Construction method, device and equipment of digital twinborn model

Abstract

According to the construction method, device and equipment of the digital twinborn model provided by the embodiment of the invention, each edge computing node carries out model training according to the initialization parameters sent by the cloud computing platform and by utilizing the locally pre-stored multi-source data fusion model, so that the trained model parameters are obtained and sent to the cloud computing platform; and then, according to global model parameters obtained through weighted average calculation of the cloud computing platform, a pre-stored multi-source data fusion model and equipment operation data obtained according to a preset frequency, generating a standard fusion data vector, and finally, according to the standard fusion data vector, constructing an equipment-level digital twin model. Through the mode, real-time monitoring and prediction of the operation state of the edge equipment are realized, and the operation reliability of the equipment is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information technology, and in particular to a method and device for constructing a digital twin model. BACKGROUND

[0002] With the accelerating evolution of distribution networks towards a "source-grid-load-storage" collaborative mode, the operation state of distribution systems is becoming more complex, the scale of multi-source data is continuously expanding, and the requirements for real-time and accuracy of equipment health management and fault diagnosis are significantly increasing. Traditional monitoring methods relying on centralized collection and single-point modeling cannot meet the real-time, high-reliability, and privacy protection requirements. Digital twin technology, as a core means for building a "virtual-real synchronous" power system, can realize equipment state perception, operation risk prediction, and optimization decision-making through real-time data-driven virtual models. Therefore, there is an urgent need for a digital twin construction method for cloud-edge collaborative architecture to achieve multi-source data hierarchical collection, federated learning privacy fusion, device-level and system-level model collaboration, and online dynamic updating and closed-loop fault diagnosis to meet the real-time, accurate, and secure intelligent development needs of future distribution systems.

[0003] In the prior art, the digital monitoring of distribution systems usually relies on a cloud centralized architecture, that is, various types of operation data are directly uploaded to the cloud platform through front-end collection devices, and the cloud platform performs data cleaning, fusion, modeling, and fault diagnosis. This type of solution generally uses traditional big data fusion methods or single machine learning models to build system state prediction and equipment health assessment, and updates the twin model in a periodic offline modeling manner.

[0004] However, the existing solution has poor data real-time performance and high cloud load, making it difficult to meet the real-time and high-precision decision-making needs of distribution networks. SUMMARY

[0005] The embodiments of the present application provide a method and device for constructing a digital twin model to solve the problem of poor data real-time performance and high cloud load in the prior art, which makes it difficult to meet the real-time and high-precision decision-making needs of distribution networks.

[0006] In a first aspect, the embodiments of the present application provide a method for constructing a digital twin model, applied to a system comprising:

[0007] sending equipment operation data to the cloud computing platform at a preset frequency;

[0008] receiving initialization parameters sent by the cloud computing platform;

[0009] According to the sample equipment operation data, the label vector, and the initialization parameter, a pre-stored multi-source data fusion model is trained to obtain trained model parameters, and the trained model parameters are sent to the cloud computing platform. The label vector is a sample data fusion vector corresponding to the sample operation data.

[0010] The global model parameters sent by the cloud computing platform are received, and a standard fusion data vector is obtained according to the equipment operation data, the global model parameters, and the pre-stored multi-source data fusion model.

[0011] A device-level digital twin model is constructed according to the standard fusion data vector.

[0012] In a possible implementation, the training of the pre-stored multi-source data fusion model according to the sample equipment operation data, the label vector, and the initialization parameter to obtain the trained model parameters comprises:

[0013] The sample equipment operation data is pre-processed to obtain standardized data. The pre-processing at least includes denoising, missing value completion, and data standardization.

[0014] The sample equipment operation data and the label vector are input into a multi-source data fusion model based on the initialization parameter, and the multi-source data fusion model based on the initialization parameter is iteratively trained with the minimization of the mean square error of the model as the training target.

[0015] When the multi-source data fusion model based on the initialization parameter meets a preset condition, the trained model parameters are obtained. The preset condition includes a maximum number of iterations and / or a minimum mean square error.

[0016] In a possible implementation, the obtaining of the standard fusion data vector according to the equipment operation data, the global model parameters, and the pre-stored multi-source data fusion model comprises:

[0017] The global model parameters are deployed to the pre-stored multi-source data fusion model to obtain a multi-source data fusion global model.

[0018] The pre-processing operation is performed on the equipment operation data to obtain standard equipment operation data.

[0019] The standard equipment operation data are input into the multi-source data fusion global model to obtain the standard fusion data vector.

[0020] In a possible implementation, the construction of the device-level digital twin model according to the standard fusion data vector comprises:

[0021] The device physical attribute is associated with the device operation data, a device real-time state is generated to construct a physical layer;

[0022] The standard fusion data vector is input into a pre-constructed long short-term memory neural network-based device mechanism model, a virtual device state is output to construct a virtual layer, and the long short-term memory neural network-based device mechanism model is trained according to historical standard fusion data vectors and the device physical attribute;

[0023] A hash table fast matching algorithm is used to establish a mapping relationship between the device real-time state and the virtual device state to construct a mapping layer;

[0024] According to the physical layer, the virtual layer and the mapping layer, a device-level digital twin model is constructed.

[0025] In a second aspect, the embodiments of the present application provide a method for constructing a digital twin model, applied to a cloud computing platform, comprising:

[0026] Receiving device operation data sent by each edge computing node;

[0027] Initializing a pre-stored multi-source data fusion model and sending initialization parameters to each edge computing node;

[0028] Receiving trained model parameters sent by each edge computing node and performing weighted processing on each trained model parameter to obtain global model parameters;

[0029] Sending global model parameters to each edge computing node;

[0030] According to the device operation data and macro data, a system-level digital twin model is constructed.

[0031] In a possible implementation, the weighted processing of each trained model parameter to obtain global model parameters comprises:

[0032] According to a pre-set data quality evaluation index, each edge computing node is assigned a corresponding aggregation weight, wherein the data quality evaluation index is determined based on one or more of node type, data integrity and historical data accuracy;

[0033] Based on each aggregated weight, the weighted average calculation is performed on each trained model parameter to obtain global model parameters.

