Power distribution box load distribution method and system based on digital twinning

Abstract

The application relates to the field of intelligent regulation and control of power systems, and provides a power distribution box load distribution method and system based on digital twinning, which comprises the following steps: acquiring current working condition data of a power distribution box; performing constraint reasoning based on a multilevel knowledge graph containing state parameter nodes, implicit influence correlation and time sequence evolution rules to generate constraint data; injecting the constraint data into a digital twinning simulation model to perform parallel simulation calculation on a candidate load distribution scheme; based on the simulation result, triggering knowledge graph update constraint data, and returning to the injection simulation model step until a preset convergence condition is met; and based on the simulation result meeting the convergence condition, determining a target load distribution scheme. Through the alternative iteration closed loop of simulation calculation and knowledge constraint, the application improves the accuracy and executability of the load distribution scheme.

Description

Technical Field

[0001] This application relates to the field of intelligent control of power systems, and in particular to a load distribution method and system for distribution boxes based on digital twins. Background Technology

[0002] Load allocation in distribution boxes is a crucial link in ensuring the safe and stable operation of low-voltage distribution networks. It typically relies on digital twin technology to construct a virtual mirror model, combined with a knowledge graph to store topology and operating rules, and generates load adjustment strategies through simulation. In a conventional technical architecture, the digital twin model is responsible for real-time mapping of physical states and parallel computation across multiple scenarios, while the knowledge graph serves as a static knowledge base providing parameter queries and relationship retrieval. The two interact through a data interface to aid decision-making.

[0003] However, in practical applications, the physical environment of the distribution box exhibits strong coupling of multiple factors and time-varying nonlinear characteristics. For example, changes in ambient temperature simultaneously affect cable current carrying capacity and air conditioning load power; the charging behavior of new energy vehicles exhibits both time-period clustering and random fluctuations; and there is a complex electromagnetic coupling relationship between harmonic content and neutral line current. Therefore, when the digital twin model performs parallel simulation, its calculation process mainly relies on preset physical equations and real-time acquired data. The implicit associations and temporal evolution patterns in the knowledge graph are only invoked discretely before or after the simulation begins, failing to be embedded as dynamic constraints within the iterative loop of the simulation calculation. Because simulation and knowledge reasoning are serial or loosely coupled on the time axis, intermediate states generated during the simulation cannot trigger real-time verification at the knowledge level. This prevents the implicit constraints in the knowledge graph from dynamically refining as the simulation progresses, leading to a deviation between mathematical optimality and physical feasibility in the final output load allocation scheme. Summary of the Invention

[0004] To address the problem that existing technologies lack tight coupling and alternating iteration between digital twin simulation and knowledge graph constraint reasoning, resulting in insufficient accuracy of load allocation schemes, this application proposes a digital twin-based distribution box load allocation method and system. This method achieves closed-loop co-evolution of simulation calculation and knowledge constraints, thereby improving the accuracy and executability of load allocation schemes.

[0005] To achieve the above objectives, this application adopts the following technical solution: Firstly, a load allocation method for distribution boxes based on digital twins includes: acquiring current operating condition data of the distribution box; performing constraint reasoning on the current operating condition data based on a pre-constructed multi-level knowledge graph to generate constraint data, whereby the multi-level knowledge graph contains state parameter nodes, implicit influence relationships between state parameter nodes, and temporal evolution patterns of state parameter nodes; injecting the constraint data into a digital twin simulation model to perform parallel simulation calculations on a preset number of candidate load allocation schemes, obtaining simulation results for each candidate load allocation scheme; triggering constraint reasoning on the multi-level knowledge graph based on the simulation results, updating the constraint data, and returning to execute the step of injecting constraint data into the digital twin simulation model until a preset convergence condition is met; and determining the target load allocation scheme from the preset number of candidate load allocation schemes based on the simulation results that meet the preset convergence condition.

[0006] The above scheme injects constraint data generated by the knowledge graph into the digital twin simulation model, and uses the simulation results to trigger the knowledge graph to update the constraint data, forming an alternating iterative closed loop of simulation calculation and constraint reasoning. This ensures that each round of simulation is carried out under dynamic knowledge constraints, and that the final target load allocation scheme simultaneously satisfies the mathematical optimization objective and the physical system safety constraints.

[0007] In conjunction with the first aspect mentioned above, in one possible implementation, the multi-level knowledge graph includes a basic entity layer and a state parameter layer. The basic entity layer contains distribution box equipment entities, phase line entities, and electrical equipment entities. The state parameter layer contains electrical state parameter nodes and environmental state parameter nodes. The basic entity layer and the state parameter layer are connected through inter-layer mapping. The implicit influence relationships within the state parameter layer include pairwise influence implicit relationships and multi-factor mutual influence implicit relationships, where multi-factor mutual influence implicit relationships are represented by hyperedges, which simultaneously connect three or more state parameter nodes.

[0008] In conjunction with the first aspect mentioned above, in one possible implementation, the process of determining pairwise influence latent associations and multi-factor mutual influence latent associations includes: acquiring the time-series data sequences of each state parameter node within a continuous time window; calculating the dynamic time warping distance between the time-series data sequences of any two state parameter nodes, and determining the initial association strength of the pairwise influence latent association based on the dynamic time warping distance; performing a Granger causality test on the time-series data sequences of the two state parameter nodes, determining the influence direction of the pairwise influence latent association based on the test results, and generating the pairwise influence latent association based on the initial association strength and influence direction; for any parameter combination consisting of three or more state parameter nodes, using the time-series data sequences of each state parameter node as input, learning the attention weights between nodes within the parameter combination through a time-series graph attention network, and determining the hyperedge connection mode and joint influence weight of the multi-factor mutual influence latent association based on the attention weights.

[0009] With reference to the first aspect, in a possible implementation, the representation of the time evolution law comprises: decomposing the time sequence data of each state parameter node into a trend component, a periodic component and a random component. Based on the trend component and the periodic component, a state transition probability matrix of each state parameter node under different time periods and different conditions is constructed. Based on the random component and the modeling of the change law of the strength of the implicit influence correlation over time, the correlation evolution law is obtained.

[0010] With reference to the first aspect, in a possible implementation, the process of constructing the digital twin simulation model comprises: obtaining electrical topology parameters, device parameters and cable parameters of the distribution box, and constructing a physical model of the distribution box. Real-time operation data of the distribution box is obtained, and the real-time operation data is mapped to each node of the physical model to construct a data model that is updated synchronously with the physical model. According to the physical model and the data model, a simulation model for parallel simulation calculation is constructed, and the simulation model integrates a three-phase power flow calculation sub-module, a harmonic analysis sub-module and a temperature rise simulation sub-module. The simulation model takes the state parameters of each node of the distribution box as input, and takes the three-phase voltage, the three-phase current, the active power, the reactive power, the harmonic content of each order and the node temperature of each node obtained by parallel simulation calculation as output.

[0011] With reference to the first aspect, in a possible implementation, the constraint data comprises a theoretical value range constraint, a parameter relationship constraint and a time sequence trend constraint, and the process of injecting the constraint data into the digital twin simulation model comprises: injecting the theoretical value range constraint by adjusting the value boundary of the corresponding parameter in the digital twin simulation model. The parameter relationship constraint is injected by adding a constraint equation or a correction coefficient between parameters in the digital twin simulation model. The time sequence trend constraint is injected by setting the parameter variation margin in the prediction time domain for the digital twin simulation model.

[0012] With reference to the first aspect, in a possible implementation, the preset convergence condition comprises: the output difference of two consecutive iterations is less than a preset threshold, or the maximum number of iterations is reached, and the process of determining the target load distribution scheme from the preset candidate load distribution schemes comprises: performing compliance checking of hard constraints and soft constraints on the simulation results that meet the preset convergence condition, the hard constraints comprising device safe operation constraints, and the soft constraints comprising power quality optimization constraints. The candidate load distribution scheme that passes the compliance checking is recorded as the target load distribution scheme.

[0013] With reference to the first aspect, in a possible implementation, after the target load distribution scheme is executed, actual operation data is obtained. Based on the deviation between the actual operation data and the simulation results, the model parameters of the digital twin simulation model and the correlation weights of the implicit influence correlations in the multi-level knowledge graph are updated simultaneously.

[0014] In a second aspect, a power distribution box load distribution system based on digital twinning is provided, comprising: the system comprises: a perception layer configured to acquire current working condition data of the power distribution box; a knowledge construction layer configured to perform constraint reasoning on the current working condition data based on a pre-constructed multi-level knowledge graph to generate constraint data, the multi-level knowledge graph comprising state parameter nodes, implicit influence correlations between the state parameter nodes, and time sequence evolution rules of the state parameter nodes; a twinning modeling layer configured to construct a digital twinning simulation model; a collaborative reasoning layer configured to inject the constraint data into the digital twinning simulation model, perform parallel simulation calculation on a preset number of candidate load distribution schemes, and obtain simulation results of the candidate load distribution schemes; based on the simulation results, trigger the multi-level knowledge graph to perform constraint reasoning, update the constraint data, and return to the step of injecting the constraint data into the digital twinning simulation model until a preset convergence condition is met; and a decision execution layer configured to determine a target load distribution scheme from the preset number of candidate load distribution schemes based on the simulation results that meet the preset convergence condition.

