronization, and coordinate alignment, the heterogeneous information is fused into a unified representation of multimodal data, providing multi-dimensional basic information for scene construction.

[0043] Specifically, in this embodiment, step S2 includes:

[0044] S2.1 Based on the equipment topology connection relationship and electrical boundary conditions in the structural constraint parameters, collect historical telemetry data, infrared thermal imaging video, partial discharge ultrasonic signals and equipment ledger point cloud data of the power distribution network equipment to construct the original multi-source dataset.

[0045] In this embodiment, the scope and focus of data collection are determined based on the equipment topology connections and electrical boundary conditions included in the structural constraint parameters extracted in step S1.2. Equipment topology connections refer to the electrical connection structures formed between distribution network equipment via busbars, cables, or overhead lines; electrical boundary conditions refer to the range and limits of electrical quantities such as voltage and current under normal operating and fault conditions. Based on the above information, four types of raw data are collected: historical telemetry data (including time-series electrical parameters such as current, voltage, active power, and reactive power), infrared thermal imaging video (a sequence of thermal radiation images reflecting the temperature distribution on the equipment surface), partial discharge ultrasonic signals (ultrasonic detection signals generated by insulation defects), and equipment ledger point cloud data (a dense spatial point set on the equipment surface obtained through three-dimensional laser scanning). These four types of data reflect the operating status and physical properties of the equipment from four dimensions: electrical, thermal, insulation, and geometric, respectively, and together constitute the original multi-source dataset.

[0046] S2.2 Based on the modeling accuracy and computational efficiency indicators in the optimization objective function, timestamp alignment and time series interpolation are performed on the original multi-source dataset, and wavelet transform is used to remove electromagnetic noise interference.

[0047] In this embodiment, based on the two indicators of modeling accuracy and computational efficiency included in the optimization objective function constructed in step S1.3, preprocessing operations are performed on the original multi-source dataset. First, the original data from various sampling frequencies and data formats are timestamped to unify the time base of each data source to the same time axis. For data channels with inconsistent time sampling intervals, time-series interpolation is used to fill in the missing data values ​​to ensure the synchronization of each data source in the time dimension. Second, to address the noise interference mixed in the data collected under the strong electromagnetic environment of the power distribution network, wavelet transform is used for denoising. Wavelet transform is a joint time-frequency domain analysis tool that can decompose a signal into different frequency components. By setting a threshold, high-frequency noise components are filtered out and the signal is reconstructed, effectively preserving the useful features in the original signal, improving data quality, and providing a clean data foundation for subsequent fusion processing.

[0048] S2.3. Spatial registration is performed on time-series telemetry data, thermal imaging textures, and ultrasonic spectral features. After mapping to a unified coordinate system, feature-level fusion is used to generate multimodal data containing electrical, thermal, and mechanical states.

[0049] In this embodiment, spatial registration and feature fusion operations are performed on three types of preprocessed heterogeneous data: time-series telemetry data (characterizing changes in electrical state), thermal imaging textures (characterizing spatial distribution of temperature field), and ultrasonic spectral features (characterizing acoustic signal characteristics caused by insulation defects). First, the three types of data are mapped to a unified device spatial coordinate system to ensure a one-to-one correspondence between the data sources in spatial location. After spatial registration, a feature-level fusion strategy is used to integrate the multi-source information. Feature-level fusion refers to combining and associating data from different modalities at the feature extraction level, rather than simply superimposing them at the original data level or decision level. Through feature-level fusion, electrical quantity features, thermal quantity features, and mechanical vibration features are associated with a unified spatial grid node, generating multimodal data that simultaneously contains electrical state information (voltage and current distribution), thermal state information (temperature field distribution), and mechanical state information (structural deformation and vibration characteristics), providing complete and spatially aligned input data for the subsequent initialization of the 3D scene.

[0050] S3. Initialize the 3DGS scene representation using the multimodal data, and adaptively configure the initial rendering attributes based on the structural constraint parameters and sensitivity indicators. Construct the initial 3D scene representation using the fused data, determine the position, size, and appearance of each basic element in the scene, and automatically adjust the display parameters and calculation priorities of each element according to the sensitivity of different parts of the device to temperature changes, so that the initial model focuses on key areas under limited computing resources.

[0051] Specifically, in this embodiment, step S3 includes:

[0052] S3.1. Initialize the position and scale of each Gaussian ellipsoid according to the point cloud information in the multimodal data, and initialize the color and opacity according to the optical properties of the device material.

[0053] In this embodiment, based on the point cloud information in the multimodal data generated in step S2.3, the basic geometric and optical parameters of each Gaussian ellipsoid in the 3DGS scene are initialized. 3DGS (3D Gaussian Splatting) is a scene representation and rendering method based on the superposition of a large number of three-dimensional Gaussian functions. Each Gaussian ellipsoid describes a local area in the scene through a position vector (three-dimensional spatial coordinates), a covariance matrix (which determines the shape, size, and orientation of the ellipsoid), a color vector (RGB three channels), and an opacity scalar (which controls the degree of contribution of the Gaussian ellipsoid to the final image). Specifically, this embodiment uses the spatial distribution of the point cloud in the multimodal data as a reference, and sets the initial position of each Gaussian ellipsoid at the point cloud cluster center. Its scale is determined according to the spatial variance of the local density of the point cloud, so that the Gaussian ellipsoid can cover the geometric details of the corresponding area. At the same time, based on the material optical properties recorded in the equipment ledger (such as the specular reflectivity of metal surfaces, the diffuse reflectance coefficient of insulating materials, and color characteristics), each Gaussian ellipsoid is assigned an initial color value and opacity value, so that the initial scene visually reflects the true material characteristics of the equipment.

[0054] S3.2 Calculate the sensitivity index of each device area under the optimization objective function, where the sensitivity weight of the high thermal sensitivity area is higher than that of the low thermal sensitivity area.