[0034] In a possible implementation, the macro data at least includes regional power grid topology data;

[0035] The system-level digital twin model is constructed according to the device operation data and the macro data, comprising:

[0036] Based on the regional power grid topology data in the macro data, the regional network topology is represented by an adjacency matrix to construct a topology layer;

[0037] Based on the device operation data and the regional network topology, power flow simulation and transient simulation under fault state are performed to obtain simulation results to construct a simulation layer; the power flow simulation calculates electrical parameters of each node by using Newton-Raphson method, and the transient simulation simulates the transient process by using a graph neural network trained based on historical fault data;

[0038] Based on the simulation results, multi-objective genetic algorithm is used to perform optimization calculation on the operation parameters of the power distribution system to generate an optimization strategy including reactive power compensation, load distribution and backup power source input timing to construct an optimization layer;

[0039] According to the topology layer, the simulation layer and the optimization layer, a system-level digital twin model is constructed.

[0040] In a third aspect, an embodiment of the present application provides a digital twin model construction device, applied to each edge computing node, comprising:

[0041] A sending module is configured to send device operation data to the cloud computing platform at a preset frequency.

[0042] A receiving module is configured to receive initialization parameters sent by the cloud computing platform.

[0043] A processing module is configured to train a pre-stored multi-source data fusion model according to sample device operation data, a label vector and the initialization parameters, obtain model parameters after training, and send the model parameters after training to the cloud computing platform, wherein the label vector is a sample data fusion vector corresponding to the sample operation data.

[0044] The processing module is further configured to receive global model parameters sent by the cloud computing platform, and obtain a standard fusion data vector according to the device operation data, the global model parameters and the pre-stored multi-source data fusion model.

[0045] A construction module is configured to construct a device-level digital twin model according to the standard fusion data vector.

[0046] In a possible implementation, the processing module is specifically configured to:

[0047] Preprocess sample device operation data to obtain standardized data, wherein the preprocessing at least includes denoising, missing value completion and data standardization.

[0048] inputting the sample equipment operation data and the label vector into the multi-source data fusion model based on the initialization parameters, performing iterative training on the multi-source data fusion model based on the initialization parameters, and taking minimizing mean square error of the model as a training target;

[0049] when the multi-source data fusion model based on the initialization parameters reaches a preset condition, obtaining trained model parameters, and the preset condition includes a maximum number of iterations and / or a minimum mean square error.

[0050] In a possible implementation, the processing module is further configured to:

[0051] deploying the global model parameters to the pre-stored multi-source data fusion model to obtain a multi-source data fusion global model;

[0052] performing the preprocessing operation on the equipment operation data to obtain standard equipment operation data;

[0053] inputting the standard equipment operation data into the multi-source data fusion global model to obtain a standard fusion data vector.

[0054] In a possible implementation, the construction module is specifically configured to:

[0055] associating the equipment physical attributes with the equipment operation data to generate an equipment real-time state to construct a physical layer;

[0056] inputting the standard fusion data vector into a pre-constructed equipment mechanism model based on a long short-term memory neural network to output a virtual equipment state to construct a virtual layer, and the equipment mechanism model based on the long short-term memory neural network is trained according to historical standard fusion data vectors and the equipment physical attributes;

[0057] establishing a mapping relationship between the equipment real-time state and the virtual equipment state by using a hash table fast matching algorithm to construct a mapping layer;

[0058] constructing an equipment-level digital twin model according to the physical layer, the virtual layer, and the mapping layer.

[0059] In a fourth aspect, an embodiment of the present application provides a digital twin model construction apparatus, applied to a cloud computing platform, and including:

[0060] a receiving module configured to receive equipment operation data sent by each edge computing node;

[0061] a sending module configured to send macro data to each edge computing node;

[0062] The sending module is further configured to initialize the pre-stored multi-source data fusion model, and send initialization parameters to each edge computing node;

[0063] The processing module is configured to receive the trained model parameters sent by each edge computing node, and perform weighted processing on each trained model parameter to obtain global model parameters;

[0064] The sending module is further configured to send the global model parameters to each edge computing node;

[0065] The constructing module is configured to construct a system-level digital twin model according to the device operation data and macro data.

[0066] In a possible implementation, the processing module is specifically configured to:

[0067] According to a pre-set data quality evaluation index, each edge computing node is assigned a corresponding aggregation weight, wherein the data quality evaluation index is determined based on one or more of node type, data integrity, and historical data accuracy;

[0068] Based on each aggregated weight, weighted average calculation is performed on each trained model parameter to obtain global model parameters.

[0069] In a possible implementation, the macro data at least includes regional power grid topology data;

[0070] The constructing module is specifically configured to:

[0071] Based on the regional power grid topology data in the macro data, an adjacency matrix is used to represent the regional network topology to construct a topology layer;

[0072] Based on the device operation data and the regional network topology structure, power flow simulation and transient simulation under fault state are performed to obtain simulation results, to construct a simulation layer; the power flow simulation is to calculate electrical parameters of each node by using Newton-Raphson method, and the transient simulation is to simulate a transient process by using a graph neural network trained based on historical fault data;

[0073] Based on the simulation results, multi-objective genetic algorithm is used to perform optimization calculation on operation parameters of the power distribution system to generate an optimization strategy including reactive power compensation, load distribution, and standby power source input timing, to construct an optimization layer;

[0074] According to the topology layer, the simulation layer, and the optimization layer, a system-level digital twin model is constructed.

[0075] In a fifth aspect, an electronic device is provided, including a memory and a processor.

[0076] The memory stores computer-executable instructions;

[0077] The processor executes the computer-executable instructions stored in the memory, so that the processor executes the first aspect and / or various possible implementation manners of the first aspect.

[0078] The method, device and equipment for constructing a digital twin model provided by the embodiments of the present application, each edge computing node first receives the initialization parameters of the multi-source data fusion model issued by the cloud computing platform, and combines the local pre-stored historical equipment operation data, the label vector and the multi-source data fusion model to perform local model training, thereby obtaining the optimized local model parameters, and uploading these parameters to the cloud computing platform; the cloud platform performs weighted averaging on the model parameters from each edge node according to the data quality or node weight, generates global model parameters, realizes unified fusion and privacy protection of different source data; subsequently, each edge node utilizes the global model parameters to combine the real-time collected equipment operation data to generate standardized fusion data vectors, providing high-quality and unified input for subsequent modeling; finally, based on these standard fusion data vectors, the edge node constructs a device-level digital twin model, realizes virtual-real mapping through physical layer modeling, mechanism modeling and data-driven correction, ensures that the virtual model can quickly and accurately reflect the operation state of the equipment, realizes real-time monitoring and prediction of the operation state of the equipment, thereby improving the operation reliability of the equipment, supporting fault early warning, and providing a high-precision data basis for system-level optimization and control. BRIEF DESCRIPTION OF DRAWINGS

[0079] The accompanying drawings, which are incorporated in and constitute a part of the specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the application.

