Multi-objective optimization decision method and system for zero-carbon reconstruction scheme of industrial building

Abstract

The application relates to the technical field of architectural design, in particular to a multi-target optimization decision-making method and system for an industrial building zero-carbon reconstruction scheme. The method aims to solve the problem that the existing reconstruction scheme only focuses on operation carbon and ignores hidden carbon and the problem of static isolation in decision-making. The method comprises the following steps: generating a digital resource passport of waste materials based on a building information model; accessing regional market data to construct a dynamic resource-demand graph; calling a multi-modal graph matching algorithm to match the optimal resource supply subgraph for the project; using a pre-trained multi-agent reinforcement learning optimization engine to generate a dynamic resource decision sequence; and finally synthesizing a zero-carbon reconstruction optimization scheme containing a technical measure list and a resource circulation action plan table.

Description

Technical Field

[0001] This application relates to the field of architectural design technology, and in particular to a multi-objective optimization decision-making method and system for zero-carbon retrofitting schemes of industrial buildings. Background Technology

[0002] Industrial buildings are a key area of ​​energy consumption and carbon emissions, and their greening and zero-carbon transformation is crucial to promoting the low-carbon transformation of the industrial sector. Existing decision support methods for industrial building renovation mainly focus on the comparison of energy-saving technologies and economic benefit analysis during the operation phase. Their optimization targets are mostly economic indicators such as initial investment and operating energy consumption costs, or "operational carbon emissions" based on operating energy consumption conversion.

[0003] With the popularization of the concept of life cycle assessment, the scope of carbon emission accounting in the construction sector has expanded from a single operational phase to encompass the entire process of building material production, transportation, construction, demolition, and waste disposal. This means that retrofitting schemes that only optimize "operational carbon" while ignoring "hidden carbon" may lead to an increase in total life cycle carbon emissions instead of a decrease. It is understood that some studies have begun to attempt to incorporate life cycle carbon emissions into building retrofitting assessment models, but these methods generally have the following limitations: First, the treatment of old building material resources is simplistic, often treating them as waste to calculate disposal costs and carbon emissions, without actively optimizing them as tradable "negative carbon resources"; second, the optimization models are mostly static and isolated single-project optimizations, unable to reflect the dynamic changes in regional secondary building material market supply and demand, prices, and logistics costs, and also failing to consider potential resource synergy opportunities between multiple concurrent retrofitting projects; third, the decision-making process has limited intelligence, making it difficult to generate dynamic, adaptive, and optimally comprehensive retrofitting action plans under complex market uncertainties, multi-project game dynamics, and sequential decision-making requirements.

[0004] Therefore, there is an urgent need for an optimization method and system for zero-carbon retrofitting of industrial buildings that can uniformly calculate and proactively optimize carbon emissions throughout the entire life cycle, deeply integrate into the regional circular economy market, and possess intelligent collaborative decision-making capabilities. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies that only focus on operational carbon, ignore the synergy of the circular economy, and employ static and isolated decision-making. It provides a multi-objective optimization decision-making method and system for zero-carbon retrofitting schemes of industrial buildings. This method aims to achieve unified quantification and proactive optimization of carbon emissions (including implicit and operational carbon) throughout the entire lifecycle of a retrofitting project. By constructing a dynamic regional resource map and coupling it with intelligent algorithms, it generates dynamically optimized retrofitting schemes for projects that satisfy both technical and economic constraints and maximize regional carbon reduction and resource recycling benefits. To achieve the above objectives, the embodiments of this application disclose the following technical solutions:

[0006] In a first aspect, embodiments of this application provide a multi-objective optimization decision-making method for zero-carbon retrofitting schemes of industrial buildings, the method comprising:

[0007] Obtain the building information model data and the list of components to be demolished for the target industrial building to be renovated. Based on the building information model data and the list of components to be demolished, generate a digital waste material resource passport for the target industrial building to be renovated. The digital waste material resource passport includes the material type, geometric dimensions, mass estimate and recyclability level characteristics of the components to be demolished.

[0008] The system accesses data interfaces from regional secondary building materials trading platforms, logistics information platforms, and resource demand data streams from one or more concurrent renovation projects. Based on the accessed data from these platforms, the system constructs and continuously updates a regional resource-demand graph. In this graph, nodes represent resource suppliers, resource demanders, or logistics service providers, and edges represent potential supply and demand relationships or transportation routes between nodes. Each node and edge is associated with a multimodal attribute set that includes material properties, spatiotemporal location, unit carbon footprint vector, and price.

[0009] In response to the renovation needs of the target industrial building, the target industrial building is added as a new demand-side node to the regional resource-demand graph, and the new demand-side node is assigned the corresponding multimodal attribute set based on the digital waste material resource passport.

[0010] The multimodal graph matching algorithm is invoked, and the newly added demand-side node is used as the query center to search for the optimal subgraph in the regional resource-demand graph. The multimodal graph matching algorithm uses material performance matching degree, transportation carbon emissions, supply time window and comprehensive cost as joint constraints to output one or more resource supply subgraphs. The resource supply subgraphs include resource supply-side nodes that can meet the transformation requirements, logistics path edges and their associated multimodal attribute sets.

[0011] The target industrial building to be transformed and its corresponding transformation requirements are instantiated as an agent. The one or more resource supply subgraphs, the digital waste material resource passport, and the ontological technical parameters of the target industrial building to be transformed are jointly input into a pre-trained multi-agent reinforcement learning optimization engine. The multi-agent reinforcement learning optimization engine runs the policy network of the agent to generate a dynamic resource decision sequence. The dynamic resource decision sequence arranges a series of action instructions in chronological order. The types of action instructions include purchasing new materials, purchasing specific recycled materials in the resource supply subgraph, and selling old materials in the digital waste material resource passport to specific nodes in the regional resource-demand graph.

[0012] Based on each action instruction in the dynamic resource decision sequence and its associated multimodal attribute set, the estimated total carbon emissions and estimated renovation cost of the target industrial building within a preset life cycle time range are calculated.

[0013] By integrating the dynamic resource decision sequence, the estimated total carbon emissions, and the estimated retrofit cost, a zero-carbon retrofit optimization plan for the target industrial building is generated. The zero-carbon retrofit optimization plan includes a list of technical measures and a resource flow action plan.

[0014] Optionally, generating the digital waste material resource passport for the target industrial building to be renovated specifically includes:

[0015] The component to be demolished was subjected to a three-dimensional laser scan to obtain point cloud data;

[0016] Based on the point cloud data, component segmentation and material identification are performed to obtain the material type and geometric dimensions of each component;

[0017] By combining the material type and the geometric dimensions, and based on a preset material density database, an estimated mass value is obtained;

[0018] Based on the material type and historical service environment data, the recyclability level characteristic data are predicted by a pre-trained damage assessment model.

[0019] Optionally, the invocation of the multimodal graph matching algorithm, using the newly added demand-side node as the query center, to perform an optimal subgraph search in the regional resource-demand graph specifically includes:

[0020] Extract the multimodal attribute set of the newly added demand-side node and encode it into a high-dimensional query vector;

[0021] The multimodal attribute sets of all resource supplier nodes in the regional resource-demand graph are extracted in parallel and encoded into high-dimensional candidate vectors respectively;

[0022] Calculate the multi-dimensional similarity between the high-dimensional query vector and each of the high-dimensional candidate vectors. The multi-dimensional similarity is a weighted comprehensive value of material property similarity, spatiotemporal distance similarity, carbon footprint difference, and economic cost similarity.

[0023] Resource supplier nodes are sorted according to the multi-dimensional similarity and feasibility filtering is performed based on the joint constraints.