[0055] In this embodiment, the sensitivity index of each device region under the optimization objective function is calculated. This sensitivity index is used to quantify the sensitivity of different regions of the device to temperature changes. Its physical meaning is: when the temperature of a certain region changes by a unit, the rate of change of the corresponding error term in the optimization objective function for that region. The calculation method is as follows: taking the temperature field prediction error minimization term in the optimization objective function constructed in step S1.3 as the benchmark, the partial derivative of this term with respect to the input characteristics of each device region (such as the current density, heat flux density, etc. of that region) is obtained, and this is used as the thermal sensitivity coefficient of that region. Based on this, combined with the inherent thermal diffusivity of the device material (reflecting the material's ability to conduct temperature changes) and the current density distribution of that region, the above thermal sensitivity coefficient is normalized, compressing the numerical range to the interval between zero and one, forming the final sensitivity index. According to the physical definition of this index, the sensitivity weight of high thermal sensitivity regions (such as conductor connection points, transformer winding hot spots) is significantly higher than that of low thermal sensitivity regions (such as the shell, supporting insulators). This difference will guide computing resources to tilt towards key regions in subsequent scenario configurations.

[0056] Furthermore, in S3.2, the calculation of the sensitivity index of each device region under the optimization objective function specifically includes:

[0057] S3.2.1. Using the partial derivatives of the temperature field prediction error minimization term in the objective function with respect to the input characteristics of each equipment region as a benchmark, calculate the thermal sensitivity coefficient of each equipment region.

[0058] In this embodiment, the thermal sensitivity coefficient of each equipment region is calculated based on the temperature field prediction error minimization term in the optimization objective function constructed in step S1.3. Specifically, the temperature field prediction error minimization term is regarded as a function with the input characteristics of each equipment region (including physical quantities such as current density, heat flux density, and material thermal conductivity of the region) as independent variables. The partial derivative of this function with respect to the input characteristics of each region is obtained, and the resulting partial derivative value is the thermal sensitivity coefficient of that region. The physical meaning of this coefficient is: when the input characteristics of a certain region undergo a unit small change, the rate at which the temperature field prediction error changes accordingly. The larger the absolute value of this coefficient, the more significant the impact of the temperature change in that region on the overall temperature field prediction accuracy, that is, the higher the sensitivity of that region at the thermal level.

[0059] In this embodiment, the first The thermal sensitivity coefficient of each equipment area is calculated using the following partial derivative formula:

[0060] ,

[0061] in, For the first The thermal sensitivity coefficient of each equipment area reflects the degree of influence of changes in the input characteristics of that area on the temperature field prediction error. The temperature field prediction error is defined in step S1.3; For the first The input feature vector of each device region ; For the first The current density in each region was obtained through finite element current field simulation. For the first The heat flux density of each region was obtained through thermal simulation. For the first The thermal conductivity of the material in each region is retrieved from the equipment material property database. The partial derivatives are calculated using an automatic differentiation method, specifically, based on the temperature field prediction error constructed in step S1.3. As a scalar output, the feature vectors of each region are used as input. As input variables, a computational graph is constructed, and all region pairs are solved in one go using the backpropagation algorithm. The gradient value.

[0062] S3.2.2. Combining the thermal diffusivity and current density distribution of the equipment materials, the thermal sensitivity coefficient is normalized to obtain a sensitivity index with a value range of zero to one. In this embodiment, the thermal sensitivity coefficient calculated in step S3.2.1 is normalized to compress its numerical range to the interval between zero and one, forming a unified sensitivity index. The normalization process comprehensively considers two types of physical factors: first, the thermal diffusivity of the equipment materials, which reflects the rate at which the material conducts temperature changes; the higher the thermal diffusivity, the faster the temperature disturbance propagates, and the wider its influence on the overall temperature field; second, the current density distribution in the region; the higher the current density, the greater the Joule heat generation rate per unit volume, and the more significant the amplitude and rate of temperature change. By combining the above two types of physical quantities to perform weighted normalization of the thermal sensitivity coefficient, the final output sensitivity index can be fairly compared between different regions of the equipment, with clear physical meaning and unified dimensions.

[0063] In this embodiment, the normalization process maps the thermal sensitivity coefficient to a sensitivity index with a value range of [0,1] using the following formula:

[0064] ,

[0065] in, For the first Sensitivity indicators for each device area; For the Sigmoid function; For the first Thermal diffusivity of materials in each region; For the first Current density in each region; This is the arithmetic mean of the thermal diffusivity of all areas of the equipment; This is the arithmetic mean of the current density across all areas of the device. The first one calculated in step S3.2.1 The thermal sensitivity coefficient of each region. First, calculate the average thermal diffusivity of all regions of the equipment. and the average current density Then calculate the ratio for each region one by one. / and / ,and After multiplication, the result is input into the Sigmoid function to complete the normalization mapping.

[0066] S3.2.3 Map the sensitivity index to a scaling factor of the initial opacity. For device regions where the sensitivity index is greater than a preset threshold, the initial opacity of the corresponding Gaussian is increased to 0.2 to 0.5 times the base value, and higher rendering computing resources are allocated to them.

[0067] In this embodiment, the sensitivity index generated in step S3.2.2 is mapped to the initial rendering attribute configuration parameters of the Gaussian ellipsoid in the 3DGS scene. The mapping rules are as follows: For device areas where the sensitivity index is greater than a preset threshold, i.e., areas where temperature changes have a significant impact on the overall temperature field prediction accuracy, the initial opacity of the corresponding Gaussian ellipsoid is increased to 0.2 to 0.5 times the base value, making the area more visually salient in the initial scene rendering, facilitating the monitoring signal's sensitive response to errors in this area during subsequent optimization. Simultaneously, higher rendering computational resource priority is allocated to these high-sensitivity areas; that is, in the parallel computational task scheduling of the rendering pipeline, priority is given to ensuring the Gaussian attribute update frequency and rendering accuracy of these areas, ensuring that computational resources are concentrated on physically critical areas. Through the above mapping configuration, a quantitative transfer from physical sensitivity to rendering attributes is achieved, enabling the initial state of the 3DGS scene to reflect the emphasis on key physical areas in both visual performance and computational resource allocation.

