Park intelligent decision system and method based on digital twinning and knowledge graph

Abstract

The application discloses a park intelligent decision system and method based on digital twinning and a knowledge graph, and belongs to the technical field of smart park management. According to the scheme, an initial knowledge graph is constructed according to field static knowledge, real-time original data streams and state snapshots in digital twinning bodies from a park Internet of Things platform, a work order system and external data sources are acquired, standardized time sequence events are generated, and the standardized time sequence events are dynamically injected into the initial knowledge graph, so that the initial knowledge graph evolves into a dynamic knowledge graph; in response to a park abnormal event or a decision request, a root cause is analyzed based on the dynamic knowledge graph and a graph attention network, candidate strategies are generated based on the root cause, the candidate strategies are converted into simulation instructions, a digital twinning body is driven to perform simulation, recommended decisions are output according to simulation results, and decision execution effect data is injected into the dynamic knowledge graph as new time sequence events. The scheme realizes data-knowledge-model collaborative analysis, and improves the explainability and reliability of decisions.

Description

Technical Field

[0001] This invention relates to the field of smart park management technology, and in particular to a smart park decision-making system and method based on digital twins and knowledge graphs. Background Technology

[0002] With the development of IoT, big data, and AI technologies, smart park management is evolving from a traditional passive response model to a proactive intelligent decision-making model. However, existing park decision support systems still exhibit many insurmountable bottlenecks when facing complex and dynamic operating environments.

[0003] Current mainstream solutions fall into two main categories. The first category is data visualization platforms, which typically build 3D visualization dashboards based on Building Information Modeling (BIM) and Internet of Things (IoT) technologies. These dashboards can intuitively display information such as real-time equipment status, energy consumption distribution, or personnel heat maps. The advantage of this type of platform lies in the intuitiveness of information presentation, but its function is limited to describing the current situation. It cannot deeply explain the reasons behind the phenomena, let alone provide managers with actionable optimization suggestions.

[0004] The second category is automated systems based on rule engines or simple AI models. For example, a system might use pre-set logic such as "automatically activate the fresh air system when the CO2 concentration in a certain area exceeds 1000 ppm and the number of people exceeds 50." While such systems can achieve a certain degree of automated control, their decision-making logic relies heavily on pre-written rules and lacks the ability to understand complex, multi-factor causal chains. Once they encounter scenarios that are not pre-set, the system becomes helpless.

[0005] To overcome these limitations, academia and industry have begun exploring the introduction of knowledge graphs and digital twin technologies into park management. Knowledge graphs excel at structurally organizing and representing static knowledge within the park domain, such as equipment manuals, operation and maintenance procedures, historical failure cases, and policies and regulations, endowing the system with a certain degree of semantic understanding and reasoning capabilities. Digital twins, by constructing a virtual model synchronized with the physical park, can simulate and extrapolate future states under different control strategies, possessing the ability for forward-looking assessment.

[0006] However, in current technological practices, knowledge graphs and digital twins are mostly used in isolation, failing to form an effective collaborative closed loop. Specifically, knowledge graphs focus on maintaining static semantic relationships, and their knowledge base lacks dynamic linkage with real-time operational data streams from the physical world. Knowledge updates in knowledge graphs lag behind, making it difficult to automatically absorb real-time data streams reflecting the latest system state, resulting in a disconnect between "static knowledge" and "dynamic reality." Conversely, while digital twins can perform high-fidelity physical simulations, their internal decision-making processes often resemble a "black box," lacking interpretability and foresight. Managers struggle to understand the logical basis behind simulation results, thus affecting the trust and adoption rate of decisions. Summary of the Invention

[0007] The purpose of this invention is to overcome the above-mentioned defects of the prior art and provide a digital twin and knowledge graph-based intelligent decision-making system and method for parks.

[0008] To achieve the above objectives, this invention provides a smart decision-making system for industrial parks based on digital twins and knowledge graphs, comprising: The knowledge graph construction module is used to integrate static domain knowledge and build an initial knowledge graph. The digital twin module is used to build and run a digital twin that is synchronized with the physical entities of the park. The digital twin integrates an energy flow model, an equipment dynamic model, a building energy consumption model, and a carbon emission calculation model. The collaborative reasoning and decision generation module is configured to execute data knowledge transformation processes and knowledge-guided simulation processes; The data knowledge process includes: real-time parsing of the raw data streams and state snapshots in the digital twin from the park's IoT platform, work order system, and external data sources to generate standardized time-series events that conform to a predetermined semantic template; using the standardized time-series events as dynamic nodes and linking them to the corresponding entity nodes in the initial knowledge graph through preset relationship types to form a dynamic knowledge graph that reflects the evolution of the park's state. The knowledge-guided simulation process includes: responding to abnormal events or decision requests in the park, acquiring multi-dimensional time-series data related to the abnormal events or decision requests from the park's IoT platform, work order system, external data sources, and digital twin; in the dynamic knowledge graph, matching preset rules and performing path reasoning based on the multi-dimensional time-series data to form a preliminary causal chain; using a graph attention network to score the importance of event nodes and knowledge nodes in the preliminary causal chain, and outputting a root cause analysis report containing the quantitative contribution of each factor; based on the root cause analysis report, retrieving relevant strategy rules from the dynamic knowledge graph to generate at least one parameterized candidate strategy; converting the candidate strategy into a simulation instruction set to drive the digital twin module to perform simulation deduction; and based on the simulation results, outputting recommended decisions to guide park management operations. The system also includes a human-machine collaboration and feedback module, which is used to push the recommended decision to the park management terminal for manual confirmation or adjustment, and to inject the decision execution effect data as a new time-series event back into the dynamic knowledge graph.