[0024] From the nodes that pass the feasibility filter, select the smallest set of nodes that meet the transformation requirements, and add the shortest logistics path edges connecting these nodes to form the resource supply subgraph.

[0025] Optionally, the pre-trained multi-agent reinforcement learning optimization engine is trained through the following steps:

[0026] Initializing the multi-agent reinforcement learning optimization engine includes constructing a simulation environment comprising multiple agents, wherein each agent simulates a virtual modification project with randomly generated ontology technical parameters and modification requirements, and the state space of the simulation environment is a dynamic summary of the regional resource-demand graph;

[0027] Define the action space for each agent, which includes action instruction types;

[0028] Define a reward function for each agent, which is a linear weighted sum of the carbon reduction benefits and economic benefits of the virtual transformation project simulated by the agent over its entire life cycle;

[0029] The multiple agents are instructed to interact in multiple rounds in the simulated environment. The policy network parameters of each agent are updated using a policy gradient algorithm until the cumulative reward of all agents converges, thus obtaining the pre-trained multi-agent reinforcement learning optimization engine.

[0030] Optionally, after generating the zero-carbon retrofit optimization scheme for the target industrial building, the method further includes:

[0031] Monitor the updates of data from the secondary building materials trading platform and the logistics information platform in the aforementioned region;

[0032] In response to the detection that key data updates exceed a preset threshold, the multi-agent reinforcement learning optimization engine is triggered to re-evaluate and adjust the dynamic resource decision sequence and generate scheme adjustment suggestions.

[0033] Optionally, the method further includes:

[0034] Obtain zero-carbon retrofit optimization schemes for all concurrent retrofit projects in the aforementioned regional resource-demand map;

[0035] By aggregating the resource flow action plans in all the zero-carbon transformation optimization schemes, simulation calculations are performed on the material flow and carbon emission changes of all resource nodes in the regional resource-demand map within a preset future time period.

[0036] Based on the changes in material flow and carbon emissions, regional resource bottlenecks and carbon emission hotspots are identified, and regional collaborative optimization suggestions are output.

[0037] Secondly, embodiments of this application provide a multi-objective optimization decision-making system for zero-carbon retrofitting schemes of industrial buildings, the system comprising:

[0038] The data acquisition and passport generation module is used to acquire the building information model data and the list of components to be demolished for the target industrial building to be renovated, and to generate a digital waste material resource passport for the target industrial building to be renovated based on the building information model data and the list of components to be demolished.

[0039] The dynamic map construction and maintenance module is used to connect to external data sources and build and update regional resource-demand maps.

[0040] The graph matching decision support module is used to call the multimodal graph matching algorithm to search for resource supply subgraphs in the resource-demand graph of the region for newly added transformation requirements;

[0041] The multi-agent optimization decision module is used to run a pre-trained multi-agent reinforcement learning optimization engine to generate a dynamic resource decision sequence based on the resource supply subgraph and the digital waste material resource passport.

[0042] The life cycle assessment module is used to calculate the estimated total carbon emissions and estimated retrofit costs based on the dynamic resource decision sequence.

[0043] The scheme synthesis and output module is used to integrate the dynamic resource decision sequence, the estimated total carbon emissions, and the estimated retrofit cost to generate and output a zero-carbon retrofit optimization scheme.

[0044] Optionally, the multi-agent optimization decision-making module includes a policy network submodule and an action sequence generation submodule;

[0045] The policy network submodule stores neural network parameters obtained through offline training. The neural network is used to output the probability distribution of each selectable action given the environmental state and the agent's own state.

[0046] The action sequence generation submodule is used to generate the dynamic resource decision sequence covering the entire transformation cycle by using a sequence decision algorithm based on the probability distribution output by the strategy network submodule.

[0047] Optionally, the system further includes a scheme dynamic adjustment module, which is used to monitor the changes in the map data in the dynamic map construction and maintenance module in real time, and when it detects that there is a significant change in the key data related to the currently generated zero-carbon transformation optimization scheme, it automatically initiates a re-decision request to the multi-agent optimization decision module and receives the new decision results to update the zero-carbon transformation optimization scheme.

[0048] Optionally, the system further includes a regional collaborative analysis module, which is used to periodically obtain all generated zero-carbon transformation optimization schemes in the region from the scheme synthesis and output module, perform data aggregation and conflict detection, and output intervention strategy suggestions to the system administrator for balancing global resources and carbon emissions based on simulation prediction of regional resource supply and demand trends.

[0049] The beneficial effects of this application are as follows:

[0050] (1) By creating a “digital waste material resource passport” and linking it with full life cycle carbon footprint data, proactive optimization of carbon emissions throughout the entire life cycle was achieved;

[0051] (2) By constructing a dynamic “regional resource-demand map”, the decision-making of a single renovation project is placed in the real-time changing regional resource market environment, which breaks through the limitations of static optimization of a single project;

[0052] (3) The combination of "multimodal graph matching algorithm" and "multi-agent reinforcement learning optimization engine" provides intelligent dynamic decision-making capabilities;

[0053] (4) The final output plan not only includes technical measures, but also includes a detailed resource circulation action plan, which clarifies the object, time, path and carbon cost of each resource transaction, thus enhancing the implementation and feasibility of the plan. Attached Figure Description

[0054] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0055] Figure 1 This is an architectural block diagram of the multi-objective optimization decision-making system for zero-carbon retrofitting of industrial buildings provided in this embodiment of the invention;

[0056] Figure 2 This is a flowchart of the multi-objective optimization decision-making method for zero-carbon retrofitting of industrial buildings provided in this embodiment of the invention;

[0057] Figure 3 This is a schematic diagram illustrating the principle of the collaborative operation of the multimodal graph matching algorithm and the multi-agent reinforcement learning optimization engine in this embodiment of the invention. Detailed Implementation

[0058] Specific embodiments of the invention will now be described in detail. Although the invention is described in conjunction with these specific embodiments, it should be understood that it is not intended to limit the invention to these specific embodiments. Rather, these embodiments are intended to cover alternative, modified, or equivalent embodiments that may be included within the spirit and scope of the invention as defined by the claims. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. The invention may be practiced without some or all of these specific details.

[0059] When used in conjunction with the terms "comprising," "method comprising," or similar language in this specification and appended claims, the singular forms "a," "some," and "the" include plural references unless the context clearly indicates otherwise. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0060] The following is in conjunction with the appendix Figure 1-3 The preferred embodiments of the present invention will be described in detail so that those skilled in the art can implement the present invention accordingly.

[0061] A multi-objective optimization decision-making method for zero-carbon retrofitting of industrial buildings, the method comprising:

[0062] S1. Obtain the building information model data and the list of components to be demolished for the target industrial building to be renovated. Based on the building information model data and the list of components to be demolished, generate a digital waste material resource passport for the target industrial building to be renovated. The digital waste material resource passport includes the material type, geometric dimensions, mass estimate and recyclability level characteristics of the components to be demolished.

[0063] The process of generating a digital waste material resource passport for the target industrial building renovation includes: performing 3D laser scanning on the components to be demolished to obtain point cloud data; segmenting and identifying the components based on the point cloud data to obtain the material type and geometric dimensions of each component; combining the material type and geometric dimensions with a pre-set material density database to estimate the mass value; and predicting the recyclability level characteristics based on the material type and historical service environment data through a pre-trained damage assessment model.

[0064] It should be noted that "Building Information Modeling (BIM) data" is a digital model containing the three-dimensional geometric information, component types, and attribute parameters of the target industrial building, typically existing in standard formats such as IFC (Industry Foundation Classes). The "List of Components to be Demolished" is a list of identified components (such as steel beams, concrete slabs, doors, and windows) planned for demolition, extracted from this model or renovation design. The "Preset Material Density Database" is a static lookup table storing the average density values ​​of various building materials. For example, referring to the "Code for Design of Building Structures" or publicly available engineering material handbooks, the density of steel can be set to 7850 kg / m³, ordinary concrete to 2400 kg / m³, and wood to be between 400-750 kg / m³ depending on the tree species.