[0068] S3.3. Based on the structural constraint parameters, limit the relative displacement range between adjacent Gaussians, and adaptively adjust the initial opacity and rendering priority of each Gaussian in combination with the sensitivity index to generate an initial 3DGS scene with physical interpretability.

[0069] In this embodiment, the initial configuration of the 3DGS scene is optimized by combining structural constraint parameters and sensitivity indices. First, based on the structural constraint parameters extracted in step S1.2, the relative displacement range between adjacent Gaussian ellipsoids is limited. That is, the spatial distance between the centers of adjacent Gaussians must not exceed the maximum value allowed by the actual geometric dimensions of the device, nor be less than the minimum value determined by the point cloud sampling resolution. This ensures that the spatial layout of the scene geometrically conforms to the actual physical structure of the device, avoiding misalignment or overlap that violates topological constraints. Second, the sensitivity indices calculated in step S3.2 are mapped to the adjustment basis of two specific configuration parameters: one is the scaling factor of the initial opacity. For device areas with sensitivity indices greater than a preset threshold (i.e., thermally sensitive areas), the initial opacity of the corresponding Gaussian ellipsoid is increased to 0.2 to 0.5 times the base value, making these areas more visually salient in the initial rendering. The other is the priority allocation of rendering computing resources. Areas with higher sensitivity indices are assigned higher rendering priority, that is, when computing resources are limited, priority is given to ensuring the rendering accuracy and update frequency of these areas. With the above configuration, an initial 3DGS scene with physical interpretability is generated. This scene is subject to structural constraints in terms of geometric layout and guided by sensitivity indicators in terms of visual weight, providing a physically reasonable initial state for the dynamic modulation of subsequent state responses.

[0070] S4. Encode real-time operating status parameters as conditional features, using the optimization objective function corresponding to the current state as a priori constraint, and inject them into the 3DGS rendering pipeline to achieve a dynamic response of Gaussian properties to the operating state. Convert the operating parameters currently collected by the device into control signals that the model can recognize, calculate the degree of influence of the operating state on each area in the scene through an attention mechanism, and dynamically update the geometric and appearance attributes of scene elements accordingly, so that the visualization model can be synchronously adjusted in real time according to the device's operating state (such as load changes, failure occurrence).

[0071] Specifically, in this embodiment, step S4 includes:

[0072] S4.1 Input the real-time collected three-phase current, voltage and ambient temperature and humidity parameters into the fully connected network and encode them into high-dimensional conditional feature vectors.

[0073] In this embodiment, the real-time collected operating status parameters of the power distribution network equipment are input into a fully connected network and encoded into a high-dimensional conditional feature vector. The fully connected network is a multi-layer neural network structure, where each neuron in each layer is connected to the output of all neurons in the previous layer. Through a composite mapping of layer-by-layer linear transformations and nonlinear activation functions, the data in the original input space is projected into a high-dimensional feature space. Specifically, the input parameters collected in this embodiment include three-phase current (real-time current amplitude and phase of phases A, B, and C), three-phase voltage (real-time voltage amplitude and phase of each phase), and ambient temperature and humidity (temperature and relative humidity of the environment where the equipment is located). After being encoded by the fully connected network, these original parameters form a high-dimensional conditional feature vector containing comprehensive information about the operating status. This feature vector serves as the conditional input for the subsequent self-attention mechanism, characterizing the current operating condition of the equipment.

[0074] S4.2 Calculate the association weights between the conditional feature vector and each Gaussian attribute through a self-attention mechanism, and use the output value of the optimization objective function corresponding to the current running state as a scaling factor.

[0075] In this embodiment, the association weights between the high-dimensional conditional feature vector generated in step S4.1 and each Gaussian attribute in the 3DGS scene are calculated using a self-attention mechanism. The self-attention mechanism is a feature interaction modeling method whose core idea is to calculate the correlation weights between elements at different positions in a sequence, enabling the model to dynamically focus on the information unit most relevant to the current condition. Specifically, this embodiment uses the high-dimensional conditional feature vector as the query vector and the sequence of Gaussian attribute feature vectors composed of the position vectors, scale parameters, and material attribute codes of each Gaussian in the scene as the key vector and value vector. Multi-head self-attention operation is performed, outputting an association weight matrix representing the degree of influence of the current operating state on each Gaussian. The larger the weight value in this matrix, the more important the corresponding Gaussian is in reflecting the current operating state of the device. For example, when an overload occurs, Gaussians located at conductor connection points or winding hotspots will be assigned higher association weights. Based on this, the output value of the optimization objective function corresponding to the current operating state is used as a scaling factor to adjust the overall association weight matrix, ensuring that the modulation amplitude is constrained by the objective function value and avoiding over-modulation that could lead to scene distortion.

[0076] Furthermore, the step of calculating the association weights between the conditional feature vector and each Gaussian attribute through a self-attention mechanism specifically includes:

[0077] S4.2.1 Encode the position, scale, and material properties of each Gaussian ellipsoid into a sequence of Gaussian attribute feature vectors. In this embodiment, the attribute information of all Gaussian ellipsoids in the 3DGS scene is encoded into a unified sequence of feature vectors, serving as the key and value vectors for the self-attention mechanism. Specifically, for each Gaussian ellipsoid in the scene, its position vector (3D spatial coordinates), scale parameters (the length of the ellipsoid's three axes and its spatial orientation represented by the covariance matrix), and material properties (including optical parameters such as surface reflectivity, diffuse reflectance coefficient, and color features) are concatenated and linearly transformed to form a feature vector of fixed length. The feature vectors of all Gaussian ellipsoids are arranged in spatial order to form a sequence of Gaussian attribute feature vectors. This sequence fully describes the geometric structure and optical property distribution of all local regions in the current 3D scene, providing a structured scene representation for subsequent correlation weight calculation.

[0078] S4.2.2 Using the high-dimensional conditional feature vector as the query vector and the Gaussian attribute feature vector sequence as the key vector and value vector, perform multi-head self-attention operation.