[0009] In the intelligent decision-making system for parks based on digital twins and knowledge graphs of the present invention, the step of using a graph attention network to score the importance of event nodes and knowledge nodes in the preliminary causal chain includes: Using the event nodes and knowledge nodes in the preliminary causal chain as input, a subgraph is constructed, and a graph attention network is used to calculate the attention weight of each node in the subgraph. The attention weight is then used as the contribution of the node to the abnormal event or decision request.

[0010] In the intelligent decision-making system for parks based on digital twins and knowledge graphs of the present invention, the step of generating standardized time-series events conforming to a predetermined semantic template and linking the standardized time-series events as dynamic nodes to corresponding entity nodes in the initial knowledge graph through a preset relationship type includes: Extract the event subject, event type, attribute value, and timestamp from the original data stream or state snapshot; Create an event node that includes the event subject, event type, attribute value, and timestamp, and map the event node to the entity node in the initial knowledge graph that corresponds to the event subject.

[0011] In the intelligent decision-making system for parks based on digital twins and knowledge graphs of the present invention, the static knowledge includes entities and relationships extracted from equipment manuals, operation and maintenance manuals, and national standards. The entity types include equipment, energy, emission factors, and policies, and the relationship types include consumption, generation, controlled by, and compliance.

[0012] In the intelligent decision-making system for parks based on digital twins and knowledge graphs of the present invention, the collaborative reasoning and decision generation module is further configured as follows: A reinforcement learning algorithm is used, combined with a multi-objective reward function considering cost, carbon emissions, and risk, to evaluate the long-term returns of each candidate strategy. The evaluation results are then used as one of the bases for outputting recommendation decisions.

[0013] This invention also provides a smart decision-making method for industrial parks based on digital twins and knowledge graphs, comprising the following steps: S1: Construct an initial knowledge graph that integrates static knowledge of the park domain; construct a digital twin corresponding to the physical entities of the park; S2: Real-time acquisition of raw data streams and state snapshots from the digital twin from the park's IoT platform, work order system, and external data sources; generation of standardized time-series events; and dynamic injection of the standardized time-series events into the initial knowledge graph, so that it evolves into a dynamic knowledge graph. S3: In response to an abnormal event or decision request in the park, acquire multi-dimensional time-series data related to the abnormal event or decision request from the park's IoT platform, work order system, external data sources, and digital twin; in the dynamic knowledge graph, match preset rules based on the multi-dimensional time-series data and perform path reasoning to form a preliminary causal chain; use a graph attention network to score the importance of event nodes and knowledge nodes in the preliminary causal chain, and output a root cause analysis report containing the quantitative contribution of each factor; based on the root cause analysis report, retrieve relevant strategy rules from the dynamic knowledge graph to generate at least one parameterized candidate decision strategy; convert the candidate decision strategy into simulation instructions to drive the digital twin to perform simulation deduction; based on the simulation results, output recommended decisions to guide park management operations; S4: Push the recommended decision to the park management terminal for manual confirmation or adjustment, and inject the decision execution effect data as a new time-series event into the dynamic knowledge graph.

[0014] In the intelligent decision-making method for parks based on digital twins and knowledge graphs of the present invention, the step of using a graph attention network to score the importance of event nodes and knowledge nodes in the preliminary causal chain includes: Using the event nodes and knowledge nodes in the preliminary causal chain as input, a subgraph is constructed, and a graph attention network is used to calculate the attention weight of each node in the subgraph. The attention weight is then used as the contribution of the node to the abnormal event or decision request.

[0015] In the intelligent decision-making method for parks based on digital twins and knowledge graphs of the present invention, the step of generating standardized time-series events and dynamically injecting the standardized time-series events into the initial knowledge graph includes: Extract the event subject, event type, attribute value, and timestamp from the original data stream or state snapshot; Create an event node that includes the event subject, event type, attribute value, and timestamp, and map the event node to the entity node in the initial knowledge graph that corresponds to the event subject.

[0016] In the intelligent decision-making method for industrial parks based on digital twins and knowledge graphs of the present invention, the generation of candidate strategies specifically includes: Retrieve strategy rules that match the current decision-making context from the dynamic knowledge graph; Based on the retrieved policy rules, candidate policy objects containing specific action objects, action parameters, and execution sequence are generated by instantiation; The process of converting candidate strategies into a simulation instruction set includes: Map the action parameters in the candidate strategy object to the control variables of the corresponding virtual object in the digital twin; Future weather forecast data and planned maintenance schedules are extracted as boundary conditions for the simulation, and equipment operating limits are extracted as constraints for the simulation.

[0017] The intelligent decision-making method for industrial parks based on digital twins and knowledge graphs of the present invention further includes: A reinforcement learning algorithm is used, combined with a multi-objective reward function considering cost, carbon emissions, and risk, to evaluate the long-term benefits of each hypothetical scheme, and the evaluation results are used as one of the bases for outputting recommendation decisions.