[0065] Specifically, a ground-based or airborne 3D laser scanner is used to scan the area to be demolished. The scanner emits a laser beam and receives the reflected signal, obtaining the spatial 3D coordinates of a large number of points on the object's surface by calculating the time difference or phase difference. Includes reflectivity and color Information, collectively forming "point cloud data," is collected. After scanning, preprocessing such as point cloud registration and denoising is performed to create complete 3D site data. Next, the point sets belonging to individual physical components are separated from the chaotic point cloud data (segmentation), and the material type of each component is determined (identification). Component segmentation can employ methods based on geometric features (such as region growing and clustering algorithms) or deep learning (such as PointNet++ and 3D convolutional neural networks). For example, walls and floors are first segmented using planar detection algorithms, and then beams and columns are separated using edge detection and clustering. Material identification can combine multi-source information, including reflectance intensity and color; geometric features; and semantic association with the BIM model: the segmented point cloud is matched with the original BIM model to directly inherit the material properties of components in the BIM model. By combining one or more of these methods, the bounding box, geometric dimensions, and inferred material type (e.g., "Q345B H-beam 600*300*12*20") of each component can be output. After obtaining the component's material type and geometry, its mass is estimated by calculating its volume and multiplying it by the material's density.

[0066] The "pre-trained residual assessment model" is a machine learning model (such as Gradient Boosting Decision Tree (GBDT), Support Vector Machine (SVM), or a simple neural network). Its construction and workflow are as follows:

[0067] The model’s input features include material type, historical service environment data (such as the corrosion level of the environment in which the component is located, the number of years it has been in service, the highest temperature it has experienced, etc., which can be obtained from building operation and maintenance records, BIM models or by inferring historical data from environmental sensors), and optional input point cloud derived features (such as the roughness of the local surface of the point cloud, the depth of the depression, etc., which can be used as indirect indicators of corrosion or damage).

[0068] This model requires supervised learning using a labeled dataset. The training data comes from a large number of known samples of waste building materials. Each sample contains the aforementioned feature inputs and a true "recyclability level" label determined through manual inspection or non-destructive testing (such as ultrasonic or magnetic particle testing). It is understood that recyclability levels can be defined in multiple categories, such as: L1 (intact, can be directly reused), L2 (slightly damaged, requires repair before use), L3 (severely damaged, can only be recycled as raw material), and L4 (worthless, requires disposal).

[0069] By inputting the features of the target component into a pre-trained model, the model will output the probability of it belonging to each recycling class, or directly output the most likely class label. For example, for a steel beam that has been in service for 20 years and is in a typical indoor environment, the model may predict that its recycling class is L1 with a 70% probability and L2 with a 30% probability.

[0070] Understandably, this solution provides precise input for calculating the "existing implicit carbon" contained in this batch of old building materials and the "carbon offsetting benefits" (avoiding carbon emissions from the production of new materials) that their reuse may bring, through accurate material identification and quality estimation. By introducing recyclability grade characteristic data, waste building materials are transformed from a "cost burden" into a "value asset." Automated scanning and intelligent identification prediction enable rapid, non-contact acquisition of large-area building information, and the use of consistent data-driven models for evaluation significantly improves data acquisition efficiency and the consistency and traceability of evaluation results.

[0071] S2. Access the data interfaces of the regional secondary building materials trading platform, the logistics information platform, and the resource demand data streams of one or more concurrent renovation projects. Based on the accessed data from the regional secondary building materials trading platform, the logistics information platform, and the resource demand data streams, construct and continuously update the regional resource-demand graph. In the regional resource-demand graph, the nodes represent resource suppliers, resource demanders, or logistics service providers, and the edges represent the potential supply and demand relationships or transportation paths between nodes. Both nodes and edges are associated with a multimodal attribute set that includes material properties, spatiotemporal location, unit carbon footprint vector, and price.

[0072] The regional secondary building materials trading platform data interface is used to connect to commercial or government-led trading platforms for surplus and recycled building materials. By calling their provided application programming interfaces (APIs), key information such as the category, specifications, quantity, price, supplier location, and contact information of listed resources like secondhand steel, demolished old doors and windows, and recycled aggregates can be retrieved periodically or triggered. The logistics information platform data interface is used to access real-time data from the logistics and transportation market, including vehicle and route data from the government's construction waste supervision platform and pricing interfaces from third-party logistics service platforms, to obtain real-time freight estimates for different routes and vehicle types. The resource demand data flow for concurrent renovation projects is the collaborative data channel within this system. When multiple industrial building renovation projects are simultaneously optimizing their plans, the "Resource Circulation Action Plan" generated by each project (i.e., the future plan for purchasing what materials and selling what used materials) will serve as dynamic demand data, flowing into the overall flow through the system's internal interface.

[0073] Specifically, based on the accessed real-time data stream, a regional resource-demand graph is constructed and continuously updated. This graph contains three types of entity nodes: resource supplier nodes represent sellers on secondhand trading platforms, recycled building material manufacturers, or renovation projects within the system planning to sell used materials; resource demander nodes represent target renovation projects initiating optimization requests within the system, or other concurrent projects with material needs; and logistics service provider nodes represent companies or vehicles that can provide transportation services. Potential supply and demand relationship edges in the regional resource-demand graph are established between supplier and demander nodes, indicating the possibility of a match between the materials available from the supplier and the materials required by the demander. Transportation path edges are established between logistics service provider nodes and resource nodes, or between supplier and demander nodes (considering logistics costs), representing the feasible logistics links between the two points and their cost and time attributes. Each node and edge is associated with a structured attribute set, i.e., a multimodal attribute set, mainly including: material attributes, spatiotemporal location, and a unit carbon footprint vector. The unit carbon footprint vector refers to the carbon emissions involved in producing or acquiring a unit quantity of the material; it is a vector containing components such as production carbon footprint and transportation carbon footprint. This data can come from accessed carbon footprint databases, industry-published carbon emission factors, or standard-based calculation models.

[0074] To ensure timely decision-making, the graph is not built all at once but is continuously updated. The system polls or monitors the "data update query interface" of the aforementioned data interfaces. When changes are detected in the source data (such as new resources being listed, old resources selling out, transportation price adjustments, or new project requirements being added), it automatically triggers attribute updates, additions, or deletions of the corresponding nodes and edges in the graph. This dynamism ensures that optimization decisions are always based on the latest market conditions.

[0075] Understandably, by using dynamic graphs and accessing the "resource demand data stream for concurrent transformation projects," the problem of "static and isolated optimization models that cannot reflect the dynamic collaboration of regional markets" has been solved.

[0076] S3. In response to the transformation needs of the target industrial building, the target industrial building is added as a new demand node to the regional resource-demand graph, and the new demand node is given a corresponding multimodal attribute set based on the digital waste material resource passport.