[0079] In this embodiment, the high-dimensional conditional feature vector generated in step S4.1 is used as the query vector, and the Gaussian attribute feature vector sequence generated in step S4.2.1 is used as the key vector and value vector, respectively, to perform multi-head self-attention operation. The multi-head self-attention mechanism projects the input features into multiple different subspaces, independently performs scaled dot product attention operations in each subspace, and then concatenates the attention results from each subspace and fuses them through a linear transformation for output. Mathematically, this involves performing a dot product between the query vector and all key vectors, scaling and Softmax normalization to obtain the attention weight distribution, and then using this weight distribution to perform a weighted summation of the value vectors. Through parallel computation of the multi-head structure, the model can capture the correlation patterns between the running state conditions and Gaussian attributes in multiple different feature subspaces, ultimately outputting a correlated feature representation that integrates multi-party attention information, making the correlation assessment more comprehensive and robust.

[0080] S4.2.3. Generate a correlation weight matrix based on the calculation results to characterize the degree of influence of electrical state and ambient temperature on each Gaussian, where the Gaussian corresponding to the overload region or hot spot region is given a higher correlation weight.

[0081] In this embodiment, a correlation weight matrix is ​​generated based on the output of the multi-head self-attention operation, characterizing the influence of the current electrical state and ambient temperature on each Gaussian ellipsoid. The row indices of this matrix correspond to the numbers of each Gaussian ellipsoid in the scene, the column indices correspond to the dimensional components in the high-dimensional conditional feature vector, and the values ​​of the matrix elements reflect the strength of the influence of the corresponding conditional feature dimension on a specific Gaussian attribute. Specifically, Gaussian ellipsoids corresponding to equipment overload areas (such as conductor sections with a sudden increase in current density) or high-temperature areas detected by thermal imaging (such as connection points or winding hotspots with abnormally high temperatures) are assigned higher values ​​in their corresponding elements of the correlation weight matrix. Through this weight allocation, a quantitative correlation mapping between the operating state conditions and the spatial location of the three-dimensional scene is achieved, enabling the adjustment magnitude of the Gaussian attribute during subsequent modulation to accurately reflect the sensitivity of the region to the current operating state.

[0082] S4.3 Multiply the associated weights element by element with the scaling factor to generate dynamic modulation coefficients. Inject the modulation coefficients as scaling bias terms for Gaussian attribute updates into the 3DGS rendering pipeline to adjust the shape, density, and color distribution of the Gaussian in real time, thereby achieving dynamic response of the Gaussian attribute to overload, short circuit, or aging conditions.

[0083] In this embodiment, the correlation weight matrix obtained in step S4.2 is multiplied element-wise with the scaling factor to generate dynamic modulation coefficients. These modulation coefficients are modulation signals that match the dimensions of the Gaussian attributes, containing adjustments to each Gaussian ellipsoid in terms of its current position, scale, color, and opacity. These modulation coefficients are injected into the 3DGS rendering pipeline as a scaling bias term for Gaussian attribute updates. That is, in addition to the regular iterative updates of each Gaussian attribute, these modulation coefficients are superimposed as an additional offset, thereby enabling real-time response of Gaussian attributes to changes in operating state. Specifically, when an overload occurs, the modulation coefficients drive the Gaussian scale of the hotspot area to expand, the color to shift towards the high-temperature color gamut, and the opacity to increase, to visually reflect the temperature rise effect; when a short circuit occurs, the modulation coefficients drive a sudden change in the Gaussian density of the corresponding area and a color shift towards the warning color gamut, to characterize the fault impact; when insulation aging occurs, the modulation coefficients drive a gradual change in the opacity of the Gaussian in the local area, reflecting the gradual degradation of insulation performance. Through the above dynamic modulation mechanism, the digital twin scene can synchronously update its visual presentation as the real-time operating state of the device changes.

[0084] S5. Using image sequences as supervision signals and physical constraints of the mechanism model as regularization terms, the Gaussian parameters and conditional neural modulation network are optimized end-to-end to obtain a state-aware 3DGS digital twin model. Using actual captured image sequences as a reference standard, and simultaneously constraining the model using physical laws, visual fidelity and physical plausibility are taken as optimization objectives. The scene geometric parameters and the weights of the state modulation module are adjusted synchronously to ultimately obtain a state-responsive digital twin model that is both visually realistic and conforms to physical laws.

[0085] Specifically, in this embodiment, step S5 includes:

[0086] S5.1 Using multi-view image sequences and infrared thermal imaging sequences as supervision signals, pixel-level photometric loss and thermal consistency loss are constructed.

[0087] In this embodiment, multi-view image sequences and infrared thermal imaging sequences are used as supervision signals to construct two types of loss functions. Multi-view image sequences refer to visible light image sequences of the device simultaneously acquired from multiple different spatial orientations, used to provide visual supervision of the device's appearance, texture, and contour information. Infrared thermal imaging sequences refer to thermal radiation image sequences of the device's surface temperature distribution simultaneously acquired from the same viewpoint, used to provide thermal supervision of the device's temperature field distribution information. Based on these supervision signals, pixel-level photometric loss and thermal consistency loss are constructed. Pixel-level photometric loss measures the model's reconstruction accuracy at the visual appearance level by comparing the color and brightness differences between the rendered image and the real image pixel by pixel. Thermal consistency loss measures the model's expression accuracy at the thermal physics level by comparing the temperature values ​​between the rendered temperature distribution and the real infrared thermal image pixel by pixel. These two types of loss functions together constitute a joint supervision signal, guiding the model to converge to the true value simultaneously in both visual and thermal dimensions.