[0080] Figure 1 The structure diagram of the power distribution system provided by the embodiments of the present application Figure 1

[0081] Figure 2 The interface diagram of the method for constructing a digital twin model provided by the embodiments of the present application

[0082] Figure 3 The structure diagram of the device for constructing a digital twin model provided by the embodiments of the present application Figure 1

[0083] Figure 4 The structure diagram of the device for constructing a digital twin model provided by the embodiments of the present application Figure 2

[0084] Figure 5 The structure diagram of the electronic device provided by the embodiments of the present application. ​​​

[0085] The present application has been shown and described with reference to the preferred embodiments. Equivalent mechanisms and methods incorporating only functionally similar to those described herein are within the scope of the application. Therefore, the application is not limited except as indicated in the appended claims. DETAILED DESCRIPTION

[0086] The exemplary embodiments will be described in detail with reference to only the drawings. The following description, which is not intended to be limiting, is provided in order to explain the concepts of the application by reference to particular embodiments. The same numbers in different drawings designate the same or similar elements.

[0087] As the power distribution network accelerates towards the "source-grid-load-storage" collaborative mode, the operation state of the power distribution system becomes more complex, the scale of multi-source data continues to expand, and the real-time and accuracy requirements of equipment health management and fault diagnosis are significantly improved. The traditional monitoring method relying on centralized collection and single-point modeling cannot meet the real-time, high-reliability, and privacy protection requirements. Digital twin technology, as a core means to build a "virtual-real synchronous" power system, can realize equipment state perception, operation risk prediction, and optimization decision-making through real-time data-driven virtual models. Therefore, a digital twin construction method for cloud-edge collaborative architecture is urgently needed to realize multi-source data hierarchical collection, federated learning privacy fusion, device-level and system-level model collaboration, and online dynamic updating and closed-loop fault diagnosis to meet the real-time, accurate, and safe intelligent development needs of future power distribution systems.

[0088] In the prior art, the digital monitoring of the power distribution system usually relies on a cloud centralized architecture, that is, various types of operation data are directly uploaded to the cloud platform through the front-end collection device, and the cloud completes data cleaning, fusion, modeling, and fault diagnosis tasks. This type of solution generally uses traditional big data fusion methods or single machine learning models to build system state prediction and equipment health assessment, and updates the twin model in a periodic offline modeling manner.

[0089] However, the existing solution can achieve basic operation monitoring in a small-scale scenario, but due to the lack of edge-side collaboration capabilities, the data real-time performance is poor, the cloud load is high, and cross-regional data fusion is easily restricted by privacy and heterogeneity, making it difficult to meet the real-time, dynamic simulation, and high-precision decision-making needs of modern power distribution networks.

[0090] Based on this, this application proposes a method for constructing a digital twin model. The inventors, using the principles of cloud-edge collaboration and federated learning, propose a method for constructing a digital twin model. This involves collecting device operation data at the edge and sending it to the cloud at a preset frequency, while simultaneously receiving model initialization parameters from the cloud. At the edge, the pre-stored multi-source data fusion model is trained locally using sample device operation data, label vectors, and initialization parameters. The trained model parameters are then generated and uploaded to the cloud. The cloud aggregates the parameters from each edge node to form global model parameters, which are then distributed to the edge to generate standard fusion data vectors. Finally, a device-level digital twin model is constructed at the edge. This method achieves collaborative data processing and model training between the edge and the cloud, significantly improving data transmission real-time performance and fusion accuracy while reducing cloud computing pressure. It also enables secure fusion of cross-regional, multi-source data and provides high-precision device status simulation and decision support under dynamic operating conditions.

[0091] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0092] Figure 1 A schematic diagram of the power distribution system provided in the embodiments of this application. Figure 1 ;like Figure 1 As shown, the system includes: a physical device layer 101, an edge device layer 102, a cloud hub layer 103, a communication interconnection layer 104, and an application service layer 105. This power distribution system is used in conjunction with the construction method of a digital twin model.

[0093] Among them, the physical equipment layer 101, as the real-world foundation of the digital twin, is the source of system data acquisition and corresponds to the actual hardware of the power distribution system. Its core components are divided into two categories: core power distribution equipment and multi-dimensional sensing equipment. The core power distribution equipment includes: 10kV-220kV transformers, high-voltage switches (circuit breakers / disconnect switches), feeder cables, distributed photovoltaic / energy storage equipment, and charging piles. Its function is to realize power transmission, distribution, and terminal access, and it is the physical object of the "virtual-real mapping" of the twin model.

[0094] The multi-dimensional sensing devices include: current / voltage sensors (monitoring electrical quantities), temperature / vibration sensors (monitoring equipment status), smart meters (monitoring user load), and ambient temperature and humidity sensors; their function is to collect real-time operating data of the equipment (such as transformer oil temperature and line current) and provide raw data input for edge computing nodes / cloud computing platforms.

[0095] The edge processing layer 102, as a "real-time response center" of digital twinning, is deployed in an edge computing node (substation / switching station) and undertakes local rapid processing and near real-time decision-making. The core components include two categories: edge hardware and processing modules and local decision-making units. The edge hardware and processing modules include: edge gateway (connecting sensing devices and cloud) and edge computing unit (small server). The functions include: pre-processing data (filtering noise, filling missing values); running a lightweight twinning model (such as a simplified state model of a device) to realize real-time mapping (delay ≤ 100 ms) of physical devices to virtual models; the functions of the local decision-making unit: rapid early warning (such as local alarm when the current exceeds the threshold); simple fault handling (such as automatic shutdown of the fault line when single-phase grounding occurs), avoiding the time delay caused by relying on the cloud,

[0096] The cloud center layer 103, as a global optimization brain of digital twinning, is deployed in a cloud computing platform and undertakes high-precision modeling and global management. The core components include two categories: global twinning model center and data management and decision-making unit. The global twinning model center includes: device full-life cycle model (integrating device historical operation / maintenance data to predict remaining life); system-level topology and power flow model (covering regional power distribution network, simulating full-network voltage and network loss); the function is to realize high-fidelity virtual restoration of physical systems, support complex simulation (such as fault transient analysis) and long-term optimization; the data management and decision-making unit includes: distributed database (storing pre-processed data and historical data uploaded by the edge) and global optimization algorithm module; the function is to fuse multi-source data (edge real-time data and historical data of power distribution management system) and make global decisions (such as cross-regional load scheduling and complex fault impact range analysis).