[0015] In combination with the first aspect described above, in a possible implementation, the collaborative reasoning layer is configured to perform an alternating iterative data flow of simulation calculation of the digital twinning simulation model and constraint reasoning of the multi-level knowledge graph, and the decision execution layer is configured to provide a feedback data flow to the knowledge construction layer and the twinning modeling layer.

[0016] The present application provides a power distribution box load distribution method and system based on digital twinning. By injecting constraint data generated by a knowledge graph into a digital twinning simulation model, and using simulation results to trigger the knowledge graph to update the constraint data, an alternating iterative mechanism of tightly coupled simulation calculation and knowledge constraint reasoning is constructed. Under this mechanism, each round of parallel calculation of the digital twinning simulation model is dynamically constrained by the implicit influence correlations and time sequence evolution rules in the knowledge graph, avoiding theoretical deviations caused by the simulation process deviating from the implicit constraints of the physical system. At the same time, the constraint reasoning of the knowledge graph is also updated and refined in real time according to the simulation intermediate results, so that the constraint conditions are self-adaptively adjusted as the simulation depth advances. This two-way feedback iterative closed loop can effectively ensure that the final output target load distribution scheme not only meets the mathematical optimization objective, but also strictly meets the equipment safe operation boundary and power quality requirements, thereby improving the accuracy and executability of the load distribution scheme in the actual physical system.

[0017] It should be understood that the description of technical features, technical solutions, advantages or similar language in this application does not imply that all features and advantages can be realized in any single embodiment. On the contrary, it can be understood that the description of a feature or advantage means that the specific technical feature, technical solution or advantage is included in at least one embodiment. Therefore, the description of technical features, technical solutions or advantages in this specification does not necessarily refer to the same embodiment. Further, the technical features, technical solutions and advantages described in this embodiment can be combined in any appropriate manner. Those skilled in the art will understand that the embodiments can be implemented without one or more specific technical features, technical solutions or advantages of a particular embodiment. In other embodiments, additional technical features and advantages can be identified in specific embodiments that do not embody all embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 A flowchart of a power distribution box load distribution method based on digital twinning provided by an embodiment of the application is shown in the figure. Figure 2 A multi-level knowledge graph architecture diagram in a power distribution box load distribution method based on digital twinning provided by an embodiment of the application is shown in the figure. Figure 3 A two-by-two influence hidden association mining flowchart in a power distribution box load distribution method based on digital twinning provided by an embodiment of the application is shown in the figure. Figure 4 A multi-factor mutual influence hidden association (hyperedge) mining flowchart in a power distribution box load distribution method based on digital twinning provided by an embodiment of the application is shown in the figure. Figure 5 A time sequence evolution rule representation flowchart in a power distribution box load distribution method based on digital twinning provided by an embodiment of the application is shown in the figure. Figure 6 A digital twinning simulation model construction and constraint injection diagram in a power distribution box load distribution method based on digital twinning provided by an embodiment of the application is shown in the figure. Figure 7 A simulation and constraint alternating iteration convergence judgment flowchart in a power distribution box load distribution method based on digital twinning provided by an embodiment of the application is shown in the figure. Figure 8 A target scheme compliance verification and double-loop feedback update flowchart in a power distribution box load distribution method based on digital twinning provided by an embodiment of the application is shown in the figure. Figure 9 An architecture diagram of a power distribution box load distribution system based on digital twinning provided by an embodiment of the application is shown in the figure. DETAILED DESCRIPTION

[0019] In the description of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. The "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" means one or more, and "multiple" means two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0020] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0021] Example 1: like Figure 1 As shown, this embodiment provides a load allocation method for distribution boxes based on digital twins. This method achieves the collaborative evolution and accurate decision-making of load allocation schemes by constructing a two-way feedback mechanism between digital twin simulation and knowledge graph-constrained reasoning. Specifically, the method includes the following steps: Step S100: Obtain the current operating status data of the distribution box.

[0022] Among them, the current operating condition data refers to a multi-dimensional data set that characterizes the operating status of the distribution box at a specific moment and its external environment. It not only covers traditional electrical operating parameters, but also includes environmental factors and equipment status information that affect load characteristics.

[0023] In some implementations, the system uses intelligent sensor networks and environmental sensing units deployed at key nodes of the distribution box to collect real-time electrical status parameters such as three-phase voltage, three-phase current, neutral current, active power, reactive power, power factor, and frequency, as well as environmental and time parameters such as terminal temperature, ambient temperature and humidity, and the current time period. The collected multi-source heterogeneous data is then cleaned, normalized, and time-aligned via an edge computing gateway to form a standardized operating condition data stream.

[0024] It should be noted that by incorporating environmental parameters and equipment status parameters into the scope of operating condition data, the physical boundary conditions of the distribution box operation can be fully reflected, providing a complete input basis for subsequent constraint reasoning of the knowledge graph and avoiding deviation of the reasoning results from the actual operating conditions due to missing information.

[0025] Step S200: Based on the pre-constructed multi-level knowledge graph, perform constraint reasoning on the current working condition data to generate constraint data. The multi-level knowledge graph includes state parameter nodes, implicit influence relationships between state parameter nodes, and temporal evolution patterns of state parameter nodes.

[0026] Among them, constraint data refers to structured information generated by knowledge graph reasoning, which is used to limit the operating boundaries and parameter relationships of digital twin simulation models. It can transform the implicit association knowledge stored in the knowledge graph into physical constraints that the simulation model can recognize.

[0027] Implicit influence relationships refer to the coupling relationships between parameters that cannot be directly derived from the physical topology but exist in historical operational data.

[0028] Temporal evolution refers to the dynamic characteristics of how parameters and their correlation strength change over time.

[0029] In some implementations, after receiving the current working condition data, the knowledge graph engine can first locate the corresponding working condition node in the graph, and then perform multi-hop reasoning along the implicit influence association edges and the path of temporal evolution.

[0030] For example, when an increase in ambient temperature is detected, not only can the derating constraint of cable current carrying capacity be inferred, but also the constraint of the indirect impact of the air conditioning load growth trend on the phase current can be inferred by combining time period information. Finally, the constraint dataset containing the theoretical value range, the equation of relationship between parameters and the predicted time domain margin is output.

[0031] It should be noted that this knowledge graph-based constraint reasoning mechanism allows the simulation model to operate without relying solely on fixed physical equations, but can dynamically incorporate implicit patterns mined from historical experience. This establishes a cognitive framework that closely resembles the real physical system before the simulation starts, effectively improving the physical realism of the simulation.

[0032] Step S300: Inject constraint data into the digital twin simulation model, perform parallel simulation calculations on the preset candidate load allocation schemes, and obtain the simulation results of each candidate load allocation scheme.

[0033] Injection refers to the process of mapping constraint data to the internal parameter boundaries, correction coefficients, or constraint equations of the simulation model, so that the simulation calculations run on a knowledge-defined track.

[0034] Candidate load allocation schemes refer to a variety of possible load adjustment strategies preset according to the current operating conditions, such as phase-to-phase load switching and equipment switching combinations.

[0035] In some implementations, the system can first dynamically adjust the boundary values ​​of the corresponding parameters in the simulation model based on the constraint data (such as correcting the upper limit of current carrying capacity in the temperature rise model), and add coupling equations between parameters (such as the correlation constraint between neutral current and harmonic content), thereby setting the parameter change margin for future periods.

[0036] Multiple simulation scenarios are launched in parallel. Each scenario corresponds to a candidate scheme, and three-phase power flow calculation, harmonic analysis, and temperature rise simulation are performed independently in the simulation model after the constraints are injected. Thus, the predicted results of voltage, current, power, harmonic content, and node temperature under each scheme can be output.

[0037] It should be noted that by injecting knowledge constraints into the simulation process in advance, it can be ensured that the deduction of all candidate solutions is carried out under the premise of conforming to physical laws and safety boundaries. This avoids the invalid solutions that are theoretically feasible but practically infeasible due to the neglect of implicit constraints in traditional simulation, thereby improving the efficiency of parallel computing and the reliability of the results.

[0038] Step S400: Based on the simulation results, trigger the multi-level knowledge graph to perform constraint reasoning, update the constraint data, and return to execute the step of injecting the constraint data into the digital twin simulation model until the preset convergence condition is met.

[0039] Alternating iteration refers to the process in which simulation calculation and knowledge-constrained reasoning are executed cyclically and mutually drive each other on the time axis. The output of each round of simulation becomes the new input for the next round of knowledge reasoning, and the updated knowledge constraints, in turn, refine the next round of simulation.

[0040] The preset convergence condition refers to the standard for determining whether the iterative process can be terminated. It is usually the difference between the outputs of two consecutive iterations being less than a preset threshold or the maximum number of iterations being reached.

[0041] In some implementations, after the Nth round of simulation is completed, the system can feed back the key state parameters from the simulation results to the knowledge graph engine. The graph engine can then reassess the applicability of the current constraints based on these new states.