[0088] In this embodiment, the pixel-level photometric loss is calculated using the following formula:

[0089] ,

[0090] The thermal uniformity loss is calculated using the following formula:

[0091] ,

[0092] in, It represents pixel-level photometric loss, is dimensionless, and measures the color difference between a rendered image and a real image. It is the set of coordinates of all valid pixels in a visible light image; For set The total number of pixels in; These are the pixel coordinates in the image plane; To render an image in pixels The RGB color vector at that location; To create realistic multi-view images at pixel level RGB color vector at the location; Let L2 be the norm of the vector; The thermal consistency loss, expressed in Kelvin squares, measures the difference between the rendered temperature distribution and the actual infrared thermal image. It is the set of coordinates of all valid pixels in the infrared thermal image; For set The total number of pixels in; To render an image in pixels The temperature value at that location is obtained by weighted summation of the temperature attributes of each Gaussian node in the 3DGS scene along the viewing direction. For true infrared thermal imaging at the pixel level The temperature value at that location.

[0093] The multi-view image sequence was simultaneously acquired by at least eight industrial cameras at different spatial orientations, while the infrared thermal imaging sequence was simultaneously acquired by an infrared thermal imager positioned at the same viewpoint. During rendering, RGB color maps and temperature maps were output through the 3DGS α-blending rendering pipeline and compared pixel-level with the visible light images and infrared thermal images, respectively. Both loss terms used the square of the L2 norm as a metric to amplify the contribution of larger deviations to the gradient.

[0094] S5.2. The heat balance equation and electric field intensity distribution output by the aforementioned mechanism model are used as physical regularization terms.

[0095] In this embodiment, the heat balance equation and electric field intensity distribution output by the mechanistic model constructed in step S1 are used as physical regularization terms to impose physical consistency constraints on the model training process. The heat balance equation describes the conservation relationship between heat generation, conduction, and dissipation inside the device, i.e., the Joule heat generated per unit volume plus the heat flowing in minus the heat flowing out equals the increase in thermal energy within that volume. This equation has a fundamental constraint on the spatial distribution of the device's temperature field. The electric field intensity distribution describes the magnitude and direction of the electric field vector at various spatial locations inside the device, and it follows the distribution law determined by the Poisson equation for the electrostatic field, i.e., the divergence of the electric field intensity equals the space charge density divided by the dielectric constant. These two types of physical laws are added to the overall loss function in the form of regularization loss, so that during the parameter optimization process, in addition to satisfying visual supervision, the deviation between the temperature and electric field distribution predicted by the Gaussian nodes and the calculation results of the mechanistic model is penalized. For overheated and high field strength regions, a larger penalty weight is assigned to the regularization term to suppress hot spots and field strength distortions that violate physical laws in the rendering results, thus ensuring the credibility of the model at the physical level.

[0096] Furthermore, the use of the heat balance equation and electric field intensity distribution output by the aforementioned mechanism model as physical regularization terms specifically includes:

[0097] S5.2.1. Based on the heat balance equation, construct a temperature field canonical loss to constrain the deviation between the predicted temperature of each Gaussian node and the temperature calculated by the mechanism model.

[0098] In this embodiment, a temperature field canonicalization loss is constructed based on the heat balance equation output by the mechanistic model. The heat balance equation describes the conservation relationship between heat generation, conduction, and dissipation within the device, specifically: the Joule heat power generated per unit volume by current passing through conductors, plus the net heat flow power flowing in from the surrounding area, equals the rate of change of stored thermal energy within that volume. This equation provides fundamental physical constraints on the spatial distribution of the device's temperature field. Based on this, the sum of the squares of the deviations between the temperature values ​​predicted by each Gaussian node in the 3DGS scene and the theoretical temperature values ​​calculated by the mechanistic model based on the heat balance equation is used as the temperature field canonicalization loss. The purpose of this loss term is to constrain the temperature predictions of the Gaussian nodes from deviating from the theoretical temperature distribution determined by the physical equation, thereby ensuring that the output of the twin model at the thermal level conforms to fundamental physical laws.

[0099] In this embodiment, the temperature field canonical loss is calculated using the following formula:

[0100] ,

[0101] in, This is the temperature field canonical loss; The total number of Gaussian ellipsoids in the 3DGS scene; For the index of the Gaussian ellipsoid; For the first The temperature values ​​predicted by each Gaussian node are calculated using the temperature properties of the Gaussian ellipsoid during the rendering process. For the first The theoretical temperature values ​​calculated by the mechanism model at each Gaussian node based on the heat balance equation. By solving the heat balance equation Obtain, among which For material density, For isobaric specific heat capacity, Thermal conductivity, For current density, Let be the electrical conductivity. For each Gaussian ellipsoid, The predicted temperatures are pre-calculated using finite element thermal simulation and stored as supervision labels. During training, the rendering pipeline simultaneously outputs the predicted temperatures for each Gaussian node. The deviation between the two is calculated using the mean square error.

[0102] S5.2.2 Construct a canonical loss of electric field intensity based on the Poisson equation of electrostatic field to constrain the gradient change of the Gaussian distribution on the surface of the charged body to conform to the distribution law of the equipotential surface of the electric field.

[0103] In this embodiment, an electric field intensity canonical loss is constructed based on the Poisson equation for electrostatics. The Poisson equation describes the intrinsic relationship between electric field intensity and space charge distribution. Mathematically, the divergence of the electric field intensity is equal to the ratio of space charge density to dielectric constant, which determines the distribution law of electric field equipotential surfaces on the surface of a charged body. That is, the surface of a conductor is an equipotential surface, and electric field lines are always orthogonal to the equipotential surfaces. Based on this, the spatial gradient change of the Gaussian distribution near the surface of a charged body in the 3DGS scene is constrained to conform to the above-mentioned electric field equipotential surface distribution law. The sum of the squares of the deviations between the measured value and the theoretical expected value of the Gaussian gradient is used as the electric field intensity canonical loss. The role of this loss term is to constrain the spatial arrangement of Gaussian ellipsoids related to the surface of the charged body to avoid gradient changes that violate the physical laws of electric fields, thus avoiding phenomena that violate physical common sense, such as distortion of electric field equipotential surfaces or field lines penetrating the surface of conductors, in the rendering results.