[0097] The communication interconnection layer 104, as a data transmission link of digital twinning, connects the physical device layer, the edge processing layer and the cloud center layer. The core components include two categories: hierarchical communication protocol and security protection unit. The hierarchical communication protocol includes edge local (industrial Ethernet / Wi-Fi, connecting sensing devices and edge gateway) and cloud-edge interconnection (5G / HTTP, transmitting abnormal data from edge to cloud and optimization instructions from cloud to edge); the function is to ensure real-time data transmission and less redundancy (such as transmitting abnormal data to cloud within 50 ms and transmitting normal data in batches); the security protection unit includes data encryption (transmission / storage encryption) and device identity authentication (to prevent fake devices from accessing); the function is to ensure the safety of power distribution data (such as user load) and control instructions (such as switch opening).

[0098] The application service layer 105, as a value landing end of digital twinning, focuses on the core needs of power distribution operation and three core scenarios: real-time monitoring visualization: based on a cloud model, device status is displayed in a three-dimensional interface (such as red indicating a fault device and green indicating a normal device), and remote viewing is supported; fault diagnosis and self-healing: edge early warning to cloud analysis of fault type / impact range and then automatic issuance of a processing scheme (such as cutting off a fault line and starting a backup power supply); and load optimization scheduling: cloud prediction of future 24-hour load, optimization of photovoltaic / energy storage operation (such as using energy storage to supply power during a load peak), and reduction of network loss.

[0099] Figure 2 The interface of the method for constructing a digital twinning model provided in the embodiments of the present application Figure 1 ; as Figure 2 shown, the method comprises:

[0100] S201, each edge computing node sends device operation data to a cloud computing platform at a preset frequency.

[0101] It should be understood that edge computing nodes are deployed at key nodes such as substations, feeder ends, distribution switches and user sides, and the edge computing nodes (edge side) are connected to multiple types of sensors (current / voltage sensors with a precision of 0.2 level, temperature sensors with a precision of ±0.5℃, vibration sensors, etc.). The edge computing nodes perform local preprocessing (such as denoising, missing value completion and standardization, which are implemented by preprocessing of S205) on the collected raw data, and send the data to a cloud computing platform (cloud side) at a preset frequency and content classification. Specifically, device status data (current, voltage, oil temperature and vibration frequency) is sampled at 5-10 Hz and summarized or reported in real time in case of abnormality at a minute / second level; load and environmental data is sampled at 1-2 Hz or 1 Hz, and abnormal data is uploaded in real time and normal data is sent in batches according to a transmission strategy. 5G-Uu or a low-delay link is used for transmission of abnormal / high-priority data, and the target end-to-end delay is ≤50 ms; HTTP / 2 is used for batch reporting of regular batch data (which can be sent after being cached for 1 hour).

[0102] It can be understood that in this way, high-frequency key data on the edge side can be transmitted to the cloud side with low delay and high reliability, the real-time monitoring capability is improved, unnecessary network resource occupation is avoided, and the load pressure of the cloud side is relieved from the source.

[0103] It should also be noted that the cloud computing platform collects required data through a power distribution management system (PMS), an energy management system (EMS) and a third-party platform, and specifically:

[0104] Historical data: 5-year operation records of the collection device, 3-year fault data (collection frequency 1 time / day); macro data: regional power grid topology (updated 1 time / month), 24-hour weather forecast (updated 1 time / hour); and planning data: power grid expansion scheme, new energy access plan (updated 1 time / quarter). After collecting these data on the cloud computing platform, the macro data required by the edge side (such as topology change information) will be sent to the edge side as needed to avoid data redundancy.

[0105] S202, the cloud computing platform receives the device operation data sent by each edge computing node.

[0106] It should be understood that after the cloud receives the device operation data sent by each edge computing node, the data also needs to be warehoused, indexed and quality evaluated. The receiving process includes data integrity check, time stamp synchronization (using NTP / PTP time setting), and labeling node type, signal-to-noise ratio and missing ratio for each data for subsequent data quality evaluation indicators. The cloud sends the original / preprocessed summary to different modules: federal learning coordinator, system-level simulation module, historical database. Data reception follows an encrypted channel (SM2) and an authentication mechanism.

[0107] It can be understood that by establishing a high-quality data receiving and management link in the cloud, a reliable and traceable data foundation is provided for subsequent model initialization, parameter aggregation and system-level simulation, reducing the simulation / decision deviation caused by errors or time sequence disorder.

[0108] S203, the cloud computing platform initializes the pre-stored multi-source data fusion model and sends the initialization parameters to each edge computing node.

[0109] It should be understood that the multi-source data fusion model pre-stored in the cloud has an input layer supporting 128 features (device status, environment, load, etc.), 3 hidden layers, and an output of a normalized fusion feature vector (64 dimensions later). At the beginning of the federal learning round, the cloud initializes the model parameters according to the global strategy (such as Xavier / He initialization or preheating parameters based on historical training), and after encrypting (SM2) the initialization parameters, it distributes them to all online edge nodes through a secure channel. The cloud also issues training hyperparameters (learning rate 0.001, local iteration round 10, batch size, etc.) and aggregation strategies (data quality weight criteria).

[0110] It can be understood that by uniform initialization, the edge nodes are locally trained under the same model architecture and initial conditions, facilitating subsequent parameter aggregation and convergence control, while supporting privacy-protected parameter distribution.

[0111] S204, each edge computing node receives the initialization parameters sent by the cloud computing platform.

[0112] It should be understood that after receiving the cloud initialization parameters, each edge computing node performs integrity and signature verification, loads the pre-stored multi-source data fusion model, and prepares a local training environment (for example, an edge running environment using TensorFlow Lite / TF Federated). The receiving triggers a local preprocessing pipeline, so that the edge can start training or inference based on the latest data immediately after receiving the parameters.

[0113] It can be understood that through the unified distribution and training of the initialization parameters, the edge model is consistent with the cloud initialization parameters, the model deviation caused by asynchronous parameters is reduced, and the efficiency and security of federated learning are improved.

[0114] S205, each edge computing node trains the pre-stored multi-source data fusion model according to the sample device running data, the label vector and the initialization parameters, obtains the trained model parameters, and sends the trained model parameters to the cloud computing platform.

[0115] The label vector is a sample data fusion vector corresponding to the sample running data.

[0116] In an implementable manner, first, a preprocessing operation is performed on the sample device running data to obtain standardized data, and the preprocessing operation at least includes denoising, missing value completion and data standardization; then, the sample device running data and the label vector are input into the multi-source data fusion model based on the initialization parameters, and the multi-source data fusion model based on the initialization parameters is iteratively trained with the minimum mean square error of the model as the training target; finally, when the multi-source data fusion model based on the initialization parameters meets the preset condition, the trained model parameters are obtained, and the preset condition includes the maximum number of iterations and / or the minimum mean square error.