[0042] For example, if a certain scheme is found to cause local temperature rise to exceed expectations, a stricter constraint on the correlation between temperature rise and current carrying capacity is triggered, generating updated constraint data C(N+1). Then, C(N+1) is injected into the simulation model again for the N+1th round of calculation. This cycle continues until the imbalance or line loss change rate between two adjacent rounds is less than 0.2% or the iteration reaches 5 times.

[0043] It should be noted that this tightly coupled alternating iterative mechanism can effectively break the limitations of serial or loosely coupled simulation and knowledge reasoning in traditional technologies, and realize positive collaborative evolution in which the deeper the simulation, the more accurate the constraints, and the more accurate the constraints, the more realistic the simulation. This ensures that the simulation results in the final convergence state not only meet the mathematical optimization objectives, but also strictly conform to the constraints of the dynamically changing physical system.

[0044] Step S500: Based on the simulation results that meet the preset convergence conditions, determine the target load allocation scheme from the preset candidate load allocation schemes.

[0045] Among them, the target load allocation scheme refers to the optimal load adjustment strategy that, after sufficient iterative verification, simultaneously satisfies the hard constraints of equipment safe operation and the soft constraints of power quality optimization.

[0046] In some implementations, the system performs a dual compliance check on the simulation results of each converged candidate scheme: First, verify whether the equipment safety constraints such as cable current carrying capacity and voltage deviation limits are violated, and eliminate unqualified solutions. Secondly, for the schemes that pass the hard constraint verification, they are sorted according to the multi-objective optimization function, such as minimizing imbalance, minimizing line loss, and minimizing the number of commutations, and the scheme with the highest comprehensive score is selected as the target scheme output.

[0047] It should be noted that, since this determination process is based on simulation results that have been repeatedly calibrated with knowledge constraints, the output target solution is not only theoretically optimal, but also highly executable and secure in the actual physical system. This effectively solves the technical problem of large deviations between simulation solutions and on-site execution effects in existing technologies.

[0048] Based on the above technical solution, this embodiment achieves deep integration and collaborative evolution of digital twin simulation deduction and knowledge graph constraint reasoning by constructing a complete closed-loop process of acquiring operating condition data, knowledge-constrained reasoning, constraint injection simulation, alternating iterative updates, and target scheme determination. This process utilizes implicit associations and temporal patterns in the knowledge graph to dynamically constrain the simulation process, and uses simulation results to drive knowledge updates in reverse, forming a tightly coupled bidirectional feedback mechanism. This ensures the accuracy, physical executability, and adaptive evolution capability of the load allocation scheme under complex operating conditions with multiple coupled factors.

[0049] Example 2: like Figures 2-5As shown, this embodiment, based on Embodiment 1, further details the specific construction method of the multi-level knowledge graph and its internal implicit associations and temporal pattern mining mechanism. This mechanism aims to transform the complex coupling relationships in the operation of the distribution box, which are difficult to describe using explicit physical equations, into computable and reasonable structured knowledge, providing high-precision cognitive constraints for digital twin simulation. Specifically, it includes the following steps: Step S210: Construct the basic entity layer and state parameter layer of the multi-level knowledge graph, and define the inter-layer mapping association and hyperedge structure.

[0050] Among them, the basic entity layer refers to the static digital abstraction of the physical composition and topological connection relationship of the distribution box.

[0051] The state parameter layer refers to the independent semantic representation of dynamic electrical quantities and environmental quantities that reflect the operating state of the system.

[0052] A hyperedge is a generalized association structure that can connect three or more state parameter nodes at the same time, thereby expressing the combined influence of multiple factors.

[0053] In some implementations, the construction of the basic entity layer can begin by extracting the distribution box equipment entity, the A / B / C three-phase line entity, and the electrical equipment entities such as air conditioners, charging piles, and lighting from the distribution box ledger, electrical drawings, and sensor configuration table, and establishing explicit topological relationships such as "connected to", "belonging to", and "installed in".

[0054] This allows electrical state parameters such as three-phase voltage, three-phase current, neutral line current, active power, reactive power, and harmonic content, as well as environmental state parameters such as terminal temperature, ambient temperature and humidity, time period, and proportion of new energy vehicles, to be incorporated into the graph as independent nodes. The two layers are then connected through inter-layer mapping to construct a state parameter layer.

[0055] For example, map "A-phase current node" to "A-phase line entity" and map "terminal temperature node" to "distribution box equipment entity".

[0056] When three or more parameters are detected to have an inseparable combined effect (such as "ambient temperature + harmonic content + load rate" jointly determining "cable temperature rise rate"), a superedge can be created to connect these nodes, thus defining the superedge structure inside the state parameter layer.

[0057] It should be noted that this dual-layer separation plus hyperedge architecture design retains the topological anchor points of physical devices, thus aligning with the digital twin model, while decoupling dynamic parameters from device entities to form an independent reasoning space. This allows the knowledge graph to focus on the discovery and reuse of general patterns between parameters, detached from specific device instances. Furthermore, the introduction of the hyperedge structure breaks through the limitation of traditional knowledge graphs, which can only express binary relationships.

[0058] Step S220: Mining the pairwise influence hidden correlations between state parameter nodes based on time series data sequences.

[0059] Among them, the pairwise influence implicit correlation refers to the dynamic causal relationship with specific directionality and intensity that exists between any two state parameter nodes. It is mainly based on the time-series dependency pattern derived from historical operation data statistics.

[0060] In some implementations, the time-series data sequence of each state parameter node within a continuous time window (e.g., the past 30 days, with a sampling interval of 5 minutes) is first obtained, and then any two nodes can be targeted. and The Dynamic Time Warping (DTW) algorithm is used to calculate the optimal alignment path distance between the two sequences. This distance can effectively overcome the time axis distortion caused by load response lag or sensor sampling asynchrony. The DTW distance is normalized and used as the initial correlation strength.

[0061] right and Perform Granger causality tests on the time series, construct a vector autoregressive model and calculate the F-statistic. If " Historical values ​​help to significantly reduce If the prediction error of the current value passes the significance test (e.g., p < 0.05, where p is the probability value), then it is determined that there exists a difference between the prediction error of the current value and the prediction error of the current value. point to The direction of influence; Therefore, by combining the initial correlation strength and influence direction, weighted directed pairwise implicit influence edges can be generated.

[0062] For example, this method can reveal that "ambient temperature" has a positive impact on "phase A current" (the increase in temperature leads to an increase in air conditioning load, which in turn increases the current), and this correlation has a time lag characteristic of about 15 minutes.

[0063] It should be noted that the combination of DTW and Granger test can strictly distinguish between "covariance" and "causality" and quantify the effects of time lag, thereby ensuring that the generated implicit associations have real physical interpretability and predictive value.

[0064] Step S230: Mining hidden associations of multiple factors interacting between state parameter nodes based on time sequence graph attention network.

[0065] Among them, the hidden association of multiple factors refers to a high-order coupling mode formed by the joint participation of three or more state parameter nodes, which cannot be replaced by the linear superposition of pairwise relationships. It is stored in the form of a hyperedge and comes with joint influence weights.

[0066] In some implementations, the temporal data sequence of potential multi-factor parameter combinations identified within a preset or sliding window is fed into the Time Series Graph Attention Network (TGAT). The TGAT network then uses a time encoder to map consecutive timestamps into location embeddings, enabling the model to perceive the precise temporal intervals between events.

[0067] Thus, a multi-head self-attention mechanism can be used to pass messages between parameter nodes and automatically learn the attention weight of each node to the target node at different time scales. At this time, the weight reflects the contribution of each factor to the joint effect under specific context conditions. For example, under the conditions of "evening peak + high temperature", the attention weight of "charging pile power" to "neutral line current" is significantly higher than that at other times.

[0068] Based on the learned attention weights, the connection method and joint influence weight of the hyperedge are determined, and the hyperedge and its weights are stored in the state parameter layer of the knowledge graph.

[0069] It should be noted that many key phenomena in the distribution box system are essentially high-order. For example, "neutral overload" is often not caused by a single harmonic source, but is the result of resonance under a specific combination of "fundamental imbalance + 3rd harmonic + 5th harmonic + zero-sequence impedance". Therefore, the traditional pairwise correlation analysis method will incorrectly disperse this high-order effect into multiple weakly correlated edges, resulting in information loss. However, this application directly learns high-order interaction features on the graph structure through the TGAT network and completely preserves this combination pattern in the form of hyperedges, so that subsequent constraint inference can accurately identify and warn of such hidden composite risks.

[0070] Step S240: Construct a temporal evolution model of state parameter nodes.

[0071] Among them, the temporal evolution law refers to the dynamic change pattern of the state parameters themselves and their correlation strength over time, including long-term trends, periodic fluctuations and random disturbances, as well as the non-stationary characteristics of the correlation in different time periods or under different conditions.

[0072] In some implementations, the system first performs STL decomposition or empirical mode decomposition on the time series data of each state parameter node to separate trend components, periodic components and random components. Based on the trend components and periodic components, the system calculates the state transition frequency of each parameter in different time periods (such as weekday daytime, weekend nighttime, summer, winter) and different condition intervals, and constructs a state transition probability matrix.