[0018] The present invention constructs an initial knowledge graph based on domain static knowledge, acquires real-time raw data streams from the park's IoT platform, work order system, and external data sources, and generates state snapshots from the digital twin, generating standardized time-series events. These standardized time-series events are then dynamically injected into the initial knowledge graph, evolving it into a dynamic knowledge graph. In response to park anomalies or decision requests, root causes are analyzed based on the dynamic knowledge graph and graph attention network. Candidate strategies are generated based on these root causes, transformed into simulation instructions, and used to drive the digital twin to perform simulations. Recommended decisions are output based on the simulation results, and the decision execution effect data is injected into the dynamic knowledge graph as new time-series events. Compared with existing technologies, the present invention has the following beneficial effects: (1) Breaking down the barriers between knowledge, data, and models: Real-time event streams from heterogeneous data sources such as IoT platforms, work order systems, and external data sources are dynamically injected into the knowledge graph, keeping the static knowledge base synchronized with the dynamic physical world, truly reflecting the current state of the system, and ensuring the timeliness and comprehensiveness of decision-making basis. By transforming state snapshots in the digital twin into event nodes in the knowledge graph, and injecting the execution effect data of decisions back into the knowledge graph as new knowledge, the semantic depth of the knowledge graph is improved, realizing a closed loop of continuous evolution of data-knowledge-model collaboration, and improving the accuracy of decision-making.

[0019] (2) Interpretable root cause localization is achieved: By making the reasoning process explicit as traversal and association on the graph, and by calculating the attention weight of each node in the causal chain through the graph attention network, the abstract cause is transformed into a specific contribution value, so that each conclusion of the AI ​​decision has a clear semantic basis that conforms to human business logic, enabling managers to clearly understand the root driving factors of abnormal events and significantly improving the credibility of the decision.

[0020] (3) It provides forward-looking decision support: it generates candidate strategies through knowledge graphs, and generates simulation boundary conditions and constraints based on external data sources and knowledge graphs. It simulates and deduces candidate strategies in digital twins, and determines the final decision based on the simulation results, effectively avoiding decision risks. Attached Figure Description

[0021] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a schematic diagram of the structure of a smart decision-making system for a park based on digital twins and knowledge graphs, provided in an embodiment of the present invention.

[0022] Figure 2 A schematic diagram illustrating the steps of the intelligent decision-making method for parks based on digital twins and knowledge graphs provided in an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] The core of this invention lies in constructing a three-in-one collaborative analysis framework of "data-model-knowledge." It transforms real-time data streams and state snapshots from digital twins into events, injects them into a knowledge graph for semantic enhancement, and constructs a temporal dynamic knowledge graph. The rules, causal chains, and constraints in the knowledge graph guide the simulation parameter settings and scenario construction of the digital twin. The execution effect data of decisions output by the digital twin is ultimately converted into events and injected into the knowledge graph. This enables a deep understanding of the park's operational status, root cause localization of abnormal events, and multi-solution deduction and evaluation of management strategies, ultimately outputting interpretable, verifiable, and executable decision recommendations.

[0025] The embodiments of the present invention will now be described in further detail with reference to the accompanying drawings. It should be understood that the embodiments described herein are for illustrative and explanatory purposes only and are not intended to limit the scope of the invention.

[0026] like Figure 1 As shown, this embodiment of the invention provides a smart decision-making system for a park based on digital twins and knowledge graphs, including a knowledge graph construction module, a digital twin module, and a collaborative reasoning and decision generation module.

[0027] The knowledge graph construction module is used to integrate domain static knowledge and construct an initial knowledge graph. The static knowledge includes entities and relationships extracted from equipment manuals, operation and maintenance manuals, and national standards. Entity types include equipment, energy, emission factors, and policies. Relationship types include consumption, generation, controlled by, compliant, and carbon emission factor. An example triplet is <CCHP Unit #3, Consumption, Natural Gas> and <Natural Gas, Carbon Emission Factor, 2.16 kgCO2 / m³>. In this embodiment of the invention, the knowledge graph can be flexibly expanded according to changes in the park's business format. For example, when new charging piles are added or new policies are introduced, entities and relationships are extracted from the corresponding static knowledge and added to the existing knowledge graph.

[0028] The digital twin module is used to construct and run a digital twin synchronized with the physical entities of the park. This digital twin integrates an energy flow model, an equipment dynamics model, a building energy consumption model, and a carbon emission calculation model. Specifically, the energy flow model describes the transmission and conversion efficiency of electricity, cooling, and heat in the pipeline network; the equipment dynamics model outputs equipment dynamic curves, such as the start-up and shutdown energy consumption of CCHP and the charging and discharging efficiency curves of energy storage systems; the building energy consumption model simulates building load based on a thermal resistance-thermal capacity network; and the carbon emission calculation model calculates total emissions based on energy consumption multiplied by a real-time carbon emission factor (grid / gas).

[0029] In this embodiment of the invention, the digital twin module constructs a digital twin synchronized with the physical entities of the park through a high-fidelity digital twin engine to support multi-physics coupled simulation and strategy deduction. The digital twin module receives key state data from the IoT platform to calibrate its own state in real time, ensuring consistency between the virtual and real worlds. It also receives simulation instruction sets from the collaborative reasoning and decision generation module and executes high-fidelity dynamic simulations, outputting various park indicators for a future period. For example, upon receiving the instruction to "shut down a chiller," it simulates the system response for energy consumption, carbon emissions, room temperature, etc., over the next 24 hours, outputting key parameters and dynamic curves.