[0104] In this embodiment, the electric field intensity canonical loss is calculated using the following formula:

[0105] ,

[0106] in, For electric field strength canonical loss; A set of Gaussian ellipsoid indices located near the surface of a charged body; For set The number of Gaussian ellipsoids in the data; The index of the Gaussian ellipsoid located on the surface of the charged body; For the first The electric field intensity vector at each Gaussian node is calculated by the spatial gradient variation of the Gaussian ellipsoid. For the first The divergence of the electric field intensity at each Gaussian node is calculated by using the finite difference method to calculate the gradient change of the Gaussian ellipsoid. For the first The space charge density at each Gaussian node was obtained through electrostatic field simulation. The dielectric constant of the region where the device is located is retrieved from the device's material property database. First, it is calculated using the spatial coordinate gradient of the Gaussian ellipsoid. The numerical value is then used to obtain the divergence value using the finite difference method. (Poisson equation for electrostatic field) As a physical constraint, the Gaussian ellipsoid is constrained to follow the fundamental law that equipotential surfaces are orthogonal to electric field lines in its spatial arrangement. The determination is made if the distance between the center coordinates of the Gaussian ellipsoid and the surface of the device's three-dimensional geometric model is less than a set threshold (e.g., 1 mm).

[0107] S5.2.3. The temperature field regularization loss and the electric field intensity regularization loss are weighted and summed to form a total physical regularization term, wherein the overheated region and the high field strength region are given a greater penalty weight to suppress the appearance of hot spots and field strength distortions that violate physical common sense in the rendering result.

[0108] In this embodiment, the temperature field regularization loss constructed in step S5.2.1 and the electric field intensity regularization loss constructed in step S5.2.2 are weighted and summed to form the total physical regularization term. The weight coefficients of each term in the weighted summation are dynamically determined based on the equipment operating conditions and regional importance. Based on this, greater penalty weights are assigned to overheated regions and high field strength regions: overheated regions refer to spatial regions where the temperature deviation in the temperature field regularization loss exceeds the safety threshold, and high field strength regions refer to spatial regions where the field strength gradient deviation in the electric field intensity regularization loss exceeds the design threshold. By applying higher weight coefficients to these regions in the regularization term, the model imposes greater penalties on hot spots and field strength distortion regions during parameter optimization, effectively suppressing anomalies that violate physical common sense in the rendering results from a physical constraint perspective, further enhancing the physical credibility and robustness of the digital twin model.

[0109] In this embodiment, the total physical regularization term is calculated using the following weighted summation formula:

[0110] ,

[0111] Among them, the regional penalty weight and They are defined as follows:

[0112] , ,

[0113] in, This is the overall physical regularization term, which is a component of the overall loss function in the joint optimization. This refers to the temperature field canonical loss defined in step S5.2.1; The electric field intensity canonical loss is defined in step S5.2.2; The dynamic penalty weights for the temperature field canonical loss; This represents the dynamic penalty weight for the electric field strength regularization loss. For the first The absolute value of the deviation between the temperature predicted by each Gaussian node and the temperature calculated by the mechanistic model. , For the first The absolute value of the deviation between the electric field intensity divergence at each Gaussian node and the theoretical value of the Poisson equation; This is the sensitivity adjustment parameter for temperature deviation; This is a sensitivity adjustment parameter for the electric field strength deviation. For the overheated region ( Spatial regions exceeding 5K) and high field strength regions ( Spatial areas exceeding the design threshold are automatically boosted using the Sigmoid function. and The value of makes these regions subject to greater penalties in the regularization term. and The weights change dynamically with the deviation of each region during training, eliminating the need for manually setting fixed weights.

[0114] S5.3 Jointly optimize the Gaussian geometric parameters of 3DGS and the weight parameters of the modulation network composed of the fully connected network and self-attention mechanism described in step S4, so that the digital twin model converges synchronously in both the visual realism and physical consistency dimensions, and obtains a state-aware 3DGS digital twin model of distribution network equipment.

[0115] In this embodiment, the pixel-level photometric loss and thermal consistency loss constructed in step S5.1, and the physical regularization term constructed in step S5.2, are weighted and combined to form a total loss function, thereby driving end-to-end joint optimization of model parameters. The optimization objects include two types of parameters: the first type consists of the geometric parameters (including position, scale, and rotation direction) and appearance parameters (including color and opacity) of each Gaussian ellipsoid in the 3DGS scene, which determine the geometric accuracy and visual presentation quality of the scene; the second type consists of the weight parameters of the conditional neural modulation network composed of a fully connected network and a multi-head self-attention mechanism in step S4, which determine the model's ability to extract and encode real-time running state features, as well as the quality of the modulation mapping of conditional features to the Gaussian attribute space. By jointly updating these two types of parameters through the backpropagation algorithm, the digital twin model converges simultaneously in both visual realism and physical consistency dimensions. When the total loss function value tends to stabilize and both types of sub-losses reach the preset threshold, the optimization process stops and the final state-aware 3DGS digital twin model is output. This model can simultaneously present a high-fidelity visual appearance and physically reliable temperature and electric field distribution under different operating conditions.

[0116] In this embodiment, the total loss function is defined as the weighted sum of the supervision loss and the physical regularization term:

[0117] ,

[0118] in, The total loss function serves as the objective function for optimizing model parameters. The pixel-level photometric loss defined in step S5.1; The thermal consistency loss defined in step S5.1; The total physical regularization term defined in step S5.2.3; These are the weighting coefficients for the thermal uniformity loss and the physical regularization term, respectively. The three losses differ in dimension; the photometric loss and physical regularization term are dimensionless, while the thermal uniformity loss is in units of K². Before substituting into the total loss function, the dimensions of each term are unified by dividing the thermal uniformity loss by the normalization constant. ( =300K, which is the room temperature reference temperature). The optimization algorithm uses AdamW with an initial learning rate of The weight decay coefficient is The training rounds were set to 50,000 iterations, using a cosine annealing learning rate scheduling strategy. The change is less than 2000 consecutive iterations When the model is deemed to have converged, training is stopped.