[0073] For example, "If the current three-phase imbalance is in the range of 5%-8% and the time is 19:00-21:00 in summer, the probability of it rising to the range of 8%-12% within the next 30 minutes is 0.72".

[0074] Then, the strength of the implicit influence correlation is modeled over time. For example, it is found that the "influence coefficient of temperature on current carrying capacity" reaches its peak between 14:00 and 16:00 every day and decays at night. This time-varying pattern is fitted into a correlation evolution function and attached to the corresponding correlation edge.

[0075] It should be noted that by explicitly modeling the temporal evolution pattern and embedding it into the graph, this application enables the knowledge system to have time-awareness capabilities, allowing constraint reasoning to automatically activate the matching evolution model based on the current moment. For example, it can automatically increase the expected growth of air conditioning load and tighten the load-carrying constraints in the summer evening, while relaxing them accordingly in the winter night, thus enabling it to achieve dynamic adaptive constraints that truly fit the actual working conditions.

[0076] Based on the above technical solutions, this embodiment solves the problem that traditional knowledge graphs cannot express high-order couplings by constructing a two-layer separated knowledge graph architecture and introducing a hyperedge structure. Internally, it mines pairwise hidden associations through a combination of DTW and Granger criterion, overcoming the interference of time delay and spurious correlations; it learns multi-factor joint weights through the TGAT network, preserving the nonlinear interaction characteristics of complex systems; and it endows the knowledge graph with dynamic time-varying perception capabilities by modeling temporal evolution laws. These technical means together constitute a high-precision, interpretable, and adaptive cognitive core, providing deep constraint support for digital twin simulation far exceeding that of static rule bases, thereby improving the accuracy and safety of load allocation schemes under complex and variable operating conditions.

[0077] Example 3: like Figure 6 As shown, this embodiment, based on Embodiment 1, further details the specific construction process of the digital twin simulation model and the implementation mechanism for injecting constraint data generated from the knowledge graph into the model. This mechanism aims to transform abstract knowledge constraints into physical boundaries and mathematical relationships that the simulation solver can recognize, ensuring that the simulation always operates on a cognitive track that conforms to actual working conditions. Specifically, it includes the following steps: Step S310: Construct a digital twin simulation model.

[0078] Among them, the digital twin simulation model refers to a comprehensive virtual entity that integrates the physical mechanism of the distribution box, real-time operation data mapping, and multi-physics field coupling calculation capabilities. It is organically composed of three parts: physical model, data model, and simulation model.

[0079] The physical model is a static mechanistic description built upon electrical topology, equipment nameplate parameters, and cable specifications.

[0080] The data model is a dynamic state mirror formed by mapping the real-time running data collected by the perception layer to each node of the physical model.

[0081] The simulation model is a set of executable algorithms built on the basis of the physical model and the data model, used to perform parallel computing. It integrates a three-phase power flow calculation submodule, a harmonic analysis submodule, and a temperature rise simulation submodule.

[0082] In some implementations, the system first parses the CAD drawings and equipment ledger of the distribution box, and establishes a physical model that includes node admittance matrix and branch impedance parameters. Then, it can access the real-time IoT data stream through OPC UA or MQTT protocol, and assign the measured values ​​of voltage, current, temperature and other parameters to the corresponding nodes of the physical model to form a data model that is updated synchronously with the field.

[0083] A simulation model is constructed by tightly coupling and integrating the three-phase power flow calculation submodule (based on the Newton-Raphson method or forward-backward substitution method), the harmonic analysis submodule (based on frequency domain scanning or harmonic power flow method), and the temperature rise simulation submodule (based on thermal circuit model or finite element simplified model). This allows each submodule to share the same set of node state vectors. For example, the conductor temperature calculated by the temperature rise submodule is fed back in real time to correct the line resistance parameters in the power flow submodule, while the losses calculated by the power flow submodule are used as heat sources input to the temperature rise submodule. The final simulation model takes the state parameters of each node as input and outputs the three-phase voltage, three-phase current, active power, reactive power, harmonic content, and node temperature obtained by parallel calculation.

[0084] It should be noted that this three-model integrated construction method can break the limitations of the separation of physical models and real-time data and the decoupling of calculations of various physical fields in traditional simulations. This allows the simulation model to reflect the inherent physical characteristics of the equipment and track the dynamic operating status of the system in real time. At the same time, through the internal coupling iteration of multiple sub-modules, it accurately captures the mutual influence effects between multiple physical quantities such as electricity, heat and harmonics, providing high-fidelity computational support for the accurate injection of subsequent constraint data.

[0085] Step S320: By adjusting the value boundaries of the corresponding parameters in the digital twin simulation model, theoretical value range constraints are injected.

[0086] Among them, the theoretical value range constraint refers to the dynamic limit value that represents the safe operating range of the parameter, which is derived from the knowledge graph based on the current environment and equipment status.

[0087] Adjusting the boundary values ​​refers to replacing static fixed values ​​with knowledge-driven dynamic thresholds in the domain of variables of the simulation solver.

[0088] In some implementations, when the knowledge graph inference generates an allowable current carrying capacity of 85A for a cable under the current ambient temperature of 35℃ and aging level, the system automatically locates the current variable corresponding to that cable in the simulation model and dynamically modifies its upper limit constraint from the default 100A (standard rated value) to 85A. Similarly, if the knowledge graph inference determines that the voltage-sensitive equipment's tolerance for voltage deviation has decreased in the current period, the upper and lower limits of the voltage variable are tightened from ±7% to ±5%. This adjustment is implemented by directly writing it into the solver's constraint configuration file through the API interface during the simulation initialization phase or before the start of each iteration.

[0089] It should be noted that by injecting the dynamic safety boundary of knowledge reasoning into the variable domain of the simulation model in advance, the search space of all candidate solutions can be forced to be limited to the physical range that is actually feasible under the current working conditions. This can effectively avoid the simulation from producing mathematically optimal but physically non-compliant invalid solutions, thereby reducing the waste of ineffective computing resources and improving the convergence efficiency of the optimization process.

[0090] Step S330: Inject constraint on the relationship between parameters by adding constraint equations or correction coefficients between parameters in the digital twin simulation model.

[0091] Among them, the parameter relationship constraint refers to the implicit coupling law between parameters that cannot be fully covered by the basic physical equations, which is mined from the knowledge graph.

[0092] Adding constraint equations or correction coefficients means making these implicit rules explicit into mathematical expressions that can be executed by the simulation solver.

[0093] In some implementations, to address the "nonlinear coupling relationship between neutral current, third harmonic content, and fundamental imbalance" discovered in the knowledge graph, the system inserts an additional constraint equation between the harmonic analysis submodule and the power flow calculation submodule of the simulation model. In the formula, For neutral line current, For fundamental unbalanced current, It is the third harmonic current. The coupling coefficient is determined by the hyperedge weights of the graph.

[0094] In other implementations, a joint correction coefficient can be introduced in the temperature rise simulation submodule to address the "derating effect of cable current carrying capacity under the combined effect of temperature and harmonics". The original formula for calculating current carrying capacity, which relied solely on temperature, was revised to... These equations and coefficients serve as supplementary constraints to the simulation model in each iteration of the solution process.

[0095] It should be noted that this application transforms the implicit associations mined from knowledge graphs into specific mathematical constraints and embeds them into the simulation kernel, thereby compensating for the representational defects of pure mechanistic models under complex working conditions. This enables the simulation results to truly reflect the system behavior under the coupling effect of multiple factors, and significantly improves the safety and accuracy of the load allocation scheme in the actual physical environment.

[0096] Step S340: By setting the parameter change margin in the prediction time domain for the digital twin simulation model, time-series trend constraints are injected.

[0097] Among them, temporal trend constraints refer to the direction and magnitude of changes in key parameters over a future period of time, predicted by the knowledge graph based on the temporal evolution law.

[0098] Setting a prediction time-domain margin refers to pre-adding a dynamic buffer amount that reflects future trends to the boundary conditions or initial conditions of the simulation model.

[0099] In some implementations, when the knowledge graph infers that "the load on the charging pile cluster is expected to increase by 20% within the next 30 minutes," the system does not directly use the current measured load value as the steady-state input in the simulation model. Instead, it constructs a load curve that grows linearly or non-linearly over time as the simulation boundary condition, or directly superimposes a 20% margin component on the current load as the worst-case scenario for verification. Simultaneously, for parameters with periodic fluctuation characteristics (such as air conditioning load), based on the periodic component model in the knowledge graph, corresponding sinusoidal or square wave modulation functions are set in the simulation time domain to simulate their natural fluctuations in future periods.

[0100] This margin setting allows the simulation to not only assess the feasibility at the current moment, but also verify the robustness of the scheme within the future prediction window.

[0101] It should be noted that this application transforms the temporal prediction capability of knowledge graphs into the dynamic boundary margin of simulation models, giving simulation a forward-looking perspective and ensuring that the output target solution is not only optimal at the current moment, but also remains stable and safe in future evolution trends, effectively solving the problem of the disconnect between static optimization and dynamic operating conditions.