[0030] The collaborative reasoning and decision generation module is configured to execute a bottom-up data knowledge generation process and a top-down knowledge-guided simulation process. The data knowledge generation process includes: real-time parsing of raw data streams and state snapshots from digital twins from the park's IoT platform, work order system, and external data sources to generate standardized time-series events that conform to a predetermined semantic template; and using these standardized time-series events as dynamic nodes, linking them to corresponding entity nodes in the initial knowledge graph through preset relationship types to form a dynamic knowledge graph reflecting the evolution of the park's state.

[0031] In this embodiment of the invention, the external data source refers to external interfaces such as meteorological APIs closely related to park management. The digital twin continuously receives raw data from the physical world and uses its built-in physical model to calibrate, fuse, and estimate the state of this data, outputting a more accurate and complete system-level state than the original sensor data. The collaborative reasoning and decision generation module can periodically or under specific conditions extract high-fidelity state snapshots from the digital twin. The state snapshots can be parsed into a series of composite events. For example, the raw data may only show "meter reading increased," while the state snapshot in the digital twin may reveal "the air conditioning system of Building A has entered full-load operation due to the increase in outdoor temperature." Injecting these composite events as new dynamic nodes into the knowledge graph can greatly enrich the event granularity and semantic depth of the knowledge graph, enabling it to not only record what happened but also the complex state of the system.

[0032] In this embodiment of the invention, multiple data streams from various sources, including an IoT platform, a work order system, and a meteorological API, are accessed via an adapter mode. The raw data undergoes timestamp alignment, data cleaning, and protocol parsing, assigning high-precision timestamps and data source identifiers to all data points. Then, the time-series data is converted into standardized time-series events. Generating standardized time-series events that conform to a predetermined semantic template, and using these standardized time-series events as dynamic nodes linked to corresponding entity nodes in the initial knowledge graph through preset relationship types, includes extracting the event subject, event type, attribute values, and timestamp from the raw data stream or state snapshot. Specifically, an independent streaming event detection engine can be established for event detection and extraction. A series of predefined event templates cover device status, performance indicators, operational behaviors, environmental changes, etc. For example: Equipment anomaly events: {Device ID, Event Type: "Overload / Overtemperature / Offline", Severity Level, Current Value, Threshold, Timestamp}; Performance limit exceedance events: {Indicator Name: "Carbon Emission Intensity", Current Value, Threshold, Trend, Timestamp}; Operation command events: {Operation Object, Action: "Start / Stop / Adjust", Target Parameter, Operator / System, Timestamp}; Environmental change events: {Weather Station ID, Parameter: "Light Intensity / Wind Speed", Change Amount, Timestamp}. Event detection logic includes: Rule-based: Configuring business rules, such as IF Load Rate > 85% THEN generating a "High Equipment Load" event; Model-based: Using lightweight machine learning models such as Isolation Forest to perform online anomaly detection on real-time data streams and trigger anomaly events; Pattern recognition: Identifying specific data sequence patterns, such as sudden increases / decreases in power, and triggering corresponding events.

[0033] After the events are extracted, event nodes are created, each containing the event subject, event type, attribute value, and timestamp. These event nodes are then mapped to the entity nodes in the initial knowledge graph corresponding to the event subject. For example, after extracting event 1 from the raw data stream: "2026-02-01 14:00, CCHP unit #3 starts, load rate 85%", event node E1 is created and associated with the entity node CCHP unit #3 in the knowledge graph, with the corresponding timestamp annotated. After extracting event 2: "Daily photovoltaic output is 30% lower than predicted", event node E2 is created and associated with the two entity nodes photovoltaic array and weather mutation factor in the knowledge graph. The relationship between event nodes and entity nodes can be, for example: (event node) - [occurred at] -> (equipment entity node), (event node) - [associated indicator] -> (indicator entity node), (event node A) - [caused / preceded by] -> (event node B). Each event node has a timestamp attribute, allowing queries to retrieve the set of events for a specific entity within a given time period based on the timestamp, thus dynamically reconstructing a snapshot of the system's state and its evolutionary trajectory. As time-series events are continuously injected, the initial knowledge graph gradually evolves into a dynamic time-series knowledge graph. The edge weights of the graph can also be dynamically updated based on event frequency.

[0034] The knowledge-guided simulation process includes: responding to abnormal events or decision requests in the park, acquiring multi-dimensional time-series data related to the abnormal events or decision requests from the park's IoT platform, work order system, external data sources, and digital twin; in the dynamic knowledge graph, matching preset rules based on the multi-dimensional time-series data and performing path reasoning to form a preliminary causal chain; using a graph attention network to score the importance of event nodes and knowledge nodes in the preliminary causal chain, and outputting a root cause analysis report containing the quantitative contribution of each factor; based on the root cause analysis report, retrieving relevant strategy rules from the dynamic knowledge graph to generate at least one parameterized candidate strategy; converting the candidate strategy into a simulation instruction set to drive the digital twin module to perform simulation deduction; and based on the simulation results, outputting recommended decisions to guide park management operations.

[0035] The generation of candidate strategies specifically includes: retrieving strategy rules matching the current decision-making context from the dynamic knowledge graph; and, based on the retrieved strategy rules, instantiating candidate strategy objects containing specific action objects, action parameters, and execution timing. The conversion of candidate strategies into simulation instruction sets includes: mapping the action parameters in the candidate strategy objects to control variables of corresponding virtual objects in the digital twin; extracting future weather forecast data and planned maintenance schedules as boundary conditions for the simulation, and extracting equipment operating limits as constraints for the simulation.