[0119] Example 2

[0120] like Figure 2 As shown, this application provides a system architecture diagram for digital twin modeling of distribution network equipment based on multi-source data fusion and 3DGS neural rendering, applied to the digital twin modeling method for distribution network equipment based on multi-source data fusion and 3DGS neural rendering as described in Embodiment 1, including:

[0121] Mechanism model construction module 210 is used to construct a state machine-driven mechanism model and output structural constraint parameters and optimization objective function.

[0122] The multi-source data acquisition and preprocessing module 220 is used to acquire multi-source data based on the structural constraint parameters and optimization objective function, and generate multimodal data through preprocessing.

[0123] The 3DGS scene initialization module 230 has its input end connected to the output end of the mechanism model construction module and the output end of the multi-source data acquisition and preprocessing module. It is used to initialize the 3DGS scene representation using the multimodal data and adaptively configure the initial rendering attributes according to the structural constraint parameters and sensitivity index.

[0124] The conditional neural modulation module 240, whose input is connected to the output of the mechanism model construction module, is used to encode real-time running state parameters into conditional features, and inject them into the 3DGS rendering pipeline with the optimization objective function corresponding to the current state as a prior constraint to achieve the dynamic response of Gaussian properties to the running state.

[0125] The joint optimization training module 250 has its input terminals connected to the output terminals of the 3DGS scene initialization module, the conditional neural modulation module, and the mechanism model construction module, respectively. It is used to optimize the Gaussian parameters and modulation network end-to-end using the image sequence as the supervision signal and the physical constraints of the mechanism model as the regularization term, so as to obtain a state-aware 3DGS digital twin model.

[0126] Figure 3 This is an electronic device provided in one embodiment of this application. For example... Figure 3 As shown, the electronic device includes at least the following components: processor 301 and memory 300, communication interface 303, and bus 302.

[0127] In this embodiment of the application, memory 300 is used to store executable instructions of processor 301, which, when configured to execute instructions, implements the method as described in the first aspect.

[0128] In embodiments of this application, a computer-readable storage medium includes instructions that instruct a device to perform the method as described in the first aspect. For example, the instructions instruct the device to perform... Figure 1 The method is shown in the process steps.

[0129] In one embodiment of this application, the program operating in the electronic device may be a program that controls a central processing unit (CPU) or similar device to achieve the functions of the above-described embodiments of the present invention (a program that enables the computer to function). Information processed by these systems is then temporarily stored in random access memory (RAM) during processing, and subsequently stored in various ROMs such as read-only memory (FlashROM) and hard disk drives (HDDs), and read, corrected, and written by the CPU as needed.

[0130] It should be noted that a portion of the electronic device described in the above embodiments can also be implemented using a computer. In this case, the program for implementing the control function can be recorded on a computer-readable recording medium, and the program recorded on the recording medium can be read into the computer and executed.

[0131] It should be noted that the computer mentioned here refers to a computer built into an electronic device, employing hardware including an operating system and peripheral devices. Furthermore, computer-readable recording media refers to removable media such as floppy disks, magneto-optical disks, ROMs, and CD-ROMs, as well as storage systems such as hard drives built into the computer.

[0132] Furthermore, computer-readable recording media can include: media that dynamically stores programs for short periods of time, such as communication lines used when transmitting programs via networks like the Internet or communication lines like telephone lines; and media that store programs for fixed periods of time, such as volatile memory inside a computer that serves as a server or client in this case. In addition, the aforementioned program can be a program used to implement the above-mentioned functions, or it can be a program that can implement the above-mentioned functions by combining them with programs already recorded in the computer.

[0133] Furthermore, the electronic device in the above embodiments can also be implemented as an assembly (system group) composed of multiple systems. Each system constituting the system group can possess some or all of the functions or functional blocks of the electronic device in the above embodiments. As a system group, it is sufficient to have all the functions or functional blocks of the electronic device.

[0134] Those skilled in the art should recognize that the above embodiments are only used to illustrate this application and are not intended to limit this application. Any appropriate changes and variations made to the above embodiments within the essential spirit and scope of this application fall within the scope of protection claimed in this application.

Claims

1. A method for digital twin modeling of power distribution network equipment based on multi-source data fusion and 3DGS neural rendering, characterized in that, Includes the following steps: S1. Construct a state machine-driven mechanism model, and output structural constraint parameters and optimization objective function; S2. Collect multi-source data based on the structural constraint parameters and optimization objective function, and generate multimodal data after preprocessing; S3. Initialize the 3DGS scene representation using the multimodal data, and adaptively configure the initial rendering attributes according to the structural constraint parameters and sensitivity index; S4. Encode the real-time running status parameters into conditional features, and use the optimization objective function corresponding to the current state as a priori constraint to inject them into the 3DGS rendering pipeline to achieve the dynamic response of Gaussian properties to the running status. S5. Using image sequences as supervision signals and physical constraints of the mechanism model as regularization terms, Gaussian parameters and conditional neural modulation networks are optimized end-to-end to obtain a state-aware 3DGS digital twin model.

2. The method according to claim 1, characterized in that, Step S1 specifically includes: Based on the physical characteristics and control logic of distribution network equipment, a finite set of states including four states—normal operation, overload, short circuit, and insulation aging—is defined, along with electrical quantity threshold events that trigger state transitions, to construct a state machine-driven mechanism model. Based on the aforementioned mechanism model, structural constraint parameters for the device's geometric dimensions, phase-to-phase distance, and heat dissipation path are extracted. With the goals of minimizing the device temperature field prediction error and maximizing the 3D rendering frame rate, a multi-objective optimization function is constructed, and the structural constraint parameters and the optimization objective function are used as prior inputs for the subsequent rendering process.