[0077] Specifically, this step is triggered by a clearly defined "renovation need" for the target industrial building. This need typically originates from the early planning and diagnostics of the project, such as specific engineering tasks identified through building information modeling (BIM) analysis, including structural reinforcement, energy-saving upgrades to the building envelope, and equipment system replacement. Upon system response, a new node representing the target industrial building to be renovated will be created in the constructed "Regional Resource-Demand Map," and its type will be labeled "Demand Side." Then, based on the previously generated "Digital Waste Material Resource Passport," a detailed, multi-dimensional "Multimodal Attribute Set" will be assembled for this node. The "Multimodal Attribute Set" includes: material attributes, spatiotemporal location, unit carbon footprint vector, and price / cost. Material attributes include supply attributes and demand attributes. Supply attributes are derived from the material type, geometric dimensions, estimated mass, and recyclability level of the components to be demolished, recorded in the "passport." Demand attributes are derived from the renovation need list. Price / cost includes supply-side value and demand-side budget. Supply-side value is the potential selling price or disposal revenue (negative cost) estimated based on the material, condition, and market conditions of the old building materials. Demand-side budget is the budget range set for the project to procure new materials. After the above attribute assignments are completed, the newly added node is officially activated and becomes part of the "Regional Resource-Demand Graph". The system will immediately perform preliminary, potential relationship calculations with all existing "Resource Supplier Nodes" and "Logistics Service Provider Nodes" in the graph based on the node's attributes (especially its "Demand Attributes" and "Location Attributes"), providing a preprocessing foundation for the next stage of efficient graph matching algorithms.

[0078] S4. Call the multimodal graph matching algorithm, take the newly added demand-side node as the query center, and search for the optimal subgraph in the regional resource-demand graph. The multimodal graph matching algorithm uses material performance matching degree, transportation carbon emissions, supply time window and comprehensive cost as joint constraints to output one or more resource supply subgraphs. The resource supply subgraphs contain resource supply-side nodes that can meet the transformation needs, logistics path edges and their associated multimodal attribute sets.

[0079] The process involves employing a multimodal graph matching algorithm, using newly added demand-side nodes as the query center, to perform optimal subgraph search within the regional resource-demand graph. Specifically, this includes: extracting the multimodal attribute sets of newly added demand-side nodes and encoding them as high-dimensional query vectors; extracting the multimodal attribute sets of all resource-supply-side nodes in the regional resource-demand graph in parallel and encoding them as high-dimensional candidate vectors; calculating the multidimensional similarity between the high-dimensional query vector and each high-dimensional candidate vector, where multidimensional similarity is a weighted composite value of material attribute similarity, spatiotemporal distance similarity, carbon footprint difference, and economic cost similarity; ranking the resource-supply-side nodes based on multidimensional similarity and performing feasibility filtering based on joint constraints; and selecting the smallest set of nodes that meet the transformation requirements from the nodes that pass the feasibility filtering, and adding the shortest logistics path edges connecting these nodes to form a resource supply subgraph.

[0080] This step addresses an optimal subgraph search problem under multiple constraints. The input is a "regional resource-demand graph" with a newly added project and a specific "modification requirement." The output is one or more "resource supply subgraphs." Each subgraph is a connected network segment containing a set of resource supplier nodes, logistics path edges connecting these supply nodes to project demand nodes, and a complete set of multimodal attributes associated with all nodes and edges.

[0081] Specifically, the algorithm first reads the multimodal attribute set of the "new demand-side node," especially its "demand attributes." A pre-trained encoder then transforms this into a fixed-length high-dimensional query vector. For example, "material type / specification" is converted into a vector using a word embedding model. Numerical features such as "demand" and "time window" are standardized and directly used as vector components. "Geographic location" is converted into a vector through geocoding. Simultaneously, the algorithm traverses all "resource supplier nodes" in the graph in parallel, encoding each node's attribute set (such as material type, inventory, location, unit price, and unit carbon footprint) into corresponding high-dimensional candidate vectors using the same encoder set. Parallel processing ensures search efficiency in large-scale maps.

[0082] It should be noted that the above word embedding model can be trained on its own using domain corpora (such as building material standards and product catalogs), or fine-tuned using a general model.

[0083] Calculate query vector With each candidate vector Multidimensional similarity between This similarity score is not a single metric, but a weighted composite value based on four key dimensions:

[0084] in, For multi-dimensional similarity, For material property similarity, For spatiotemporal distance similarity, For carbon footprint differences, For economic cost similarity, , , , These are the weighting coefficients.

[0085] The aforementioned weighting coefficients are dynamically configured based on the priorities set by the project proponent, or optimized through machine learning methods using historical project data. For example, a demonstration project with carbon neutrality as its primary goal could be set... (Carbon footprint weight) is significantly higher than others.

[0086] Based on overall similarity All candidate nodes are sorted in descending order. Then, filtering is performed based on the following joint constraints: Material performance matching condition (strength, fire resistance rating, etc., must meet design specifications); Transportation carbon emission condition (estimated transportation carbon emissions must not exceed a preset threshold); Supply time window condition (supplier's preparation time + transportation time must be within the project schedule requirements); Comprehensive cost condition (individual and estimated total costs must be within budget). Nodes that do not meet any of these constraints will be eliminated.

[0087] For the filtered set of nodes, the algorithm needs to solve a combinatorial optimization problem: how to cover all project requirements with the minimum set of nodes (to reduce management complexity and potential risks) while minimizing overall cost (or maximizing overall similarity). This is a variation of the classic set coverage problem. Given that it is an NP-hard problem, heuristic algorithms (such as greedy algorithms) or metaheuristic algorithms (such as genetic algorithms) are used in practice to find an approximate optimal solution. The algorithm tries different combinations of nodes and, for each combination, calls a path planning algorithm (such as Dijkstra's algorithm) to calculate the shortest (or lowest carbon) logistics path to the project site, adding it as an edge. Finally, it outputs one or more "resource supply subgraphs" with the highest total utility.

[0088] S5. Instantiate the target industrial building to be transformed and its corresponding transformation requirements into an intelligent agent, and input one or more resource supply subgraphs, digital waste material resource passports and the ontology technical parameters of the target industrial building to a pre-trained multi-agent reinforcement learning optimization engine. The multi-agent reinforcement learning optimization engine runs the policy network of the intelligent agent to generate a dynamic resource decision sequence. The dynamic resource decision sequence arranges a series of action instructions in chronological order. The types of action instructions include purchasing new materials, purchasing specific recycled materials in the resource supply subgraph, and selling old materials in the digital waste material resource passport to specific nodes in the regional resource-demand graph.

[0089] The pre-trained multi-agent reinforcement learning optimization engine is trained through the following steps: Initializing the multi-agent reinforcement learning optimization engine, including constructing a simulation environment with multiple agents, where each agent simulates a virtual transformation project with randomly generated ontology technical parameters and transformation requirements, and the state space of the simulation environment is a dynamic summary of the regional resource-demand graph; defining the action space of each agent, which includes action instruction types; defining the reward function of each agent, which is a linear weighted sum of the carbon emission reduction benefits and economic benefits of the virtual transformation project simulated by the agent over its entire life cycle; having multiple agents interact in the simulation environment in multiple rounds, updating the policy network parameters of each agent through a policy gradient algorithm, until the cumulative rewards of all agents converge, thus obtaining the pre-trained multi-agent reinforcement learning optimization engine.

[0090] Specifically, the abstract concept of "target industrial building renovation and its renovation needs" is transformed into an entity—an intelligent agent—within the reinforcement learning framework capable of learning and decision-making. One or more resource supply subgraphs, a digital waste material resource passport, and the ontological technical parameters of the target industrial building renovation are used as inputs. Through these inputs, the system constructs the initial state for the intelligent agent and embeds it into the environment defined by the "regional resource-demand graph," enabling it to interact with other intelligent agents (representing other concurrent projects) and the environment (the dynamically changing market).

[0091] The training process of the pre-trained optimization engine is completed offline. First, a computational simulation environment is created, and then dozens to hundreds of virtual agents are initialized simultaneously within this environment. Each agent is randomly assigned different "ontology technical parameters" (such as random budgets, project durations, and geographical locations) and "modification requirements" (such as randomly generated material lists) to cover various possible project types. The state space is the "environmental state" observed by each agent at each decision point, a dynamic summary of the regional resource-demand graph. This is not the complete graph data, but rather compressed key features, such as average price trends of various materials, supply-demand ratios, and changes in the distribution of major suppliers, to ensure computational efficiency. The action space is the set of all possible operations that the agent can execute at each decision point, i.e., the action instruction types. Specifically, these include: {Purchase new material A, Purchase recycled material B (from a resource supply subgraph), Sell used material C to market node X, Pause and wait, ...}.