[0036] In this embodiment of the invention, the step of using a graph attention network to score the importance of event nodes and knowledge nodes in the preliminary causal chain includes: taking the event nodes and knowledge nodes in the preliminary causal chain as input, constructing a subgraph, and using a graph attention network to calculate the attention weight of each node in the subgraph, and using the attention weight as the contribution of the node to the abnormal event or decision request.

[0037] In this embodiment of the invention, the goal of the knowledge-guided simulation process is to proactively drive the digital twin to conduct business-oriented, forward-looking simulations in response to abnormal events or decision requests within the park, utilizing domain knowledge from the knowledge graph. When the system detects an anomaly, such as a sudden increase in the park's total carbon emissions, or receives a decision request, such as "How to reduce next month's electricity bill?", it pulls multi-dimensional time-series data for the relevant time period from the IoT platform, work order system, external data sources, and the digital twin. In the dynamic knowledge graph, subgraph matching and path reasoning are performed based on the pulled data. For example, if it discovers that during the period of sudden carbon emission increase, there are simultaneously event nodes such as "decrease in photovoltaic output," "start of diesel generator," and "overtime work in a certain building," and these are connected into a preliminary causal chain through relationships in the knowledge graph, such as "photovoltaic panels - generation - green electricity," "green electricity - consumption - park," and "overtime work - increase - electricity load," the preliminary causal chain is extracted into an independent subgraph and input into a pre-trained graph attention network model. The graph attention network model calculates the attention weight of each node in the subgraph through a multi-layer message passing mechanism. The higher the weight, the greater the contribution of the node to the target anomaly. Finally, an interpretable root cause analysis report is generated, for example: "The main reasons for this carbon emission exceedance are: 1) a sudden drop in photovoltaic output (contribution 65%); 2) diesel generator startup (contribution 25%); 3) overtime production in Building A (contribution 10%)." Based on the root cause analysis report, candidate strategies are generated from the dynamic knowledge graph, driving the digital twin module to simulate these strategies. Specifically, candidate strategy rules are generated by matching the static strategy library and dynamic case library of the knowledge graph. Based on the retrieved strategy rules, candidate strategy objects containing specific action objects, action parameters, and execution timing are instantiated. An example of a strategy object is: {Strategy type: "Energy storage peak shaving", Action: Discharge, Power: 30kW, Start time: NOW+5min, Duration: 2h, Expected goal: Reduce carbon emissions}. Then, the action parameters in the candidate strategy objects are mapped to the control variables of the corresponding virtual objects in the digital twin; future weather forecast data and planned maintenance schedules are extracted as boundary conditions for the simulation, and equipment operating limits are extracted as constraints for the simulation. The digital twin module determines the simulation starting point, sets boundary conditions and control variables, and introduces random disturbances into the simulation based on historical uncertainty patterns recorded in the knowledge graph, and then begins to execute the simulation task. After the simulation is completed, the collaborative reasoning and decision generation module collects the simulation result data of each strategy, evaluates them, and outputs the final decision.

[0038] In some embodiments of the present invention, the collaborative reasoning and decision generation module is further configured to: employ a reinforcement learning algorithm, combined with a multi-objective reward function considering cost, carbon emissions, and risk, to evaluate the long-term benefits of each candidate strategy, and use the evaluation results as one of the bases for outputting a recommendation decision. The simulation results of a digital twin typically only represent the system response in the short term. Through reinforcement learning algorithms, the long-term benefits of each candidate strategy can be evaluated, and then the evaluation results can be combined with the simulation results to select the candidate strategy with the best overall performance in terms of both short-term response and long-term benefits.

[0039] In this embodiment of the invention, the intelligent decision-making system for the park based on digital twins and knowledge graphs further includes a human-machine collaboration and feedback module, which is used to push the recommended decision to the park management terminal for manual confirmation or adjustment, and to inject the decision execution effect data after confirmation or adjustment as a new time-series event back into the dynamic knowledge graph.

[0040] Managers can view root cause analysis reports, simulation results, and recommended decisions on the terminal, and then confirm, adjust, or reject them. Through the human-machine collaboration and feedback module, manual confirmation and adjustment of the system's recommended decisions can be achieved, improving decision reliability. Regardless of the manager's choice, the final execution effect data of that decision will be captured and, as a new standardized time-series event, injected back into the dynamic knowledge graph through a data knowledge process to optimize subsequent inference rules. If a decision is adopted and performs well, the entire "strategy-condition-result" chain can be structurally extracted as a new strategy rule and injected into the knowledge graph's strategy library. Similarly, if a decision's execution leads to equipment overload or a surge in energy consumption, this negative example can be injected into the knowledge graph as a taboo rule or risk warning to prevent future recurrence.

[0041] Through the aforementioned "bottom-up" data eventification and dynamic injection, and the "top-down" knowledge-guided simulation and closed-loop verification, the solution of this invention achieves bidirectional coupling between digital twins and knowledge graphs, forming a complete intelligent decision-making system with perception, analysis, decision-making, and learning capabilities. This is fundamentally different from the simple data transfer or model invocation between digital twins and knowledge graphs in existing technologies.