3. The method according to claim 1, characterized in that, Step S2 specifically includes: Based on the equipment topology connection relationship and electrical boundary conditions in the structural constraint parameters, historical telemetry data, infrared thermal imaging video, partial discharge ultrasonic signals and equipment ledger point cloud data of power distribution network equipment are collected to construct the original multi-source dataset. Based on the modeling accuracy and computational efficiency indices in the optimization objective function, timestamp alignment and time series interpolation are performed on the original multi-source dataset, and wavelet transform is used to remove electromagnetic noise interference. Spatial registration of time-series telemetry data, thermal imaging textures, and ultrasonic spectral features is performed and mapped to a unified coordinate system. Then, feature-level fusion is used to generate multimodal data that includes electrical, thermal, and mechanical states.

4. The method according to claim 1, characterized in that, Step S3 specifically includes: The position and scale of each Gaussian ellipsoid are initialized based on the point cloud information in the multimodal data, and the color and opacity are initialized based on the optical properties of the device material. Calculate the sensitivity index of each device region under the optimization objective function, where the sensitivity weight of the high thermal sensitivity region is higher than that of the low thermal sensitivity region; Based on the structural constraint parameters, the relative displacement range between adjacent Gaussians is limited, and the initial opacity and rendering priority of each Gaussian are adaptively adjusted in combination with the sensitivity index to generate an initial 3DGS scene with physical interpretability.

5. The method according to claim 4, characterized in that, The calculation of the sensitivity index of each device region under the optimization objective function specifically includes: Based on the partial derivatives of the temperature field prediction error minimization term in the objective function with respect to the input characteristics of each equipment region, the thermal sensitivity coefficient of each equipment region is calculated. By combining the thermal diffusivity and current density distribution of the equipment materials, the thermal sensitivity coefficient is normalized to obtain a sensitivity index with a value range of zero to one. The sensitivity index is mapped to a scaling factor of the initial opacity. For device regions where the sensitivity index is greater than a preset threshold, the initial opacity of the corresponding Gaussian value is increased to 0.2 to 0.5 times the base value, and higher rendering computing resources are allocated to them.

6. The method according to claim 1, characterized in that, Step S4 specifically includes: The real-time collected three-phase current, voltage, and ambient temperature and humidity parameters are input into a fully connected network and encoded into high-dimensional conditional feature vectors. The association weights between the conditional feature vector and each Gaussian attribute are calculated using a self-attention mechanism, and the output value of the optimization objective function corresponding to the current running state is used as a scaling factor. The associated weights are multiplied element-wise by the scaling factor to generate dynamic modulation coefficients. These modulation coefficients are then injected into the 3DGS rendering pipeline as scaling bias terms for Gaussian attribute updates. This allows for real-time adjustment of the shape, density, and color distribution of the Gaussian attribute, enabling dynamic responses of the Gaussian attribute to overload, short circuit, or aging conditions.

7. The method according to claim 6, characterized in that, The calculation of the association weights between the conditional feature vector and each Gaussian attribute using a self-attention mechanism specifically includes: The position, scale, and material properties of each Gaussian are encoded into a sequence of Gaussian attribute feature vectors. Using the high-dimensional conditional feature vector as the query vector and the Gaussian attribute feature vector sequence as the key vector and value vector, multi-head self-attention operation is performed; Based on the calculation results, a correlation weight matrix is ​​generated to characterize the degree of influence of electrical state and ambient temperature on each Gaussian, where Gaussians corresponding to overload regions or hot spots are given higher correlation weights.

8. The method according to claim 1, characterized in that, Step S5 specifically includes: Using multi-view image sequences and infrared thermal imaging sequences as supervision signals, pixel-level photometric loss and thermal consistency loss are constructed. The heat balance equation and electric field intensity distribution output by the aforementioned mechanism model are used as physical regularization terms; By jointly optimizing the Gaussian geometric parameters of 3DGS and the weight parameters of the modulation network composed of the fully connected network and self-attention mechanism described in step S4, the digital twin model converges synchronously in both visual realism and physical consistency, thus obtaining a state-aware 3DGS digital twin model of distribution network equipment.

9. The method according to claim 8, characterized in that, The use of the heat balance equation and electric field intensity distribution output by the aforementioned mechanism model as physical regularization terms specifically includes: Based on the aforementioned thermal balance equation, a temperature field canonical loss is constructed to constrain the deviation between the predicted temperature of each Gaussian node and the temperature calculated by the mechanism model. Based on the Poisson equation for electrostatic field, a canonical loss for electric field intensity is constructed to constrain the gradient change of the Gaussian distribution on the surface of the charged body to conform to the distribution law of the equipotential surface of the electric field. The temperature field regularization loss and the electric field intensity regularization loss are weighted and summed to form a total physical regularization term, in which overheated regions and high field intensity regions are given greater penalty weights to suppress hot spots and field intensity distortions that violate physical common sense in the rendering results.

10. A digital twin modeling system for power distribution network equipment based on multi-source data fusion and 3DGS neural rendering, applied to the method described in any one of claims 1 to 9, characterized in that, The system includes: The mechanism model construction module is used to build state machine-driven mechanism models and output structural constraint parameters and optimization objective functions; The multi-source data acquisition and preprocessing module is used to acquire multi-source data based on the structural constraint parameters and optimization objective function, and generate multimodal data through preprocessing. The 3DGS scene initialization module has its input end connected to the output end of the mechanism model construction module and the output end of the multi-source data acquisition and preprocessing module. It is used to initialize the 3DGS scene representation using the multimodal data and adaptively configure the initial rendering attributes according to the structural constraint parameters and sensitivity index. The conditional neural modulation module, whose input is connected to the output of the mechanism model construction module, is used to encode real-time running state parameters into conditional features, and inject them into the 3DGS rendering pipeline with the optimization objective function corresponding to the current state as a prior constraint to achieve the dynamic response of Gaussian properties to the running state. The joint optimization training module has its input terminals connected to the output terminals of the 3DGS scene initialization module, the conditional neural modulation module, and the mechanism model construction module, respectively. It is used to optimize the Gaussian parameters and modulation network end-to-end using the image sequence as the supervision signal and the physical constraints of the mechanism model as the regularization term, so as to obtain a state-aware 3DGS digital twin model.

Images