[0117] It should be understood that the edge computing node first performs a preprocessing operation on the sample device running data, specifically: for current or voltage data, a Kalman filter and wavelet transform combined algorithm is used, first smoothing high-frequency fluctuations through Kalman filter, and then removing white noise through db4 wavelet decomposition, with a target denoising rate of ≥90%; for temperature or vibration data: a moving average filter (window size 5 sampling points) is used to filter out transient interference; then missing value completion processing is performed, for short-term missing data (i.e. data missing for no more than 5 sampling points): linear interpolation method is used, with a completion time of ≤100ms; for long-term missing data (data missing for more than 5 sampling points): cubic spline interpolation method is used, combined with historical data correction, with a target completion error of ≤3%; finally, standardization processing is performed, i.e. all device running data is mapped to [0,1], wherein the minimum / maximum value of each type of data is based on the rated parameters and safety margin of the device (such as the maximum current being 1.2 times the rated current), to avoid dimensional influence.

[0118] The specific training process is: the preprocessed sample data and the corresponding label vector (i.e. the "data fusion vector" of the sample) are input into the local multi-source data fusion model, and the objective is to minimize the mean square error MSE. Training is performed using a local iteration of 10 rounds, a learning rate of 0.001, and the model parameters are sparsified and quantized during the training process to compress to ≤5MB. After reaching the preset condition (maximum number of iterations or MSE≤0.05), the edge computing node transmits the trained parameters back to the cloud after encryption. Among them, the federated learning framework uses TensorFlow Federated to coordinate the training process and ensures that only parameters are transmitted, not raw data.

[0119] It can be understood that sufficient preprocessing and local training at the edge not only improve the fitting ability of the local model to the local regional characteristics, but also protect data privacy by only uploading parameters, significantly reducing cross-domain data transmission volume and cloud preprocessing pressure, while ensuring that the fusion error is controlled within the design threshold.

[0120] S206, the cloud computing platform receives the trained model parameters sent by each edge computing node, and performs weighted processing on each trained model parameter to obtain global model parameters.

[0121] In an implementable manner, first, according to a pre-set data quality evaluation index, each edge computing node is assigned a corresponding aggregation weight, and then based on each aggregated weight, a weighted average calculation is performed on each trained model parameter to obtain global model parameters.

[0122] The data quality evaluation index is determined based on one or more of node type, data integrity, and historical data accuracy.

[0123] It should be understood that after receiving the trained model parameters from each edge computing node, the cloud computing platform first needs to verify the completeness and reliability of the parameters. Then, based on preset data quality evaluation indicators, specifically, it first sets basic weights according to node type (e.g., substation, distribution box, and user side). For example, due to its comprehensive monitoring equipment and high data quality, the basic weight of a substation node is set to 0.8, that of a distribution box node to 0.5, and that of a user side node to 0.2. Subsequently, considering the data completeness of each node, the missing rate and noise rate are mapped to completeness scores, and the basic weights are adjusted accordingly. For example, if a substation node has an extremely low missing rate and a low noise rate, its completeness score is high, thus maintaining its final weight at a high level. Finally, the weights are further fine-tuned based on the accuracy of each node's historical data (e.g., historical model error performance). For example, if a user side node has a consistently large historical prediction error, its final aggregated weight is reduced from the original. Through the above methods, a comprehensive weight reflecting node type, data collection quality, and historical reliability can be obtained, which is used for the weighted aggregation of global model parameters, making the aggregation result more stable, reliable, and possessing higher generalization ability. Furthermore, a weighted average method is used to fuse the parameters to obtain a global model.

[0124] It is understandable that weighted aggregation based on data quality can improve the robustness of the global model, ensure that high-quality data nodes contribute more to the global model, improve the performance of the aggregated model at key nodes, and reduce the adverse effects of abnormal nodes on the global model, thereby improving the overall fusion accuracy and system reliability.

[0125] S207. The cloud computing platform sends global model parameters to each edge computing node.

[0126] It should be understood that the cloud computing platform distributes the aggregated global model parameters to all edge computing nodes, still using encrypted transmission (SM2) and including version numbers and expiration information. Simultaneously, it should be noted that the cloud also distributes new training / inference strategies, thresholds, and monitoring metrics (such as MSE, latency requirements, and triggering conditions). Distribution can employ gRPC (low latency, high concurrency) or a reliable distribution mechanism based on message queues, supporting breakpoint resumption and retry strategies (3 retries, 10-second intervals).

[0127] Understandably, this approach ensures that edge devices can promptly acquire and deploy globally optimized model parameters, maintaining consistency between local inference results and the global perspective, thereby improving cross-regional consistency and model generalization capabilities.

[0128] S208. Each edge computing node receives the global model parameters sent by the cloud computing platform and obtains a standard fused data vector based on the device operation data, the global model parameters, and the pre-stored multi-source data fusion model.

[0129] In one possible approach, the global model parameters are first deployed to a pre-stored multi-source data fusion model to obtain a multi-source data fusion global model; then, preprocessing operations are performed on the device operation data to obtain standard device operation data; finally, the standard device operation data is input into the multi-source data fusion global model to obtain a standard fusion data vector.

[0130] It should be understood that edge computing nodes deploy global model parameters to a locally pre-stored multi-source data fusion model, using an established local preprocessing pipeline to denoise, complete, and standardize the latest device operation data to obtain standard device operation data. This standard device data is then input into the deployed multi-source data fusion global model, outputting a standard fusion data vector (64-dimensional). This vector contains device status characteristics, environmental characteristics, and load characteristics. The edge computing node performs a confidence assessment on the output vector (based on model output variance or historical error) and uses the fusion vector as input to a local twin model or reports it to the cloud computing platform for system-level simulation.

[0131] Understandably, the standard fused data vectors generated by edge computing nodes ensure the uniformity and high quality of data features, enabling seamless integration between device-level and system-level models and reducing the risks of bias and privacy leaks caused by heterogeneous data fusion.

[0132] S209. The cloud computing platform constructs a system-level digital twin model based on equipment operation data and macro data.