[0042] like Figure 2 As shown, the present invention also provides a smart decision-making method for industrial parks based on digital twins and knowledge graphs, including the following steps: S1: Construct an initial knowledge graph that integrates static knowledge of the park domain; construct a digital twin corresponding to the physical entities of the park.

[0043] S2: Execute the data knowledge process: acquire raw data streams and state snapshots from the digital twin in real time from the park's IoT platform, work order system and external data sources, generate standardized time-series events, and dynamically inject the standardized time-series events into the initial knowledge graph, so that it evolves into a dynamic knowledge graph.

[0044] In this embodiment of the invention, generating standardized time-series events and dynamically injecting them into the initial knowledge graph includes: extracting the event subject, event type, attribute value, and timestamp from the original data stream or state snapshot; creating event nodes containing the event subject, event type, attribute value, and timestamp; and mapping the event nodes to entity nodes in the initial knowledge graph corresponding to the event subject.

[0045] S3: In response to abnormal events or decision requests in the park, execute the knowledge-guided simulation process. Multidimensional time-series data related to the abnormal event or decision request is obtained from the park's IoT platform, work order system, external data sources, and digital twin. In the dynamic knowledge graph, preset rules are matched based on the multidimensional time-series data, and path reasoning is performed to form a preliminary causal chain. A graph attention network is used to score the importance of event nodes and knowledge nodes in the preliminary causal chain, outputting a root cause analysis report containing the quantified contribution of each factor. Based on the root cause analysis report, relevant strategy rules are retrieved from the dynamic knowledge graph to generate at least one parameterized candidate decision strategy. The candidate decision strategy is converted into simulation instructions to drive the digital twin to perform simulation deduction. Based on the simulation results, recommended decisions are output to guide park management operations.

[0046] In this embodiment of the invention, the step of using a graph attention network to score the importance of event nodes and knowledge nodes in the preliminary causal chain includes: taking the event nodes and knowledge nodes in the preliminary causal chain as input, constructing a subgraph, and using a graph attention network to calculate the attention weight of each node in the subgraph, and using the attention weight as the contribution of the node to the abnormal event or decision request.

[0047] In this embodiment of the invention, the generation of the candidate strategy specifically includes: retrieving strategy rules matching the current decision-making context from the dynamic knowledge graph; and, based on the retrieved strategy rules, instantiating and generating a candidate strategy object containing specific action objects, action parameters, and execution timing. The conversion of the candidate strategy into a simulation instruction set includes: mapping the action parameters in the candidate strategy object to control variables of the corresponding virtual objects in the digital twin. Future weather forecast data and planned maintenance schedules are extracted as boundary conditions for the simulation, and equipment operating limits are extracted as constraints for the simulation.

[0048] S4: Execute the human-machine collaboration and feedback closed-loop process: push the recommended decision to the park management terminal for manual confirmation or adjustment, and feed back the decision execution effect data after confirmation or adjustment as a new time-series event to step S2 to update the dynamic knowledge graph.

[0049] In some embodiments of the present invention, the intelligent decision-making method for parks based on digital twins and knowledge graphs further includes: using a reinforcement learning algorithm, combined with a multi-objective reward function of cost, carbon emissions, and risk, to evaluate the long-term benefits of each proposed scheme, and using the evaluation results as one of the bases for outputting recommendation decisions.

[0050] Example 1 In this embodiment, the intelligent decision-making method for parks according to the present invention is described in detail, taking the generation of abnormal carbon emissions and emission reduction strategies in parks as an example.

[0051] Step 1: Construct a dynamic knowledge graph for the park Static knowledge extraction: Extract entities and relationships from equipment manuals, maintenance manuals, and national standards (such as GB / T 32950-2016) to construct an ontology. Entity types: equipment (e.g., "CCHP Unit #3"), energy (e.g., "natural gas"), emission factors, policies (e.g., "Park Carbon Peaking Action Plan"); Relationship types: Consumption, Generation, Controlled by, Compliant with, Carbon emission factor is.

[0052] Example triplet: <CCHP Unit #3, Consumption, Natural Gas>, <Natural Gas, Carbon Emission Factor: 2.16 kg CO2 / m³>.

[0053] Dynamic event injection: Events are parsed from data streams such as IoT platforms, work order systems, meteorological APIs, and state snapshots in digital twins, and mapped to entity nodes in the knowledge graph. For example, for event 1: "2026-02-01 14:00, CCHP unit #3 starts up, load rate 85%", an event node E1 is created and associated with the CCHP unit #3 node, with a timestamp added. For event 2: "Daily photovoltaic output is 30% lower than predicted", an event node E2 is created and associated with the photovoltaic array and weather change factor nodes.

[0054] Step 2: Construct a high-fidelity digital twin Based on the park's BIM model, the following sub-models are integrated: Energy flow model: describes the transmission and conversion efficiency of electricity, cooling, and heat in the pipeline network; Equipment dynamic model: Outputs dynamic performance parameter curves of equipment, such as the start-up and shutdown energy consumption of CCHP and the charging and discharging efficiency curves of energy storage systems; Building energy consumption model: Simulating building load based on thermal resistance-thermal capacity (RC) network; Carbon emission calculation model: Total emissions are calculated based on energy consumption × real-time carbon emission factor (grid / gas).

[0055] The constructed digital twin can receive external commands and simulate the system response for the next 24 hours.