[0102] Based on the above technical solution, this embodiment constructs a digital twin integrating physical data and simulation models. It then deeply injects three types of constraint data—the theoretical value range generated from the knowledge graph, the relationships between parameters, and the time-series trend—into the simulation model kernel by adjusting variable boundaries, adding constraint equations / correction coefficients, and setting prediction time-domain margins. This achieves comprehensive and deep-level constraints of knowledge on the simulation. Simultaneously, this injection mechanism transforms the simulation model from a purely mathematical calculator detached from cognition into an intelligent inference engine with environmental perception and empirical reasoning capabilities. This effectively ensures that the load allocation scheme generation process always conforms to the dynamic safety boundaries and implicit coupling laws of the actual physical system, thereby improving the physical executability and long-term operational stability of the scheme.

[0103] Example 4: like Figures 7-8 As shown, this embodiment, based on Embodiment 1, further details the specific implementation mechanism of iterative convergence determination, scheme compliance verification, and dual-path update based on execution feedback. This mechanism constitutes the decision-making exit and adaptive evolution core of the load distribution system, ensuring that the output target scheme not only meets real-time safety constraints but also continuously optimizes system accuracy through historical execution data. Specifically, it includes the following steps: Step S410: Determine whether the alternating iteration process meets the preset convergence condition.

[0104] Among them, the preset convergence condition refers to the quantitative criterion used to terminate the alternating iterative cycle of digital twin simulation and knowledge graph constrained reasoning. It can effectively balance computational efficiency and scheme accuracy, and prevent the system from falling into invalid oscillations or overcomputation.

[0105] In some implementations, the system employs a dual convergence criterion, wherein: The first criterion is the output difference threshold, which monitors the changes in key indicators (such as three-phase imbalance, maximum cable temperature, and overall line loss rate) output in two consecutive iterations. When the absolute value of the change is less than the preset threshold (for example, the change in imbalance is less than 0.2% and the change in temperature is less than 0.5℃), it is determined that the system has entered the steady-state convergence range. The second criterion is a maximum limit on the number of iterations (e.g., set to 5 or 8 times), as a safety circuit breaker mechanism to prevent infinite loops caused by data anomalies or model conflicts.

[0106] When any criterion is met, the iteration stops and the simulation results of the current round are locked as the candidate set to be verified.

[0107] It should be noted that relying solely on the difference threshold may lead to difficulty in convergence under drastic fluctuations in operating conditions, while relying solely on the upper limit of the number of iterations may result in a coarse solution that is not sufficiently optimized. Therefore, this application effectively combines the two through a dual convergence mechanism, ensuring that the system can accurately approximate the physical reality under normal operating conditions, while also guaranteeing the real-time response capability and controllability of computing resources under extreme and abnormal conditions.

[0108] Step S420: Perform compliance checks on the simulation results that meet the preset convergence conditions using hard and soft constraints, and record the candidate load allocation schemes that pass the compliance checks as the target load allocation schemes.

[0109] Among them, hard constraints refer to the inviolable limits that characterize the physical safety boundaries of the distribution box and the mandatory national standards.

[0110] Soft constraints refer to quantifiable evaluation indicators that characterize the goals of power quality optimization and the expectations of economic operation.

[0111] Compliance checks refer to the final safety and applicability screening of all candidate solutions after iterative convergence and before solution output.

[0112] In some implementations, the system first performs hard constraint verification, comparing the simulation results of each candidate scheme with the equipment safety operation constraints, including but not limited to: whether the current of each phase cable exceeds the corrected current carrying capacity at the current ambient temperature, whether the node voltage deviation is within the allowable range of ±7%, whether the number of opening and closing of the switching equipment exceeds the limit, and whether the neutral line current is lower than the phase line rated value. Any scheme that violates any hard constraint is directly eliminated and the reason for the violation is recorded.

[0113] Subsequently, soft constraint evaluation can be performed on the remaining schemes that have passed the hard constraint verification. A comprehensive score is calculated based on the power quality optimization constraints, such as weighted calculation of indicators such as three-phase imbalance (target ≤5%), total harmonic distortion rate (target ≤4%), expected line loss rate and commutation operation cost. The scheme with the best comprehensive score is selected as the target load allocation scheme output.

[0114] It should be noted that since simulation results only possess true physical credibility after sufficient knowledge constraints have been injected and a steady state has been reached, compliance checks are placed after iterative convergence rather than before. Furthermore, in collaboration with hard and soft constraints, the hierarchical processing mechanism of these constraints effectively ensures logical isolation between the safety baseline and the optimization ceiling. This avoids the potential risk in traditional multi-objective optimization where improper weight settings lead to sacrificing safety for economic performance, thus guaranteeing that the output target solution is absolutely safe and relatively optimal in the actual physical system.

[0115] Step S430: After the target load allocation scheme is executed, obtain the actual operating data; based on the deviation between the actual operating data and the simulation results, update the model parameters of the digital twin simulation model and the association weights of implicit influence associations in the multi-level knowledge graph.

[0116] Among them, actual operation data refers to the actual response status of the distribution box collected by the perception layer after the target plan is issued and executed.

[0117] Deviation refers to the difference between the actual operating data and the simulation prediction value corresponding to the target scheme output in step S420.

[0118] Dual-path update refers to a parallel feedback mechanism that utilizes this deviation to simultaneously drive the calibration of physical parameters of the digital twin model and the adjustment of cognitive weights in the knowledge graph.

[0119] In some implementations, the system continuously collects actual three-phase current, voltage, and temperature data within a preset observation window (e.g., 30 minutes) after the scheme is executed, and calculates the normalized deviation between these data and the simulation predictions. If the deviation exceeds the allowable range (e.g., current deviation > 3%), a dual-path update engine is triggered. One path is to use the deviation as an identification signal input to the parameter calibration module of the digital twin simulation model, and use the least squares method or Kalman filter algorithm to invert and correct the static parameters such as line impedance, equipment heat capacity, and contact resistance in the physical model, so that the simulation benchmark at the next moment is closer to the current state of the physical equipment. Another approach is to feed back the deviation and its corresponding operating context (time period, temperature, load type) to the association evolution module of the multi-level knowledge graph, and dynamically adjust the weight of the related implicit influence associations according to the direction and magnitude of the deviation.

[0120] For example, if the actual temperature rise is higher than the simulation prediction, the weight of the correlation between ambient temperature and cable temperature rise can be increased, or the joint influence coefficient between harmonic content and neutral current over-edge can be corrected.

[0121] These two update paths do not run independently, but rather have a mutual reference verification mechanism: after the twin model parameters are updated, the system will backtrack to check whether the adjustment of the correlation weights in the graph can be explained by the new parameters. If the new parameters can explain most of the bias, the excessive adjustment of the graph weights will be suppressed to prevent overfitting; otherwise, if the bias cannot be eliminated after the graph weights are adjusted, it will indicate that there may be unmodeled physical factors, triggering a deep diagnosis of the model structure.

[0122] It should be noted that this dual-path collaborative update mechanism can effectively construct a closed-loop evolutionary system from execution, feedback, dual-path calibration to mutual verification. This allows the physical fidelity of the digital twin model and the cognitive accuracy of the knowledge graph to improve synchronously with the accumulation of operating time. This effectively solves the problem of long-term operational accuracy decay caused by equipment aging, environmental changes and load characteristic drift in power distribution systems, and endows the system with continuous adaptive evolution capabilities.

[0123] Based on the above technical solutions, this embodiment improves the decision-making exit and evolutionary closed loop of the load allocation method by constructing a dual convergence determination, hierarchical compliance verification, and dual-path collaborative update mechanism. This ensures the stability and timeliness of the computation process through dual convergence criterion; while hierarchical verification of hard and soft constraints ensures the optimality of the output scheme under the premise of physical safety. Secondly, the dual-path mutual verification update mechanism based on execution deviation effectively breaks the limitation of model and knowledge evolution being separate, enabling the system to continuously self-correct and learn in actual operation, improving the accuracy, security, and adaptability of the load allocation strategy under long-term, variable operating conditions.

[0124] Example 5: like Figure 9 As shown, this embodiment provides a load distribution system for a distribution box based on digital twins. The system adopts a five-layer integrated collaborative architecture, deeply integrating digital twin simulation and knowledge graph reasoning in functional modules and data paths to support the execution of the aforementioned method embodiment. Specifically, the system includes a perception layer, a knowledge construction layer, a twin modeling layer, a collaborative inference layer, and a decision execution layer. Step S510: Configure the sensing layer to obtain the current operating status data of the distribution box.

[0125] The perception layer serves as the physical world interface and data standardization entry point for the system. It is responsible not only for collecting heterogeneous data from multiple sources but also for data preprocessing based on knowledge graph entity rules.

[0126] In some implementations, the perception layer integrates an intelligent sensor network, an environmental perception unit, a new energy information collection interface, and an edge computing gateway.

[0127] The edge computing gateway has a built-in lightweight entity recognition rule base that is periodically distributed by the knowledge building layer. At the data acquisition site, it completes the cleaning, normalization and semantic annotation of raw signals such as three-phase voltage, current, temperature and humidity, and transforms unstructured electrical signals into standardized chemical condition data streams with device entity tags and timestamps. The data is then distributed to the upper layer via MQTT or OPC UA protocol.