[0133] In one feasible approach, firstly, based on regional power grid topology data from macroscopic data, an adjacency matrix is ​​used to represent the regional network topology to construct a topology layer. Then, based on equipment operation data and the regional network topology, power flow simulation and transient simulation under fault conditions are performed to obtain simulation results, thus constructing a simulation layer. Next, based on the simulation results, a multi-objective genetic algorithm is used to optimize the operating parameters of the power distribution system, generating optimization strategies that include reactive power compensation, load allocation, and the timing of backup power supply activation, thus constructing an optimization layer. Finally, based on the topology layer, simulation layer, and optimization layer, a system-level digital twin model is constructed.

[0134] Among them, macroscopic data includes at least regional power grid topology data; power flow simulation uses the Newton-Raphson method to calculate the electrical parameters of each node, and transient simulation uses a graph neural network trained based on historical fault data to simulate transient processes.

[0135] It should be understood that the cloud computing platform uses preprocessed device operation data reported from the edge to obtain a standard fused data vector, along with macroscopic data (such as regional power grid topology, historical operation records, 24-hour weather forecasts, and expansion / access plans) as input to construct a three-layer system-level twin model. Specifically, it uses an adjacency matrix to represent the regional network topology (nodes: substations, switches, user distribution boxes), receives new / deleted feeder information from the edge, updates the adjacency matrix within one minute, and synchronizes it to construct the topology layer. Then, the Newton-Raphson method is used for power flow simulation in normal scenarios, where the algorithm's convergence accuracy... The simulation step size is 1 minute, with output node voltage deviation ≤2% and line loss error ≤1%. For fault transient simulation, a graph neural network (GNN) trained based on historical fault data (≥500 cases) is used, which generates transient current / voltage curves with a step size of 0.1 s and a simulation error ≤5% to construct the simulation layer. Then, based on the simulation results, the simulation is performed using... A multi-objective genetic algorithm simultaneously optimizes network loss, reliability, and load balancing (objectives include reducing network loss by 5% and improving power supply reliability by 8%), outputting optimization strategies (such as reactive power compensation capacity configuration, feeder load redistribution, and backup power supply activation timing) to construct an optimization layer. Finally, these three layers are integrated through versioning and a message bus to build a system-level digital twin model. The cloud computing platform distributes simulation and optimization results to edge computing nodes on demand for execution, thereby achieving high-precision simulation, real-time decision-making, and executable strategy distribution for the regional power distribution network, meeting reliability and real-time performance indicators.

[0136] Understandably, by constructing a high-precision system-level twin, regional power flow and transient simulation, global optimization decision-making and strategy planning can be realized, improving the ability to predict and control large-scale, multi-node linkage events, and providing macro-constraints and strategy guidance for equipment-level twins.

[0137] S210: Each edge computing node constructs a device-level digital twin model based on the standard fused data vector.

[0138] In one feasible approach, the physical attributes of the device are first associated with the device's operational data to generate the device's real-time state and construct the physical layer. Then, a standard fused data vector is input into a pre-built device mechanism model based on a long short-term memory neural network to output a virtual device state and construct a virtual layer. Next, a hash table fast matching algorithm is used to establish a mapping relationship between the device's real-time state and the virtual device state to construct a mapping layer. Finally, based on the physical layer, the virtual layer, and the mapping layer, a device-level digital twin model is constructed.

[0139] Among them, the device mechanism model based on long short-term memory neural network is trained based on historical standard fusion data vectors and device physical attributes.

[0140] It should be understood that edge computing nodes associate equipment physical attribute libraries (model, rated capacity, number of winding turns, core loss, etc.) with real-time operating data to generate a real-time equipment status table (updated at a frequency of 5Hz) to construct the physical layer. Standard fused data vectors are input into a pre-trained equipment mechanism model based on a Long Short-Term Memory (LSTM) network (this model is trained based on historical standard fused vectors and equipment physical attributes), outputting virtual equipment status to construct the virtual layer. The goal of correcting the mechanism model error is to ensure that the deviation between virtual and measured values ​​is ≤3% (e.g., oil temperature deviation ≤2℃). A hash table fast matching algorithm is used with the equipment ID + timestamp as the key to establish a physical-virtual mapping relationship, achieving bidirectional synchronization (the virtual model is updated within 100ms of physical changes, and virtual predictions are synchronized to the physical layer early warning module). Finally, based on the physical layer, virtual layer, and mapping layer, an equipment-level digital twin model is constructed, and outputs such as equipment health scores and short-term load predictions are provided for local control or reported to the cloud (reporting frequency 1 time / minute).

[0141] Understandably, achieving high-precision device-level twins at the edge can enable near real-time device health assessment, short-term load forecasting, and local early warning decisions, reducing real-time dependence on the cloud and improving the system's self-healing and local control capabilities in network outages or high-latency scenarios.

[0142] It should also be noted that although S201–S210 cover the main data flow and model building process, the dynamic updates and fault closure implementation process of this digital twin model in subsequent use still require further explanation. Specifically:

[0143] Local dynamic updates for edge computing nodes: Edge computing nodes can use incremental SVM to perform triggered updates to the device-level model (trigger condition: oil temperature deviation). or current deviation Furthermore, incremental training is performed using only about 100 new data points, with an update time of ≤100 ms. After the update, the deviation falls back to within the threshold, thereby enabling the model to quickly adapt to sudden working conditions and avoiding frequent full retraining.

[0144] Global updates on the cloud computing platform: Scheduled updates in the cloud (from 02:00 to 04:00 every day) perform Transformer-based system-level parameter optimization based on the full amount of fused data (approximately 1e8 records) for the day; Triggered updates are initiated immediately in the event of topology changes or major failures, with update time ≤30 minutes, requiring the system-level simulation error to be ≤2% after the update, to ensure that the system-level model maintains high accuracy and adapts to seasonal / topological changes.

[0145] Fault diagnosis closed-loop: Edge computing nodes monitor in real time and issue early warning packets according to rules (oil temperature > 95℃, current > 1.2 times rated current, vibration > 50Hz, etc., response ≤ 100 ms); the cloud uses early warning packets, system simulation, and a historical fault database (≥1000 cases) as input to diagnose fault type, cause, and impact range using GNN, with a diagnosis accuracy of ≥95% and a diagnosis time of ≤10 s. The cloud issues processing strategies and records execution feedback. After the historical database is updated, the subsequent diagnosis accuracy of similar faults can be improved by 3%, thereby realizing end-to-end fault management from early detection to closed-loop handling, significantly improving the reliability and operation and maintenance efficiency of the power distribution system.