[0092] The reward function designed in this scheme It directly serves the dual goals of zero-carbon transformation, and its form is a linear weighted sum:

[0093] in, For carbon emission reduction benefits, For economic benefits, , These are the weighting coefficients, and , .

[0094] Multiple agents interact in parallel within a shared simulated environment. Based on their observed states, they output action probabilities through their internal policy networks (deep learning models, such as attention-based neural networks) and execute actions. Actions alter their own states and indirectly change the environmental state by influencing resource supply and demand, thus affecting other agents. Multi-agent collaboration employs policy gradient algorithms suitable for multi-agent environments, such as MAPPO (Multi-Agent Proximal Policy Optimization). The algorithm involves agents collecting experience trajectories (states, actions, rewards) through extensive trial and error, using this data to calculate policy gradients, and updating the parameters of their policy networks to favor action sequences that yield higher cumulative rewards in the future. After millions of rounds of simulated interaction training, training converges when the average cumulative reward across all agents no longer increases significantly and the policies stabilize. At this point, the policy networks have internalized the ability to make decisions that balance immediate gains with long-term rewards and self-needs with environmental constraints under complex market dynamics and multi-agent game dynamics. This set of trained policy networks constitutes the "pre-trained multi-agent reinforcement learning optimization engine."

[0095] For a new real-world project (already instantiated as an agent), the system loads a pre-trained optimization engine. The engine inputs the agent's current state (the initial state formed by its input information) into the corresponding policy network, which then directly outputs a dynamic resource decision sequence covering the entire project lifecycle.

[0096] S6. Based on each action instruction in the dynamic resource decision sequence and its associated multimodal attribute set, calculate the estimated total carbon emissions and estimated renovation cost of the target industrial building within the preset full life cycle time range.

[0097] The dynamic resource decision sequence is a list of action instructions arranged in time. The associated multimodal attribute set refers to the fact that each action instruction operates on a resource object (whether it's newly purchased / recycled materials or sold used materials) and is associated with a complete set of attributes, including the unit carbon footprint vector and price. The "preset full lifecycle timeframe" includes the building material production and transportation phase, the construction phase, the operation and maintenance phase, and the demolition and disposal phase.

[0098] Estimated total carbon emissions = Σ(carbon emissions from all "purchasing" activities) + Σ(carbon reduction benefits from all "selling" activities) + net carbon emissions during the building's operation phase after renovation

[0099] Estimated renovation cost = Σ(cost of all "purchasing" activities) + Σ(revenue from all "selling" activities) + other fixed construction costs

[0100] For "procurement" actions, carbon emissions from material production = quantity procured × "production carbon emission factor" in the unit carbon footprint vector of the material; carbon emissions from material transportation = quantity procured × transportation distance × "transportation carbon emission factor" in the unit carbon footprint vector; the total carbon emission contribution of this action = carbon emissions from material production + carbon emissions from material transportation. Material cost = quantity procured × unit price of material (from the attribute set); logistics cost = calculated based on logistics route and freight rate (or directly obtained from the attribute set); the total cost contribution of this action = material cost + logistics cost. For "sale" actions (selling used materials), carbon emission reduction benefit = quantity sold × (corresponding production carbon emission factor of virgin new material - slight carbon emission factor of used material reprocessing); the cost calculation is based on the economic benefit (negative cost) generated by this action, economic benefit = quantity sold × unit price of used material (from market transaction attributes or appraised price). The impact of this action on the total cost is negative (cost reduction). For calculating carbon emissions during the building's operation phase, based on the performance parameters of the purchased and installed energy-saving equipment and building envelope, combined with climate data of the building's location, standard building energy consumption simulation software (such as EnergyPlus, DeST) or simplified estimation methods are used to calculate the annual operational carbon emission reduction after the renovation relative to the baseline scenario. This reduction is then multiplied by the building's lifespan to obtain the total operational carbon emission reduction.

[0101] S7. Integrate dynamic resource decision-making sequences, estimate total carbon emissions, and estimate retrofit costs to generate a zero-carbon retrofit optimization plan for the target industrial building. The zero-carbon retrofit optimization plan includes a list of technical measures and a resource flow action plan.

[0102] The system includes a built-in or connected "Building Energy Conservation and Zero-Carbon Retrofit Technology Measures Library." When the decision sequence includes instructions such as "Purchase and install XX model photovoltaic panels," "Purchase and replace YY type low-emissivity double-glazed windows," or "Purchase and install ZZ energy efficiency level air source heat pumps," the system will retrieve the corresponding standard descriptions, technical parameters, construction process key points, and typical energy-saving / energy-capacity data from the technology measures library to form an itemized "Technical Measures List." The Resource Circulation Action Plan Table expands each action instruction in the decision sequence into a record with complete management attributes. The core fields for each record include: Time Sequence Number: Execution order; Planned Time Window: Suggested specific week or date range for execution; Action Type: Procurement (new / recycled materials), Sale, Logistics Scheduling, etc.; Resource Description: Material name, specifications, quantity; Related Parties: Supplier / Demander name, contact information (extracted from graph node attributes); Logistics Route: Origin and destination points, suggested transportation method; Key Attributes: Unit Price, Total Price, Unit Carbon Footprint, Estimated Carbon Emissions / Carbon Reduction for this Action; Status Tracking: Reserved field for subsequent project execution progress management.

[0103] Furthermore, after generating the zero-carbon transformation optimization scheme for the target industrial building, it also includes: monitoring the updates of data from the regional secondary building materials trading platform and logistics information platform; in response to the detection that the key data updates exceed the preset threshold, triggering the multi-agent reinforcement learning optimization engine to re-evaluate and adjust the dynamic resource decision sequence, and generating scheme adjustment suggestions.

[0104] Among them, the data from the regional secondary building materials trading platform focuses on the market changes of specific material categories and specifications planned for purchase or sale in the "Resource Circulation Action Plan" of the plan, such as price fluctuations, inventory increases or decreases, and newly listed homogeneous or alternative resources; the data from the logistics information platform focuses on the real-time status changes of the logistics routes planned for use in the plan, such as freight adjustments for specific routes, tight transportation capacity, or abnormal events that may affect transportation time (such as traffic control and weather disaster warnings).

[0105] Specifically, after generating the initial plan, the system automatically extracts a list of key resource items to be monitored (such as "H-beam Q345B" and "road transportation from City A to City B") and their initial attribute values ​​(such as initial quotation and initial freight). Subsequently, the system queries the latest data of these specific items on an external platform at a set frequency (such as hourly) through a data interface and compares it with the initial values. "Preset thresholds" are set for different attribute types:

[0106] Price fluctuation threshold: A percentage change threshold is set for material prices and logistics costs. For example, a critical update is considered when the price of a key material planned for procurement in the plan increases by more than 15%, or when the price of an alternative material with equivalent technical performance decreases by more than 20%.

[0107] Supply status threshold: Set an absolute or percentage threshold for inventory quantity. For example, a reassessment is triggered when the supplier's inventory specified in the plan changes from sufficient to less than 50% of the planned purchase quantity, or when it is completely sold out.

[0108] Carbon footprint information update threshold: When the accessed carbon footprint database publishes an update that causes a significant change in the unit carbon footprint factor of a major material in the plan (such as a change of more than 10%), which may affect the overall carbon emission assessment and ranking of the plan, a reassessment is triggered.