[0128] It should be noted that by embedding knowledge-driven preprocessing logic into the perception layer, a data distribution mechanism of one-time collection and dual-path supply is realized. This provides real-time driving data for the twin modeling layer and semantically labeled sample data for the knowledge construction layer, ensuring data consistency and semantic understandability for subsequent simulation and reasoning from the source.

[0129] Step S520: Configure the knowledge construction layer, which is used to perform constraint reasoning on the current working condition data based on the pre-built multi-level knowledge graph, and generate constraint data. The multi-level knowledge graph includes state parameter nodes, implicit influence relationships between state parameter nodes, and temporal evolution rules of state parameter nodes.

[0130] Among them, the knowledge construction layer is the cognitive core and dynamic knowledge base of the system, which includes a graph storage engine, a hidden association mining engine, a temporal evolution modeling component, and a constraint inference engine.

[0131] In some implementations, the graph storage engine uses an attribute graph model to persistently store the basic entity layer and state parameter layer data.

[0132] The latent association mining engine can receive real-time data streams from the perception layer and simulation data streams from the twin modeling layer, and then use DTW, Granger causality test and TGAT network to continuously update pairwise latent associations and hyperedge weights.

[0133] The temporal evolution modeling component is based on sliding window data and maintains the state transition probability matrix and associated evolution function of each node.

[0134] The constraint inference engine responds to the request from the co-inference layer, starts from the current working node, performs multi-hop inference along the graph path, and outputs structured constraint data including the theoretical value range, the equations of relationship between parameters, and the prediction time domain margin.

[0135] It should be noted that this layer is not a static knowledge repository, but a living cognitive system with online learning and dynamic evolution capabilities. It can continuously absorb measured data and simulation data to correct implicit correlations and temporal patterns, ensuring that the generated constraint data always reflects the latest operating characteristics and physical boundaries of the distribution box.

[0136] Step S530: Configure the twin modeling layer for building a digital twin simulation model.

[0137] Among them, the twin modeling layer is the physical mirror and computing base of the system, responsible for maintaining the virtual mapping body that evolves synchronously with the physical distribution box.

[0138] In some implementations, this layer integrates a physical model builder, a data mapper, a simulation solver, and a parameter calibration module.

[0139] The physical model builder parses the electrical topology and equipment ledger to generate node admittance matrices and thermal parameters.

[0140] The data mapper binds the real-time operating condition data uploaded by the perception layer to the corresponding node of the physical model.

[0141] The simulation solver encapsulates algorithms for three-phase power flow calculation, harmonic analysis, and temperature rise simulation, and supports parallel computing in multiple scenarios.

[0142] The parameter calibration module receives feedback data from the decision execution layer and corrects the model parameters using the least squares method or Kalman filtering inversion.

[0143] It should be noted that there is a horizontal data path between the twin modeling layer and the knowledge construction layer. The topological relationships in the knowledge graph guide the construction of the physical model, the hidden correlations are transformed into simulation correction coefficients, and the parameter deviations found in the simulation trigger the update of the graph attributes. This two-way feedback mechanism enables the digital twin to have both the rigor of the mechanistic model and the adaptability of the data-driven model.

[0144] Step S540: Configure the collaborative inference layer to inject constraint data into the digital twin simulation model, perform parallel simulation calculations on a preset number of candidate load allocation schemes, and obtain the simulation results of each candidate load allocation scheme; based on the simulation results, trigger the multi-level knowledge graph to perform constraint reasoning, update the constraint data, and return to execute the step of injecting constraint data into the digital twin simulation model until the preset convergence condition is met.

[0145] The collaborative inference layer is the system's decision engine and tightly coupled execution unit, which includes a parallel computing scheduler, constraint injection interface, iterative controller, and compliance verifier.

[0146] In some implementations, the parallel computing scheduler automatically generates multiple candidate load allocation schemes and starts multi-threaded simulation tasks based on the current operating conditions.

[0147] The constraint injection interface dynamically writes the constraint data output by the knowledge building layer into the variable boundaries and equation system of the simulation solver.

[0148] The iterative controller monitors the difference between the simulation output and the constraint update, and determines the convergence state based on a preset threshold or the maximum number of iterations.

[0149] After convergence, the compliance validator performs hierarchical verification of candidate solutions using both hard and soft constraints.

[0150] It should be noted that this layer breaks the architectural limitations of serial execution of simulation and inference in traditional systems, and can achieve alternating driving and mutual calibration of the two through an internal tight coupling mechanism. This ensures that each round of simulation is carried out under dynamic knowledge constraints, and that each inference is based on the latest simulation intermediate state, thereby guaranteeing the physical authenticity and computational efficiency of the decision-making process at the system level.

[0151] Step S550: Configure the decision execution layer to determine the target load allocation scheme from a preset number of candidate load allocation schemes based on the simulation results that meet the preset convergence conditions.

[0152] Among them, the decision execution layer is the system's strategy exit and evolutionary feedback source, responsible for converting the optimal calculation results into executable instructions and collecting execution effects.

[0153] In some implementations, this layer includes a multi-objective optimization sorter, an instruction generator, an effect monitor, and a bias analyzer.

[0154] The multi-objective optimization sorter scores the compliance-validated schemes based on indicators such as imbalance, line loss, and commutation cost.

[0155] The instruction generator compiles the optimal solution into control sequences for devices such as intelligent phase-changing switches.

[0156] The performance monitor continuously collects actual operational data after the plan is implemented.

[0157] The deviation analyzer calculates the difference between the measured values ​​and the simulation predictions, and distributes the deviation information to the twin modeling layer and the knowledge construction layer.

[0158] It should be noted that this layer is the starting point for the closed-loop evolution of the system. Through a structured deviation feedback mechanism, it drives the continuous self-correction of the upper-level model and knowledge, enabling the system to have the adaptive ability to continuously improve its accuracy as the running time increases.

[0159] Step S560: The collaborative inference layer executes the simulation calculation of the digital twin simulation model and the constrained reasoning of the multi-level knowledge graph, performing an alternating iterative data flow.

[0160] Among them, the alternating iterative data flow refers to the bidirectional, cyclic, and tightly coupled data transmission path formed between the simulation calculation engine and the constraint inference engine within the collaborative inference layer.

[0161] In some implementations, after the simulation engine completes a round of parallel computation, the data stream pushes the key state parameters (such as node voltage, current, temperature, and harmonic content) of each candidate scheme to the constraint inference engine in real time via an internal high-speed bus.

[0162] The constraint inference engine uses this as new input to re-evaluate the applicability of the current constraints and generate an updated constraint dataset, which is then injected back into the parameter configuration area of ​​the simulation calculation engine via the same bus.

[0163] The simulation engine automatically loads new constraints and resolves them in the next round of calculation.

[0164] The above process is executed cyclically on a millisecond timescale until the iterative controller issues a convergence signal.

[0165] It should be noted that this structured alternating iterative data flow design binds simulation and inference into an inseparable execution whole at the system architecture level, avoiding information lag and constraint failure caused by module decoupling. This ensures the dynamic consistency and physical security of the load distribution scheme generation process from both hardware and software architecture perspectives.

[0166] Step S570: The decision execution layer provides feedback data flow to the knowledge construction layer and the twin modeling layer.

[0167] Among them, the feedback data flow refers to the dual data channel through which the decision-making execution layer transmits information on the deviation between the actual operating effect and the expected goal to the knowledge construction layer and the twin modeling layer simultaneously and in a targeted manner after the plan is implemented.

[0168] In some implementations, the data stream contains two parallel branches: The first branch points to the parameter calibration module of the twin modeling layer, and transmits the normalized deviation vector between the measured electrical quantities and the simulation prediction quantities, which is used to drive the online identification and correction of physical model parameters. The second branch points to the association evolution module of the knowledge construction layer, and transmits the deviation vector and its corresponding working context (time period, temperature, load type), which is used to trigger the enhancement or attenuation adjustment of the implicit influence association weight. The data packets of both branches carry a unified timestamp and scheme ID to enable the receiver to perform cross-modal alignment and mutual verification.

[0169] It should be noted that this structured design of dual-path feedback data flow enables the system to distinguish between two different types of problems: model parameter errors and cognitive pattern deviations. Through a two-layer mutual reference verification mechanism, it prevents overfitting or erroneous updates from a single path, thereby building a robust closed-loop evolution capability at the system architecture level and ensuring that the load distribution system continuously improves decision-making accuracy and environmental adaptability during long-term operation.

[0170] Based on the above technical solution, this embodiment constructs a five-layer integrated architecture consisting of a perception layer, a knowledge construction layer, a twin modeling layer, a collaborative inference layer, and a decision execution layer. By combining the alternating iterative data flow within the collaborative inference layer with the dual-path feedback data flow from the decision execution layer, it achieves tight coupling and collaborative evolution of digital twin simulation and knowledge graph reasoning at the system level. This architecture ensures that the functional modules are no longer isolated subsystems, but rather interwoven into an organic whole through structured data paths. This guarantees that the load allocation decision-making process is always conducted under the dual protection of dynamic knowledge constraints and real-time physical feedback, thereby improving the system's decision-making accuracy, execution security, and adaptive evolution capabilities under complex actual working conditions.