[0146] Understandably, through Figure 2 The embodiments described herein demonstrate that this application implements a multi-source data fusion and digital twin construction mechanism that integrates cloud and edge computing. This mechanism enables high-quality joint modeling of cross-regional device data without uploading original data. Specifically, edge computing nodes utilize sample data for local training and only report model parameters. Combined with a cloud-based weighted aggregation strategy based on data quality, the global model possesses both wide-area robustness and consideration of regional operational characteristics. Edge computing nodes can generate standard fused data vectors in real-time based on the global model and construct device-level digital twin models, thereby obtaining highly timely and accurate device status descriptions. The cloud computing platform constructs system-level digital twin models based on standard fused vectors and macro-topology data, achieving automated operation of power flow simulation, fault transient prediction, and multi-objective optimization decision-making. This enables the power distribution system to possess unified perception, rapid diagnosis, and intelligent control capabilities from local devices to the global network. Therefore, this application effectively solves the problems of insufficient real-time performance, high cloud load, and difficulty in cross-domain privacy sharing in existing technologies, significantly improving the accuracy and real-time performance of power distribution network operation monitoring, fault analysis, and strategy decision-making.

[0147] The digital twin model construction method provided in this application involves each edge computing node first receiving initialization parameters of a multi-source data fusion model from a cloud computing platform. These parameters are then combined with pre-stored historical device operation data, label vectors, and the multi-source data fusion model for local model training, resulting in optimized local model parameters. These parameters are then uploaded to the cloud computing platform. The cloud platform performs a weighted average of the model parameters from each edge node based on data quality or node weight to generate global model parameters, achieving unified fusion of data from different sources and privacy protection. Subsequently, each edge node uses the global model parameters combined with real-time collected device operation data to generate standardized fusion data vectors, providing high-quality, unified input for subsequent modeling. Finally, based on these standardized fusion data vectors, the edge nodes construct a device-level digital twin model. Through physical layer modeling, mechanism modeling, and data-driven correction, a virtual-real mapping is achieved, ensuring that the virtual model can quickly and accurately reflect the device's operating status. This enables real-time monitoring and prediction of the device's operating status, thereby improving device reliability, supporting fault early warning, and providing a high-precision data foundation for system-level optimization and control.

[0148] Figure 3 Schematic diagram of the structure of the digital twin model construction apparatus provided in the embodiments of this application Figure 1 ;like Figure 3 As shown, the device includes:

[0149] The sending module 301 is used to send device operation data to the cloud computing platform at a preset frequency;

[0150] The receiving module 302 is used to receive initialization parameters sent by the cloud computing platform;

[0151] The processing module 303 is used to train the pre-stored multi-source data fusion model based on the sample device operation data, label vector and initialization parameters, to obtain the trained model parameters, and send the trained model parameters to the cloud computing platform. The label vector is the sample data fusion vector corresponding to the sample operation data.

[0152] The processing module 303 is also used to receive global model parameters sent by the cloud computing platform, and obtain a standard fused data vector based on the device operation data, global model parameters and pre-stored multi-source data fusion model;

[0153] Module 304 is used to build a device-level digital twin model based on standard fused data vectors.

[0154] In one possible implementation, the processing module 303 is specifically used for:

[0155] The sample equipment operation data is preprocessed to obtain standardized data. The preprocessing includes at least denoising, missing value completion, and data standardization.

[0156] The sample device operation data and label vectors are input into the multi-source data fusion model based on initialization parameters. The training objective is to minimize the mean square error of the model. The multi-source data fusion model based on initialization parameters is trained iteratively.

[0157] When the multi-source data fusion model based on the initialization parameters reaches the preset conditions, the trained model parameters are obtained. The preset conditions include the maximum number of iterations and / or the minimum mean square error.

[0158] In one possible implementation, the processing module 303 is further configured to:

[0159] Deploy the global model parameters to the pre-stored multi-source data fusion model to obtain the multi-source data fusion global model;

[0160] Perform preprocessing operations on the equipment operation data to obtain standard equipment operation data;

[0161] Standard equipment operation data is input into the multi-source data fusion global model to obtain a standard fusion data vector.

[0162] In one possible implementation, the construction module 304 is specifically used for:

[0163] By associating the physical attributes of the equipment with the equipment's operational data, the real-time status of the equipment is generated to construct the physical layer.

[0164] The standard fused data vector is input into a pre-built device mechanism model based on a long short-term memory neural network, and the virtual device state is output to construct a virtual layer. The device mechanism model based on the long short-term memory neural network is trained based on historical standard fused data vectors and device physical attributes.

[0165] A hash table fast matching algorithm is used to establish a mapping relationship between the real-time status of the device and the status of the virtual device, so as to construct a mapping layer;

[0166] A device-level digital twin model is constructed based on the physical layer, virtual layer, and mapping layer.

[0167] Figure 4 Schematic diagram of the structure of the digital twin model construction apparatus provided in the embodiments of this application Figure 2 ;like Figure 4 As shown, the device includes:

[0168] The receiving module 401 is used to receive device operation data sent by each edge computing node;

[0169] The sending module 402 is used to send macroscopic data to each edge computing node;

[0170] The sending module 402 is also used to initialize the pre-stored multi-source data fusion model and send the initialization parameters to each edge computing node;

[0171] The processing module 403 is used to receive the trained model parameters sent by each edge computing node, and to perform weighted processing on each trained model parameter to obtain the global model parameters.

[0172] The sending module 402 is also used to send global model parameters to each edge computing node;

[0173] Module 404 is used to build a system-level digital twin model based on equipment operation data and macro data.

[0174] In one possible implementation, the processing module 403 is specifically used for:

[0175] Based on pre-defined data quality assessment indicators, each edge computing node is assigned a corresponding aggregation weight, wherein the data quality assessment indicators are determined based on at least one or more of node type, data integrity, and historical data accuracy.

[0176] Based on each aggregated weight, a weighted average is calculated for each trained model parameter to obtain the global model parameters.

[0177] In one possible implementation, the macro data includes at least regional power grid topology data;

[0178] Module 404 is used specifically for:

[0179] Based on regional power grid topology data from macro data, an adjacency matrix is ​​used to represent the regional network topology in order to construct a topology layer;

[0180] Based on equipment operation data and regional network topology, power flow simulation and transient simulation under fault conditions are performed to obtain simulation results and construct the simulation layer. Power flow simulation uses the Newton-Raphson method to calculate the electrical parameters of each node, and transient simulation uses a graph neural network trained based on historical fault data to simulate transient processes.

[0181] Based on simulation results, a multi-objective genetic algorithm is used to optimize the operating parameters of the power distribution system, generating optimization strategies that include reactive power compensation, load allocation, and the timing of backup power supply activation, in order to construct an optimization layer.