[0056] Step 3: Collaborative Reasoning and Decision Generation Scenario trigger: The system detected that "at 14:00 today, the carbon emission intensity in the park suddenly increased to 0.85kgCO2 / kWh, exceeding the threshold of 0.6kgCO2 / kWh".

[0057] Root cause localization (bottom-up reasoning): Extract key data for abnormal periods from IoT platforms, work order systems, meteorological APIs, and digital twins: CCHP operating at full capacity, photovoltaic output only 40% of the predicted value, and increased grid electricity purchases. Initiate subgraph matching and path reasoning in the knowledge graph: Matching rule: IF Device = CCHP AND Status = Running THEN Consumption = Natural Gas AND Production = CO2. Reasoning path: CCHP operation → Consumption of natural gas → Production of CO2; Photovoltaic output ↓ → Grid electricity purchase ↑ → Grid carbon emissions ↑. Use graph attention networks to score the importance of relevant event nodes and knowledge nodes, and output a root cause report: "Main causes of excessive carbon emissions: 1. CCHP forced to operate at full capacity due to insufficient photovoltaic output (contribution 62%); 2. Grid carbon emission factor increased on the same day (contribution 28%)." Strategy Generation and Derivation (Top-Down Guided): Retrieve relevant strategy rules from the knowledge graph: Rule 1: IF carbon emissions exceed limits AND PV is insufficient THEN, prioritize energy storage discharge; Rule 2: IF there is a large meeting, THEN, ensure air conditioning and delay power consumption for non-critical equipment. Generate candidate strategy A: Initiate 50kW energy storage discharge for 2 hours; and candidate strategy B: Initiate 50kW energy storage discharge and postpone data center backup tasks for 2 hours. Convert the candidate strategies into simulation instruction sets to drive the digital twin for simulation. Based on the historical statistics of PV output fluctuations recorded in the knowledge graph, add random perturbations to the simulation. The simulation results are as follows: Candidate strategy A: carbon emissions decrease to 0.72kgCO2 / kWh, but SOC decreases to 18% (below the safety threshold); Candidate strategy B: carbon emissions decrease to 0.58kgCO2 / kWh, SOC remains at 25%, and room temperature fluctuation <0.5℃. From the simulation results, the SOC of candidate strategy B meets the safety threshold requirement and is better than candidate strategy A. The long-term benefits of each option were evaluated using a reinforcement learning algorithm. The results showed that candidate strategy B achieved a better balance between carbon reduction effect and system risk. Candidate strategy B was ultimately recommended, and an interpretable report was generated: "Recommendation: 1. Start energy storage discharge of 30kW; 2. Delay the data center's nighttime backup task by 2 hours. It is expected to reduce carbon emissions by 15% without affecting critical business operations." Step 4: Human-Machine Collaboration and Feedback Loop Decision suggestions are pushed to the park management app, supporting manual adjustments, such as the option to force the retention of data center backup tasks. Execution results confirmed or modified by the administrator are injected into the knowledge graph as new events to optimize subsequent inference rules. Assuming the administrator adopts candidate strategy B, and carbon emissions drop to 0.58 kgCO2 / kWh after 2 hours, the system automatically generates a feedback event: {Type: Strategy execution result, Associated strategy: Strategy B_ID, Expected effect: Carbon emissions ≤ 0.60, Actual effect: 0.58, Status: Success}. This event is injected into the knowledge graph and associated with the generation node and simulation prediction node of candidate strategy B.

[0058] After pilot testing in actual industrial parks, the solution implemented in this invention improved emergency response efficiency by 40%, increased management decision adoption rate by 65%, and shortened the average handling time for carbon emission anomalies to less than 15 minutes.

[0059] The above are merely specific embodiments of the present invention and should not be construed as limiting the scope of the present invention. Equivalent variations made by those skilled in the art based on this invention, as well as changes well-known to those skilled in the art, should still fall within the scope of the present invention.

Claims

1. A smart decision-making system for industrial parks based on digital twins and knowledge graphs, characterized in that: include: The knowledge graph construction module is used to integrate static domain knowledge and build an initial knowledge graph. The digital twin module is used to build and run a digital twin that is synchronized with the physical entities of the park. The digital twin integrates an energy flow model, an equipment dynamic model, a building energy consumption model, and a carbon emission calculation model. The collaborative reasoning and decision generation module is configured to execute data knowledge transformation processes and knowledge-guided simulation processes; The data knowledge process includes: real-time parsing of the raw data streams and state snapshots in the digital twin from the park's IoT platform, work order system, and external data sources to generate standardized time-series events that conform to a predetermined semantic template; using the standardized time-series events as dynamic nodes and linking them to the corresponding entity nodes in the initial knowledge graph through preset relationship types to form a dynamic knowledge graph that reflects the evolution of the park's state. The knowledge-guided simulation process includes: responding to abnormal events or decision requests in the park, acquiring multi-dimensional time-series data related to the abnormal events or decision requests from the park's IoT platform, work order system, external data sources, and digital twin; in the dynamic knowledge graph, matching preset rules and performing path reasoning based on the multi-dimensional time-series data to form a preliminary causal chain; using a graph attention network to score the importance of event nodes and knowledge nodes in the preliminary causal chain, and outputting a root cause analysis report containing the quantitative contribution of each factor; based on the root cause analysis report, retrieving relevant strategy rules from the dynamic knowledge graph to generate at least one parameterized candidate strategy; converting the candidate strategy into a simulation instruction set to drive the digital twin module to perform simulation deduction; and based on the simulation results, outputting recommended decisions to guide park management operations. The system also includes a human-machine collaboration and feedback module, which is used to push the recommended decision to the park management terminal for manual confirmation or adjustment, and to inject the decision execution effect data as a new time-series event back into the dynamic knowledge graph.