[0171] Example 6: This embodiment provides a specific application example of a digital twin-based distribution box load allocation method in a commercial complex's distribution box during peak summer electricity consumption. This example aims to further illustrate the practical implementation of the abstract technical features described in the aforementioned embodiments under complex real-world environments through specific operating parameters and business processes.

[0172] It should be understood that this embodiment is only an illustrative application scenario demonstration and not a limitation on the scope of protection of this application. Other types of distribution boxes or power consumption scenarios can also be adapted with reference to the logic of this embodiment. Step S610: Obtain the current operating status data of the power distribution box in the commercial complex.

[0173] Among them, the current operating condition data refers to a multi-dimensional data set that characterizes the real-time operating status of the distribution box and its external environment under a specific application scenario. Its specific content is determined by the load characteristics of the power supply object.

[0174] In some implementations, the power supply range of a distribution box in a commercial complex covers the underground dining area, the ground floor shops, and the charging pile group in the underground parking lot.

[0175] During a typical evening peak period at 19:30 on a certain day in July during summer, the system collected the following specific data through intelligent sensors and environmental sensing units deployed on the incoming side of the distribution box and each branch circuit: the three-phase voltages of A / B / C were 221V, 219V, and 223V, respectively; the three-phase currents were 185A, 142A, and 210A, respectively; the neutral current was 68A; the active power was 135kW; the reactive power was 45kVar; the power factor was 0.95; the frequency was 50Hz; the terminal temperature was 62℃; the ambient temperature was 35℃; the relative humidity was 75%; and the current time period was marked "Summer workday 19:30".

[0176] It should be noted that in this commercial complex scenario, due to the access of catering equipment and charging piles, the load has significant nonlinear and time-varying characteristics. Therefore, the collection of operating condition data not only includes steady-state electrical quantities, but also must cover terminal temperatures that reflect the heat accumulation effect and ambient temperature and humidity that affect heat dissipation and load levels. In this way, these multi-dimensional data can be used together to form the real physical boundary input for knowledge reasoning and simulation.

[0177] Step S620: Perform constraint reasoning based on multi-level knowledge graph to generate constraint data.

[0178] Among them, constraint reasoning refers to the process of using a pre-constructed multi-level knowledge graph customized for the commercial complex scenario to perform multi-hop deductions starting from the current working condition node and following the implicit influence association and temporal evolution law path, thereby generating dynamic constraint information that is adapted to the current moment.

[0179] In some implementations, the multi-level knowledge graph basic entity layer in this scenario includes power distribution box equipment entities, A / B / C three-phase line entities, air conditioning unit entities, lighting circuit entities, charging pile group entities, and catering equipment entities.

[0180] The state parameter layer contains independent nodes for each phase current, neutral current, terminal temperature, ambient temperature and humidity, harmonic content, etc.

[0181] The specific path of its constraint reasoning includes: starting from the node with an ambient temperature of 35℃, inferring that the air conditioning cooling power has increased by 12% compared to the baseline value by following the positive implicit correlation between temperature and air conditioning load; then predicting the increment of phase A current by following the correlation between air conditioning load and phase A current.

[0182] Starting from the current time period of 19:30, and considering the implicit correlation between the time period and the charging power of the charging piles (weight 0.92), it can be inferred that the charging pile group is in the peak output period. Furthermore, due to historical wiring habits, phase C carries 70% of the charging pile load, so there is a risk of overload in phase C current.

[0183] Next, the super-edge correlation between the air conditioner compressor harmonics, neutral current and fundamental wave imbalance was activated, and it was deduced that when the air conditioner is running at full load due to the current high temperature of 35℃, the content of the third harmonic will have a superposition effect with the fundamental wave imbalance, so that the theoretical value of the neutral current increases by 15% compared with the simple arithmetic sum.

[0184] The final constraint dataset includes: the current carrying capacity of phase A cable needs to be reduced to 88% of the rated value due to temperature rise; the neutral current safety threshold needs to be dynamically corrected according to the real-time harmonic content; and the trend margin of a 20% expected increase in the charging pile load in the next 30 minutes.

[0185] It should be noted that this constraint reasoning based on scenario-based graphs can transform the multi-factor coupling effects of commercial complexes, such as temperature, air conditioning, harmonics, and neutral line, into quantified simulation boundaries, avoiding the problem of general rules failing under specific complex working conditions and ensuring the physical authenticity of the starting point for subsequent simulation deductions.

[0186] Step S630: Inject constraint data into the digital twin simulation model and perform parallel simulation calculations on the preset candidate load allocation schemes.

[0187] Among them, constraint injection and parallel simulation refer to mapping the dynamic constraints obtained by reasoning to the internal parameters of the simulation solver, and comparing and deduce multiple possible load adjustment strategies within the same time window, so as to evaluate the expected performance of each scheme under the current knowledge constraints.

[0188] In some implementations, the digital twin simulation model of the distribution box consists of three parts: The physical model constructs a node admittance matrix based on the actual electrical topology and cable specifications of the distribution box; The data model is mapped in real time to the operating data such as voltage, current, and temperature collected in step S610; The simulation model integrates sub-modules for three-phase power flow calculation, harmonic analysis, and temperature rise simulation.

[0189] The specific operation of constraint injection includes: dynamically adjusting the upper limit of current carrying capacity of phase A cable from the default 250A to 220A (i.e., injecting constraints within the theoretical value range).

[0190] Add a neutral current correction equation to the harmonic analysis submodule. The coupling coefficient The value is 1.15 (i.e., the constraint on the relationship between the injected parameters).

[0191] A 20% linear growth slope is superimposed on the charging pile load input end as the prediction boundary for the next 30 minutes (i.e., injecting time series trend constraints).

[0192] Based on this, the system can launch simulations of three candidate schemes in parallel: Scheme 1 is to maintain the status quo; Scheme 2 is to switch the No. 3 charging pile group from phase C to phase B; Scheme 3 is to switch the No. 5 catering circuit from phase A to phase C and distribute some charging piles to phase A.

[0193] Each scheme is calculated independently in the model after the constraints are injected, and the corresponding predicted results such as three-phase unbalance, maximum cable temperature, neutral current and harmonic distortion rate are output.

[0194] It should be noted that by injecting scenario-specific implicit constraints into the simulation kernel in advance, the parallel computing process can always run on a track that conforms to the actual physical laws of the commercial complex during the summer peak season. This effectively eliminates false feasible solutions that seem optimal under the ideal model but cause neutral line overload or cable overheating under high temperature and high harmonic environment.

[0195] Step S640: Based on the simulation results, trigger the multi-level knowledge graph to perform constraint reasoning, update the constraint data, and return to execute the injection and simulation steps until the preset convergence conditions are met.

[0196] Among them, alternating iteration to convergence refers to the process in which simulation calculation and knowledge-constrained reasoning cyclically feed back to each other on the time axis and gradually approach the real steady state. The termination criterion is that the output difference is less than the threshold or the maximum number of iterations is reached.

[0197] In some implementations, after the first round of simulation, the system will find that the current in phase B of scheme two increases significantly due to the connection of the charging pile, causing the predicted temperature of phase B terminals to reach 68°C. At this point, this new state is fed back to the knowledge graph and triggers secondary inference.

[0198] The second reasoning is that the graph will determine, based on the implicit correlation from terminal temperature to contact resistance to local temperature rise, that the increase in contact resistance at the B-phase connection point at 68℃ may lead to additional heat generation, and thus generate updated constraint data C2, reducing the heat dissipation coefficient of the B-phase temperature rise model by 5%.

[0199] C2 is then injected into the model for a second round of simulation, yielding a new prediction result P2. This process continues until the change in three-phase imbalance output in two consecutive iterations is less than 0.2% and the change in the highest temperature is less than 0.5℃, or until the maximum number of iterations (5) is reached, at which point the process is forcibly terminated.

[0200] In this scenario, the system converged after three rounds of iterations. The final predicted imbalance of Scheme 2 was corrected from the initial estimate of 4.8% to 5.3%, and the neutral line current was corrected from 62A to 67A.

[0201] It should be noted that this tightly coupled alternating iterative mechanism enables the simulation model to dynamically absorb the deep understanding of the thermal nonlinear characteristics of the equipment from the knowledge graph, avoiding the problem of underestimating safety risks due to ignoring the positive feedback effect of temperature and resistance in a single static simulation. This ensures that the converged simulation results truly reflect the steady-state behavior of the power distribution box in the commercial complex under high temperature and heavy load conditions.

[0202] Step S650: Based on the simulation results that meet the preset convergence conditions, determine the target load allocation scheme from the preset candidate load allocation schemes.

[0203] Among them, determining the target scheme refers to performing hierarchical compliance verification and multi-objective optimization on the final simulation results of each candidate scheme after iterative convergence, and making a decision output process based on the execution feedback-driven dual-path evolution of the system.