[0109] Once any monitored data change exceeds its corresponding preset threshold, the system determines that a "critical data update" has occurred and automatically triggers the subsequent reassessment process.

[0110] Reassessment does not involve re-running the entire optimization process from scratch, but rather iterates efficiently based on a "snapshot" of the current project state. When generating the initial solution, the system saves the complete project state at that time, including remaining modification needs, remaining budget, actions already performed, and a summary of current market environment perception. When a reassessment is triggered, the system uses the latest state as the new starting point for decision-making. The system inputs the current state, the updated "regional resource-demand map" (containing the latest market data), and still feasible alternative "resource supply subgraphs" into the "pre-trained multi-agent reinforcement learning optimization engine." Based on the new environmental information, the engine re-runs the policy network, quickly generating a new "dynamic resource decision sequence (revised version)" starting from the current moment. The system compares the old and new decision sequences (or the latter half starting from the current node), analyzes the differences, and automatically generates easy-to-understand "solution adjustment suggestions." For example:

[0111] "Warning: The price of material X, which was originally planned to be purchased in the fourth week, has increased by 20%. It is recommended to immediately implement alternative plan A (to use local recycled material Y), which is expected to increase costs by 5%, but will keep the carbon emission reduction target unchanged."

[0112] "Opportunity alert: We have discovered a newly listed recycled building material Z, which has matching performance and a lower carbon footprint. We recommend adjusting our plans accordingly, as it is expected to reduce CO2 emissions by an additional 5 tons while lowering costs by 2%."

[0113] "Urgent change: The original logistics route has been interrupted for some reason. It is recommended to switch to the backup route B. The project is expected to be delayed by 2 days and the carbon emissions from transportation will increase by 3%."

[0114] In a more complete implementation, the method further includes: obtaining zero-carbon transformation optimization schemes for all concurrent transformation projects in the regional resource-demand map; aggregating resource flow action plans from all zero-carbon transformation optimization schemes, simulating and calculating material flow and carbon emission changes for all resource nodes in the regional resource-demand map within a preset future time period; and based on material flow and carbon emission changes, identifying regional resource bottlenecks and carbon emission hotspots, and outputting regional collaborative optimization suggestions.

[0115] Specifically, the system periodically (e.g., weekly) or on demand retrieves the "Zero-Carbon Retrofit Optimization Plans" from the internal database for all concurrent retrofit projects that are using the system for optimization and have generated final plans. The key extraction target is the "Resource Flow Action Plan" in each plan. From these plans, the system parses the future actions (procurement / sale), resource details (materials, specifications, quantity), execution time window, location of related parties, and estimated carbon emissions for each record. This essentially obtains a "schedule" of when, where, and what kind of resource inputs and outputs each project will make in the future. The system integrates all project schedules into a regional-level spatiotemporal material flow model according to a unified material classification standard and geographical coordinates. This model records how much inventory of each key building material (such as recycled steel, cement, and glass) is expected to increase or decrease at which nodes (construction sites, recycling plants, and trading markets) in the region each week (or each day). The system employs discrete event simulation or agent-based simulation methods, using the aggregated regional spatiotemporal material flow model and the current state of the "regional resource-demand map" as initial conditions for dynamic extrapolation. The simulation engine simulates the execution of all planned resource actions. For example, simulating week 3, Project A "transports" 10 tons of used steel from its construction site node to recycler node X as planned; simultaneously, Project B "transports" 15 tons of new steel from supplier node Y. By accumulating all such events, the system can calculate the net change in inventory at each node and the transportation load on major logistics routes within a future preset period (such as the next quarter). While simulating material flow, the system simultaneously accumulates the carbon emissions generated by each transportation event (based on distance, mode of transport, and load capacity), as well as the production carbon emissions corresponding to each "purchasing new materials" action. For the action of "selling used materials for reuse," its carbon emission reduction benefits are included. Finally, the simulation outputs the net carbon emission change in the region due to these renovation projects during the specified period, as well as a spatial distribution heat map of carbon emissions.

[0116] Based on simulation results, the system performs intelligent analysis and generates decision support information. Analyzing material flow data, the system can provide early warnings of potential supply-demand imbalances. For example, the simulation reveals that in the fifth week of the future, the total demand for recycled coarse aggregate planned by multiple projects in the region will exceed the sum of current market suppliers' total inventory and expected new supply. The system then identifies a regional-level supply bottleneck for "recycled coarse aggregate" in a specific future time window, which may lead to price spikes or project delays. Analyzing the spatial distribution of carbon emissions, the system can locate anomalies. For example, the simulation reveals that because multiple projects plan to purchase from a single, distant cement plant M, the carbon emission density on a main road connecting plant M and the regional center is abnormally high, becoming a "carbon emission hotspot." Based on these identifications, the system does not directly modify the plans of individual projects, but instead provides neutral, suggestive collaborative optimization prompts to regional managers or all project stakeholders. For example:

[0117] Regarding resource bottlenecks: "Warning: Regional demand for recycled coarse aggregate is expected to exceed supply by 30% next quarter. Recommendations: 1) Coordinate staggered procurement among projects; 2) Guide new suppliers to join; 3) Issue a collective procurement intention to attract external resources."

[0118] Regarding carbon emission hotspots: "Note: The carbon emission concentration of building materials transportation routes heading east exceeds the average by 150%. Recommendations: 1) Assess and introduce local alternative suppliers heading west; 2) Recommend that some projects switch to lower-carbon alternative materials."

[0119] Regarding collaboration opportunities: "Opportunity: The high-quality doors and windows removed from Project A are highly compatible with the new demand from Project C. Direct collaboration could reduce the carbon emissions from logistics and production for both parties by approximately 15 tons. We recommend facilitating the connection."

[0120] This application also provides a multi-objective optimization decision-making system for zero-carbon retrofitting schemes of industrial buildings. The system includes: a data acquisition and passport generation module, used to acquire building information model data and a list of components to be demolished for the target industrial building to be retrofitted, and generate a digital waste material resource passport for the target industrial building to be retrofitted based on the building information model data and the list of components to be demolished; a dynamic graph construction and maintenance module, used to access external data sources and construct and update regional resource-demand graphs; a graph matching decision support module, used to call a multimodal graph matching algorithm to search for resource supply subgraphs in the regional resource-demand graph for new retrofitting needs; a multi-agent optimization decision-making module, used to run a pre-trained multi-agent reinforcement learning optimization engine to generate a dynamic resource decision sequence based on the resource supply subgraph and the digital waste material resource passport; a full life cycle assessment module, used to calculate the estimated total carbon emissions and estimated retrofitting costs based on the dynamic resource decision sequence; and a scheme synthesis and output module, used to integrate the dynamic resource decision sequence, the estimated total carbon emissions, and the estimated retrofitting costs to generate and output a zero-carbon retrofitting optimization scheme.

[0121] Furthermore, the multi-agent optimization decision-making module includes a policy network submodule and an action sequence generation submodule. The policy network submodule stores neural network parameters obtained through offline training. The neural network is used to output the probability distribution of each selectable action given the environmental state and the agent's own state. The action sequence generation submodule is used to generate a dynamic resource decision sequence covering the entire transformation cycle by using a sequence decision algorithm based on the probability distribution output by the policy network submodule.