[0182] A system-level digital twin model is constructed based on the topology layer, simulation layer, and optimization layer.

[0183] The digital twin model construction apparatus provided in this application embodiment can perform the above-described... Figure 1 The methods provided in the method embodiments have similar implementation principles and technical effects, and will not be described in detail here.

[0184] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 5 As shown, the electronic device 50 provided in this embodiment includes at least one processor 501 and a memory 502. Optionally, the device 50 further includes a communication component 503. The processor 501, memory 502, and communication component 503 are connected via a bus 504.

[0185] In a specific implementation, at least one processor 501 executes computer execution instructions stored in memory 502, causing at least one processor 501 to perform the above-described method.

[0186] The specific implementation process of processor 501 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0187] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0188] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0189] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0190] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0191] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0192] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0193] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0194] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0195] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0196] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0197] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0198] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A method for constructing a digital twin model, applied to various edge computing nodes, characterized in that, include: Send device operation data to the cloud computing platform at a preset frequency; Receive initialization parameters sent by the cloud computing platform; Based on the sample device operation data, the label vector, and the initialization parameters, the pre-stored multi-source data fusion model is trained to obtain the trained model parameters, and the trained model parameters are sent to the cloud computing platform. The label vector is the sample data fusion vector corresponding to the sample operation data. The system receives global model parameters sent by the cloud computing platform and obtains a standard fused data vector based on the device operation data, the global model parameters, and the pre-stored multi-source data fusion model. Based on the standard fused data vector, a device-level digital twin model is constructed.

2. The method according to claim 1, characterized in that, The step of training the pre-stored multi-source data fusion model based on sample device operating data, label vectors, and the initialization parameters to obtain trained model parameters includes: Preprocessing operations are performed on the sample equipment operation data to obtain standardized data. The preprocessing operations include at least denoising, missing value completion, and data standardization. The sample device operation data and label vectors are input into a multi-source data fusion model based on initialization parameters. The training objective is to minimize the mean square error of the model. The multi-source data fusion model based on initialization parameters is then iteratively trained. When the multi-source data fusion model based on the initialization parameters reaches the preset conditions, the trained model parameters are obtained. The preset conditions include the maximum number of iterations and / or the minimum mean square error.

3. The method according to claim 1 or 2, characterized in that, The step of obtaining a standard fused data vector based on the device operation data, the global model parameters, and the pre-stored multi-source data fusion model includes: The global model parameters are deployed to the pre-stored multi-source data fusion model to obtain the multi-source data fusion global model. The preprocessing operation is performed on the equipment operating data to obtain standard equipment operating data; The standard equipment operation data is input into the multi-source data fusion global model to obtain a standard fusion data vector.

4. The method according to claim 1, characterized in that, The step of constructing a device-level digital twin model based on the standard fused data vector includes: By associating the physical attributes of the equipment with the equipment's operational data, the real-time status of the equipment is generated to construct the physical layer. The standard fused data vector is input into a pre-built device mechanism model based on a long short-term memory neural network, and the virtual device state is output to construct a virtual layer. The device mechanism model based on the long short-term memory neural network is trained based on historical standard fused data vectors and the physical attributes of the device. A hash table fast matching algorithm is used to establish a mapping relationship between the real-time status of the device and the status of the virtual device, so as to construct a mapping layer; A device-level digital twin model is constructed based on the physical layer, the virtual layer, and the mapping layer.

5. A method for constructing a digital twin model, applied to a cloud computing platform, characterized in that, include: Receive device operation data sent by each edge computing node; Initialize the pre-stored multi-source data fusion model and send the initialization parameters to each of the edge computing nodes; Receive the trained model parameters sent by each of the edge computing nodes, and perform weighted processing on each of the trained model parameters to obtain global model parameters; The global model parameters are sent to each of the edge computing nodes; Based on the equipment operation data and macroscopic data, a system-level digital twin model is constructed.

6. The method according to claim 5, characterized in that, The weighted processing of the trained model parameters to obtain global model parameters includes: Based on pre-defined data quality assessment indicators, each edge computing node is assigned a corresponding aggregation weight, wherein the data quality assessment indicators are determined based on at least one or more of node type, data integrity, and historical data accuracy. Based on each aggregated weight, a weighted average is calculated on the trained model parameters to obtain the global model parameters.

7. The method according to claim 5, characterized in that, The macro data includes at least regional power grid topology data; The step of constructing a system-level digital twin model based on the equipment operation data and macroscopic data includes: Based on the regional power grid topology data in the aforementioned macro data, an adjacency matrix is ​​used to represent the regional network topology in order to construct a topology layer; Based on the equipment operation data and the regional power grid topology, power flow simulation and transient simulation under fault conditions are performed to obtain simulation results and construct a simulation layer. The power flow simulation uses the Newton-Raphson method to calculate the electrical parameters of each node, and the transient simulation uses a graph neural network trained based on historical fault data to simulate the transient process. Based on simulation results, a multi-objective genetic algorithm is used to optimize the operating parameters of the power distribution system, generating optimization strategies that include reactive power compensation, load allocation, and the timing of backup power supply activation, in order to construct an optimization layer. A system-level digital twin model is constructed based on the topology layer, the simulation layer, and the optimization layer.

8. A device for constructing a digital twin model, applied to various edge computing nodes, characterized in that, include: The sending module is used to send device operation data to the cloud computing platform at a preset frequency; A receiving module is used to receive initialization parameters sent by the cloud computing platform; The processing module is used to train a pre-stored multi-source data fusion model based on sample device operation data, label vectors, and the initialization parameters, to obtain trained model parameters, and to send the trained model parameters to the cloud computing platform. The label vector is the sample data fusion vector corresponding to the sample operation data. The processing module is also used to receive global model parameters sent by the cloud computing platform, and obtain a standard fused data vector based on the device operation data, the global model parameters and the pre-stored multi-source data fusion model. A construction module is used to build a device-level digital twin model based on the standard fused data vector.

9. A device for constructing a digital twin model, applied to a cloud computing platform, characterized in that, include: The receiving module is used to receive device operation data sent by each edge computing node; The sending module is used to send macroscopic data to each of the edge computing nodes; The sending module is also used to initialize a pre-stored multi-source data fusion model and send the initialization parameters to each edge computing node; The processing module is used to receive the trained model parameters sent by each of the edge computing nodes, and to perform weighted processing on each of the trained model parameters to obtain global model parameters. The sending module is also used to send global model parameters to each of the edge computing nodes; The module is used to construct a system-level digital twin model based on the device's operating data and macroscopic data.

10. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-7.

Images