2. The intelligent decision-making system for industrial parks based on digital twins and knowledge graphs according to claim 1, characterized in that, The method of using a graph attention network to score the importance of event nodes and knowledge nodes in the preliminary causal chain includes: Using the event nodes and knowledge nodes in the preliminary causal chain as input, a subgraph is constructed, and a graph attention network is used to calculate the attention weight of each node in the subgraph. The attention weight is then used as the contribution of the node to the abnormal event or decision request.

3. The intelligent decision-making system for industrial parks based on digital twins and knowledge graphs according to claim 1, characterized in that, The generation of standardized time-series events conforming to a predetermined semantic template; Using the standardized time-series events as dynamic nodes, and linking them to corresponding entity nodes in the initial knowledge graph through preset relationship types includes: Extract the event subject, event type, attribute value, and timestamp from the original data stream or state snapshot; Create an event node that includes the event subject, event type, attribute value, and timestamp, and map the event node to the entity node in the initial knowledge graph that corresponds to the event subject.

4. The intelligent decision-making system for industrial parks based on digital twins and knowledge graphs according to claim 1, characterized in that, The static knowledge includes entities and relationships extracted from equipment manuals, operation and maintenance manuals, and national standards. The entity types include equipment, energy, emission factors, and policies, and the relationship types include consumption, generation, control by, and compliance.

5. The intelligent decision-making system for industrial parks based on digital twins and knowledge graphs according to claim 1, characterized in that, The collaborative reasoning and decision generation module is also configured as follows: A reinforcement learning algorithm is used, combined with a multi-objective reward function considering cost, carbon emissions, and risk, to evaluate the long-term returns of each candidate strategy. The evaluation results are then used as one of the bases for outputting recommendation decisions.

6. A smart decision-making method for industrial parks based on digital twins and knowledge graphs, characterized in that, Includes the following steps: S1: Construct an initial knowledge graph that integrates static knowledge of the park domain; construct a digital twin corresponding to the physical entities of the park; S2: Real-time acquisition of raw data streams and state snapshots from the digital twin from the park's IoT platform, work order system, and external data sources; generation of standardized time-series events; and dynamic injection of the standardized time-series events into the initial knowledge graph, so that it evolves into a dynamic knowledge graph. S3: In response to an abnormal event or decision request in the park, obtain multi-dimensional time-series data related to the abnormal event or decision request from the park's IoT platform, work order system, external data source and digital twin; In the dynamic knowledge graph, a preliminary causal chain is formed by matching preset rules with the multi-dimensional time-series data and performing path reasoning. A graph attention network is then used to score the importance of event nodes and knowledge nodes in the preliminary causal chain, outputting a root cause analysis report containing the quantified contribution of each factor. Based on the root cause analysis report, relevant strategy rules are retrieved from the dynamic knowledge graph to generate at least one parameterized candidate decision strategy. The candidate decision strategy is then converted into simulation instructions to drive the digital twin to perform simulation deduction. Based on the simulation results, recommended decisions are output to guide park management operations. S4: Push the recommended decision to the park management terminal for manual confirmation or adjustment, and inject the decision execution effect data as a new time-series event into the dynamic knowledge graph.

7. The intelligent decision-making method for industrial parks based on digital twins and knowledge graphs according to claim 6, characterized in that, The method of using a graph attention network to score the importance of event nodes and knowledge nodes in the preliminary causal chain includes: Using the event nodes and knowledge nodes in the preliminary causal chain as input, a subgraph is constructed, and a graph attention network is used to calculate the attention weight of each node in the subgraph. The attention weight is then used as the contribution of the node to the abnormal event or decision request.

8. The intelligent decision-making method for industrial parks based on digital twins and knowledge graphs according to claim 6, characterized in that, The step of generating standardized time-series events and dynamically injecting the standardized time-series events into the initial knowledge graph includes: Extract the event subject, event type, attribute value, and timestamp from the original data stream or state snapshot; Create an event node that includes the event subject, event type, attribute value, and timestamp, and map the event node to the entity node in the initial knowledge graph that corresponds to the event subject.

9. The intelligent decision-making method for industrial parks based on digital twins and knowledge graphs according to claim 6, characterized in that, The generation of the candidate strategy specifically includes: Retrieve strategy rules that match the current decision-making context from the dynamic knowledge graph; Based on the retrieved policy rules, candidate policy objects containing specific action objects, action parameters, and execution sequence are generated by instantiation; The process of converting candidate strategies into a simulation instruction set includes: Map the action parameters in the candidate strategy object to the control variables of the corresponding virtual object in the digital twin; Future weather forecast data and planned maintenance schedules are extracted as boundary conditions for the simulation, and equipment operating limits are extracted as constraints for the simulation.

10. The intelligent decision-making method for industrial parks based on digital twins and knowledge graphs according to claim 6, characterized in that, The method further includes: A reinforcement learning algorithm is used, combined with a multi-objective reward function considering cost, carbon emissions, and risk, to evaluate the long-term benefits of each hypothetical scheme, and the evaluation results are used as one of the bases for outputting recommendation decisions.

Images