[0122] Specifically, this module adopts an architecture that separates "policy learning" and "sequence planning." The policy network submodule and the action sequence generation submodule work collaboratively to complete the computational task from environmental perception to long-term action planning. The policy network submodule is a deep neural network model with fixed parameters, trained offline on a large scale. This neural network receives environmental state data (i.e., a dynamic summary of the "regional resource-demand map") and the agent's own state data (i.e., a specific state vector of the current transformation project, including remaining budget, remaining construction period, currently held resources (used materials), and a list of unmet demands). Internally, the network uses multi-layer nonlinear transformations (such as feedforward neural networks or neural networks with attention mechanisms) to perform high-order representation and computation on the fused input states. Its final output is not a specific action, but a discrete probability distribution covering all possible actions. This probability distribution reflects the expected long-term benefits (defined by the reward function) of different actions in the current state. The parameters of this network are obtained through the aforementioned multi-agent reinforcement learning in a simulated environment and are stored offline in this submodule. In practical deployment, this network only performs forward inference calculations and does not undergo online training, thus ensuring the speed and stability of decision response. The inputs to the action sequence generation submodule include: the action probability distribution calculated by the policy network for the current state, the agent's long-term goals (such as total budget, final carbon reduction target), and the project's temporal constraints (such as key project milestones). To achieve long-term optimality, this submodule does not simply select the action with the highest probability at each step (a greedy strategy), but instead uses a sequential decision-making algorithm for rolling planning. A typical implementation is an algorithm combining Monte Carlo Tree Search (MCTS) with the policy network. Its workflow is briefly described as follows: Starting from the current state, using the probability distribution output by the policy network as a reference, multiple rapid "deductions" or "simulations" are performed on various possible future action sequences. In each simulation, a simplified value assessment model is used (or the simulation is run until its end) to estimate the cumulative reward (i.e., the combined value of carbon reduction and economic benefits) that the action sequence may bring. Based on the evaluation results of multiple simulations, the long-term value estimates of each action in the current state are updated retrospectively. After multiple rounds of simulation, the action with the highest long-term value estimate is selected as the final decision for the current step. Once this action is determined, the system state is updated, and this process is repeated to gradually generate a sequence covering the entire cycle in a "rolling" manner. The final output is a specific, timestamped list of action instructions, namely the "dynamic resource decision sequence".

[0123] Furthermore, the system also includes a dynamic adjustment module, which monitors changes in the map data in the dynamic map construction and maintenance module in real time. When it detects significant changes in key data related to the currently generated zero-carbon transformation optimization plan, it automatically sends a re-decision request to the multi-agent optimization decision module and receives new decision results to update the zero-carbon transformation optimization plan.

[0124] Specifically, this module continuously subscribes to updates of the multimodal attribute sets of nodes and edges in the "Regional Resource-Demand Graph" by establishing an internal data bus or event listening interface with the dynamic graph construction and maintenance module. It does not monitor all data, but rather creates a list of key data to be monitored based on the currently effective "Zero-Carbon Transformation Optimization Scheme" (especially the "Resource Circulation Action Plan" within it). This list precisely corresponds to the specific resources, suppliers, and logistics routes planned for transaction in the scheme. When a monitored update event involves an item in the key data list, the module initiates evaluation logic. Evaluation is achieved by comparing the new data with the baseline values ​​recorded in the scheme and checking whether the difference exceeds a "preset threshold." If so, it is considered a significant change. Once a significant change is confirmed, the module sends a structured re-decision request to the multi-agent optimization decision-making module. The multi-agent optimization decision-making module, based on the latest state and environment, quickly runs its decision algorithm to generate a new "Dynamic Resource Decision Sequence" and related evaluation data starting from the current time point. The dynamic adjustment module receives this result and calls the function of the scheme synthesis and output module to integrate the new sequence with the unchanged parts of the project (such as the completed technical measures) to generate an updated version of the "zero-carbon transformation optimization scheme".

[0125] In a more complete implementation, the system further includes a regional collaborative analysis module, which periodically obtains all generated zero-carbon transformation optimization schemes within the region from the scheme synthesis and output module, performs data aggregation and conflict detection, and outputs intervention strategy suggestions to the system administrator for balancing global resources and carbon emissions based on simulation prediction of regional resource supply and demand trends.

[0126] Specifically, this module automatically retrieves all generated and active "zero-carbon transformation optimization schemes" within a specified geographical area from the backend database of the scheme synthesis and output module at a set cycle (e.g., weekly or bi-weekly). The extracted data is the "resource flow action plan" for each scheme. The module parses and standardizes all project plans, extracting information such as material types and quantities, planned transaction time windows, supplier and demand geographical locations, estimated transportation routes, and carbon emissions. These micro-level, discrete project plans are integrated into a unified regional spatiotemporal material flow database, forming a "macro-plan map" of regional building material resource flow and carbon emissions over a future period. Based on the aggregated data, the module runs a conflict detection algorithm, monitoring resource demand conflicts and spatiotemporal and logistical conflicts. Resource demand conflict analysis examines the planned total demand for a specific type of resource (e.g., specific specifications of recycled steel or photovoltaic modules) from all projects within the region during a specific future period (e.g., next month), comparing it with the estimated total supply in the current regional market (reflected through a resource-demand map). When demand consistently and significantly exceeds supply, the system issues a warning of potential regional supply bottlenecks, which could lead to collective price increases or construction delays. Spatiotemporal and logistical conflicts are detected by examining whether the logistics plans for multiple projects are excessively concentrated at key transportation nodes or specific routes, potentially causing logistical congestion; or by checking whether the demolition and new construction plans for multiple projects are closely aligned in time, to assess the spatiotemporal matching efficiency of construction waste generation and resource utilization.

[0127] To more dynamically assess risks, the module incorporates or invokes a discrete event simulation engine. This engine starts with the current regional resource-demand map and uses the aggregated "Resource Flow Action Plan" for each project as input to drive future events. It simulates material flow, changes in node inventory, transportation network load, and the accompanying carbon emission accumulation process over a future period (e.g., a quarter). Through simulation, it can quantitatively predict: the supply-demand gap curves for different resource categories, the transportation load and carbon emission density distribution of major logistics corridors, and the total net carbon emissions and their temporal distribution for the region as a whole due to these transformation projects, under a given plan.

[0128] Based on conflict detection results and simulation prediction data, the module does not directly modify any individual project plan, but generates a macro-level intervention strategy recommendation report for system administrators.

[0129] The detailed description of the above specific embodiments fully illustrates the feasibility, preferred implementation, and technical effects achieved by the technical solution of the present invention. Those skilled in the art can make several modifications and substitutions based on the above description without departing from the principles and spirit of the present invention, and these modifications and substitutions should also be considered within the scope of protection of the present invention.

[0130] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention. The actual method is not limited to this. In conclusion, if those skilled in the art are inspired by this description and design similar methods and embodiments without departing from the spirit of the present invention, they should all fall within the protection scope of the present invention.