[0204] In some implementations, the system first performs hard constraint checks on the three converged schemes: Option 3 was directly rejected because the C-phase current exceeded the revised current carrying capacity limit of 220A after being superimposed with the charging pile load. Both Scheme 1 and Scheme 2 passed the safety hard constraint verification, including cable current carrying capacity, voltage deviation, and number of switching operations.

[0205] Subsequently, a soft constraint assessment was conducted: Although the imbalance of Scheme 2 (5.3%) was slightly higher than the improvement of Scheme 1 (6.1%), its overall line loss rate was less than 1.2%, and its neutral current of 67A was lower than that of Scheme 1 (72A), making it closer to the power quality optimization target. After weighted scoring, Scheme 2 scored the highest and was determined as the target load allocation scheme, namely, "switching the No. 3 charging pile group from phase C to phase B".

[0206] After the scheme was implemented, the system continuously monitored the actual operating data and found that the actual imbalance was 5.1%, which was only 0.2% different from the simulation prediction of 5.3%. This small deviation was used for dual-path updates: on the one hand, the impedance parameters of phase B in the twin model were fine-tuned to improve the accuracy of the next simulation; on the other hand, the confidence weight of the associated edge of the charging pile access to phase B temperature rise in the knowledge graph was enhanced.

[0207] It should be noted that this decision-making mechanism based on hierarchical verification of hard and soft constraints and dual-path feedback updates not only ensures that the output load distribution scheme is absolutely safe and relatively optimal under the extreme condition of summer peak in commercial complexes, but also enables the system to accumulate experience from each actual execution, continuously improving the decision-making accuracy and adaptability for similar complex scenarios.

[0208] Based on the above technical solution, this embodiment applies the load allocation method based on the co-evolution of digital twins and knowledge graphs to the summer peak electricity consumption scenario of the distribution box of a commercial complex, and verifies the effectiveness of the method in handling complex working conditions such as multi-type load coupling, strong environmental sensitivity and high incidence of hidden risks.

[0209] This method effectively utilizes scenario-based knowledge graphs to capture the combined effects of factors such as temperature, harmonics, and time of day on electrical parameters. Through constraint injection and alternating iteration mechanisms, these implicit patterns are deeply integrated into the entire simulation process, ensuring that the final load allocation scheme meets both the hard requirements for safe equipment operation and achieves soft optimizations in power quality and economic operation. Simultaneously, the execution feedback-driven dual-path update mechanism endows the system with the ability to continuously learn and evolve during long-term operation, improving the power supply reliability and operational intelligence of the commercial complex's power distribution system during extreme peak periods.

[0210] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A load distribution method for distribution boxes based on digital twins, characterized in that, include: Obtain the current operating status data of the distribution box; Based on a pre-constructed multi-level knowledge graph, constraint reasoning is performed on the current working condition data to generate constraint data. The multi-level knowledge graph includes state parameter nodes, implicit influence relationships between the state parameter nodes, and the temporal evolution rules of the state parameter nodes. The constraint data is injected into the digital twin simulation model, and parallel simulation calculations are performed on a preset number of candidate load allocation schemes to obtain the simulation results of each candidate load allocation scheme. Based on the simulation results, the multi-level knowledge graph is triggered to perform constraint reasoning, update the constraint data, and return to the step of injecting the constraint data into the digital twin simulation model until the preset convergence condition is met. Based on the simulation results that satisfy the preset convergence conditions, the target load allocation scheme is determined from the preset number of candidate load allocation schemes.

2. The load allocation method for distribution boxes based on digital twins according to claim 1, characterized in that, The multi-level knowledge graph includes a basic entity layer and a state parameter layer; The basic entity layer includes the distribution box equipment entity, the phase wire entity, and the electrical equipment entity; The state parameter layer includes electrical state parameter nodes and environmental state parameter nodes; The basic entity layer and the state parameter layer are connected through inter-layer mapping. The implicit influence relationships within the state parameter layer include pairwise influence implicit relationships and multi-factor mutual influence implicit relationships, wherein the multi-factor mutual influence implicit relationships are represented by hyperedges, and the hyperedges simultaneously connect three or more state parameter nodes.

3. The load allocation method for distribution boxes based on digital twins according to claim 2, characterized in that, The process of determining the implicit relationships of pairwise influence and the implicit relationships of mutual influence among multiple factors includes: Obtain the time-series data sequence of each state parameter node within a continuous time window; For any two state parameter nodes, calculate the dynamic time warping distance between their time-series data sequences, and determine the initial association strength of the pairwise influence implicit association based on the dynamic time warping distance. Granger causality test is performed on the time series data sequences of the two state parameter nodes. The influence direction of the pairwise influence hidden association is determined based on the test results. The pairwise influence hidden association is generated based on the initial association strength and the influence direction. For any parameter combination consisting of three or more state parameter nodes, the temporal data sequence of each state parameter node is used as input. The attention weights between nodes in the parameter combination are learned through a temporal graph attention network. Based on the attention weights, the hyperedge connection mode and joint influence weight of the hidden association of the mutual influence of multiple factors are determined.

4. The load allocation method for distribution boxes based on digital twins according to claim 3, characterized in that, The methods for representing the temporal evolution law include: The time-series data of each state parameter node is decomposed into trend components, periodic components, and random components. Based on the trend component and the periodic component, construct the state transition probability matrix of each state parameter node under different time periods and conditions; The correlation evolution law is obtained by modeling the change law of the strength of the correlation between the random components and the implicit influence over time.

5. The load allocation method for distribution boxes based on digital twins according to claim 1, characterized in that, The construction process of the digital twin simulation model includes: Obtain the electrical topology parameters, equipment parameters, and cable parameters of the distribution box, and construct the physical model of the distribution box; Obtain the real-time operating data of the distribution box, map the real-time operating data to each node of the physical model, and construct a data model that is synchronously updated with the physical model; Based on the physical model and the data model, a simulation model for parallel simulation calculation is constructed. The simulation model integrates a three-phase power flow calculation submodule, a harmonic analysis submodule, and a temperature rise simulation submodule. The simulation model takes the state parameters of each node of the distribution box as input and the three-phase voltage, three-phase current, active power, reactive power, harmonic content and node temperature of each node obtained by the parallel simulation calculation as output.

6. The load allocation method for distribution boxes based on digital twins according to claim 5, characterized in that, The constraint data includes theoretical value range constraints, parameter relationship constraints, and time series trend constraints. The process of injecting the constraint data into the digital twin simulation model includes: By adjusting the value boundaries of the corresponding parameters in the digital twin simulation model, the theoretical value range constraint is injected; By adding constraint equations or correction coefficients between parameters in the digital twin simulation model, the relationship constraints between the parameters are injected. The time-series trend constraint is injected by setting a margin for parameter changes in the prediction time domain for the digital twin simulation model.

7. The load allocation method for distribution boxes based on digital twins according to claim 6, characterized in that, The preset convergence conditions include: the output difference between two consecutive iterations is less than a preset threshold, or the maximum number of iterations is reached. The process of determining the target load allocation scheme from the preset number of candidate load allocation schemes includes: The simulation results that meet the preset convergence conditions are subjected to compliance checks of hard constraints and soft constraints. The hard constraints include equipment safety operation constraints, and the soft constraints include power quality optimization constraints. The candidate load allocation scheme that passes the compliance check is denoted as the target load allocation scheme.

8. The load allocation method for distribution boxes based on digital twins according to claim 1, characterized in that, After the target load allocation scheme is executed, actual operating data is obtained; Based on the deviation between the actual operating data and the simulation results, the model parameters of the digital twin simulation model and the association weights of the implicit influence associations in the multi-level knowledge graph are updated simultaneously.

9. A load distribution system for distribution boxes based on digital twins, characterized in that, The system, applied to the digital twin-based distribution box load allocation method according to any one of claims 1-9, comprises: The perception layer is used to acquire the current operating status data of the distribution box; The knowledge construction layer is used to perform constraint reasoning on the current working condition data based on a pre-built multi-level knowledge graph to generate constraint data. The multi-level knowledge graph includes state parameter nodes, implicit influence relationships between the state parameter nodes, and the temporal evolution rules of the state parameter nodes. The twin modeling layer is used to construct digital twin simulation models; The collaborative inference layer is used to inject the constraint data into the digital twin simulation model, perform parallel simulation calculations on a preset number of candidate load allocation schemes, and obtain the simulation results of each candidate load allocation scheme. Based on the simulation results, the multi-level knowledge graph is triggered to perform constraint reasoning, update the constraint data, and return to execute the step of injecting the constraint data into the digital twin simulation model until the preset convergence condition is met. The decision execution layer is used to determine the target load allocation scheme from the preset candidate load allocation schemes based on the simulation results that meet the preset convergence conditions.

10. The distribution box load allocation system based on digital twin according to claim 9, characterized in that, The collaborative inference layer is used to execute the alternating iterative data flow of simulation calculation of the digital twin simulation model and constraint reasoning of the multi-level knowledge graph; The decision execution layer is used to provide feedback data streams to the knowledge construction layer and the twin modeling layer.

Images