Claims

1. A multi-objective optimization decision-making method for zero-carbon retrofitting of industrial buildings, characterized in that, The method includes: Obtain the building information model data and the list of components to be demolished for the target industrial building to be renovated. Based on the building information model data and the list of components to be demolished, generate a digital waste material resource passport for the target industrial building to be renovated. The digital waste material resource passport includes the material type, geometric dimensions, mass estimate and recyclability level characteristics of the components to be demolished. The system accesses data interfaces from regional secondary building materials trading platforms, logistics information platforms, and resource demand data streams from one or more concurrent renovation projects. Based on the accessed data from these platforms, the system constructs and continuously updates a regional resource-demand graph. In this graph, nodes represent resource suppliers, resource demanders, or logistics service providers, and edges represent potential supply and demand relationships or transportation routes between nodes. Each node and edge is associated with a multimodal attribute set that includes material properties, spatiotemporal location, unit carbon footprint vector, and price. In response to the renovation needs of the target industrial building, the target industrial building is added as a new demand-side node to the regional resource-demand graph, and the new demand-side node is assigned the corresponding multimodal attribute set based on the digital waste material resource passport. The multimodal graph matching algorithm is invoked, and the newly added demand-side node is used as the query center to search for the optimal subgraph in the regional resource-demand graph. The multimodal graph matching algorithm uses material performance matching degree, transportation carbon emissions, supply time window and comprehensive cost as joint constraints to output one or more resource supply subgraphs. The resource supply subgraphs include resource supply-side nodes that can meet the transformation requirements, logistics path edges and their associated multimodal attribute sets. The target industrial building to be transformed and its corresponding transformation requirements are instantiated as an agent. The one or more resource supply subgraphs, the digital waste material resource passport, and the ontological technical parameters of the target industrial building to be transformed are jointly input into a pre-trained multi-agent reinforcement learning optimization engine. The multi-agent reinforcement learning optimization engine runs the policy network of the agent to generate a dynamic resource decision sequence. The dynamic resource decision sequence arranges a series of action instructions in chronological order. The types of action instructions include purchasing new materials, purchasing specific recycled materials in the resource supply subgraph, and selling old materials in the digital waste material resource passport to specific nodes in the regional resource-demand graph. Based on each action instruction in the dynamic resource decision sequence and its associated multimodal attribute set, the estimated total carbon emissions and estimated renovation cost of the target industrial building within a preset life cycle time range are calculated. By integrating the dynamic resource decision sequence, the estimated total carbon emissions, and the estimated retrofit cost, a zero-carbon retrofit optimization plan for the target industrial building is generated. The zero-carbon retrofit optimization plan includes a list of technical measures and a resource flow action plan.

2. The multi-objective optimization decision-making method for zero-carbon retrofitting of industrial buildings according to claim 1, characterized in that, The generation of the digital waste material resource passport for the target industrial building to be renovated specifically includes: The component to be demolished was subjected to a three-dimensional laser scan to obtain point cloud data; Based on the point cloud data, component segmentation and material identification are performed to obtain the material type and geometric dimensions of each component; By combining the material type and the geometric dimensions, and based on a preset material density database, an estimated mass value is obtained; Based on the material type and historical service environment data, the recyclability level characteristic data are predicted by a pre-trained damage assessment model.

3. The multi-objective optimization decision-making method for zero-carbon retrofitting of industrial buildings according to claim 1, characterized in that, The invocation of the multimodal graph matching algorithm, using the newly added demand-side node as the query center, performs an optimal subgraph search within the regional resource-demand graph, specifically including: Extract the multimodal attribute set of the newly added demand-side node and encode it into a high-dimensional query vector; The multimodal attribute sets of all resource supplier nodes in the regional resource-demand graph are extracted in parallel and encoded into high-dimensional candidate vectors respectively; Calculate the multi-dimensional similarity between the high-dimensional query vector and each of the high-dimensional candidate vectors. The multi-dimensional similarity is a weighted comprehensive value of material property similarity, spatiotemporal distance similarity, carbon footprint difference, and economic cost similarity. Resource supplier nodes are sorted according to the multi-dimensional similarity and feasibility filtering is performed based on the joint constraints. From the nodes that pass the feasibility filter, select the smallest set of nodes that meet the transformation requirements, and add the shortest logistics path edges connecting these nodes to form the resource supply subgraph.

4. The multi-objective optimization decision-making method for zero-carbon retrofitting of industrial buildings according to claim 1, characterized in that, The pre-trained multi-agent reinforcement learning optimization engine is obtained through the following steps: Initializing the multi-agent reinforcement learning optimization engine includes constructing a simulation environment comprising multiple agents, wherein each agent simulates a virtual modification project with randomly generated ontology technical parameters and modification requirements, and the state space of the simulation environment is a dynamic summary of the regional resource-demand graph; Define the action space for each agent, which includes action instruction types; Define a reward function for each agent, which is a linear weighted sum of the carbon reduction benefits and economic benefits of the virtual transformation project simulated by the agent over its entire life cycle; The multiple agents are instructed to interact in multiple rounds in the simulated environment. The policy network parameters of each agent are updated using a policy gradient algorithm until the cumulative reward of all agents converges, thus obtaining the pre-trained multi-agent reinforcement learning optimization engine.

5. The multi-objective optimization decision-making method for zero-carbon retrofitting of industrial buildings according to claim 1, characterized in that, After generating the zero-carbon retrofit optimization scheme for the target industrial building, the method further includes: Monitor the updates of data from the secondary building materials trading platform and the logistics information platform in the aforementioned region; In response to the detection that key data updates exceed a preset threshold, the multi-agent reinforcement learning optimization engine is triggered to re-evaluate and adjust the dynamic resource decision sequence and generate scheme adjustment suggestions.

6. The multi-objective optimization decision-making method for zero-carbon retrofitting of industrial buildings according to claim 1, characterized in that, The method further includes: Obtain zero-carbon retrofit optimization schemes for all concurrent retrofit projects in the aforementioned regional resource-demand map; By aggregating the resource flow action plans in all the zero-carbon transformation optimization schemes, simulation calculations are performed on the material flow and carbon emission changes of all resource nodes in the regional resource-demand map within a preset future time period. Based on the changes in material flow and carbon emissions, regional resource bottlenecks and carbon emission hotspots are identified, and regional collaborative optimization suggestions are output.

7. A multi-objective optimization decision-making system for zero-carbon retrofitting of industrial buildings, characterized in that, The system includes: The data acquisition and passport generation module is used to acquire the building information model data and the list of components to be demolished for the target industrial building to be renovated, and to generate a digital waste material resource passport for the target industrial building to be renovated based on the building information model data and the list of components to be demolished. The dynamic map construction and maintenance module is used to connect to external data sources and build and update regional resource-demand maps. The graph matching decision support module is used to call the multimodal graph matching algorithm to search for resource supply subgraphs in the resource-demand graph of the region for newly added transformation requirements; The multi-agent optimization decision module is used to run a pre-trained multi-agent reinforcement learning optimization engine to generate a dynamic resource decision sequence based on the resource supply subgraph and the digital waste material resource passport. The life cycle assessment module is used to calculate the estimated total carbon emissions and estimated retrofit costs based on the dynamic resource decision sequence. The scheme synthesis and output module is used to integrate the dynamic resource decision sequence, the estimated total carbon emissions, and the estimated retrofit cost to generate and output a zero-carbon retrofit optimization scheme.

8. The multi-objective optimization decision-making system for zero-carbon retrofitting of industrial buildings according to claim 7, characterized in that, The multi-agent optimization decision-making module includes a policy network submodule and an action sequence generation submodule; The policy network submodule stores neural network parameters obtained through offline training. The neural network is used to output the probability distribution of each selectable action given the environmental state and the agent's own state. The action sequence generation submodule is used to generate the dynamic resource decision sequence covering the entire transformation cycle by using a sequence decision algorithm based on the probability distribution output by the strategy network submodule.

9. The multi-objective optimization decision-making system for zero-carbon retrofitting of industrial buildings according to claim 7, characterized in that, The system also includes a scheme dynamic adjustment module, which is used to monitor the changes in the map data in the dynamic map construction and maintenance module in real time, and when it detects that there is a significant change in the key data related to the currently generated zero-carbon transformation optimization scheme, it automatically initiates a re-decision request to the multi-agent optimization decision module and receives the new decision results to update the zero-carbon transformation optimization scheme.

10. The multi-objective optimization decision-making system for zero-carbon retrofitting of industrial buildings according to claim 7, characterized in that, The system also includes a regional collaborative analysis module, which periodically obtains all generated zero-carbon transformation optimization schemes within the region from the scheme synthesis and output module, performs data aggregation and conflict detection, and outputs intervention strategy suggestions to the system administrator to balance global resources and carbon emissions based on simulation prediction of regional resource supply and demand trends.

Images