A big data enabled urban construction planning dynamic optimization method

Abstract

The application discloses a kind of big data empowerment city construction planning dynamic optimization method, it is related to wisdom city technical field, including, fusion multi-source heterogeneous real-time data stream, construct and physical city synchronous evolution digital twin;Analysis city overall planning text, generate planning intent knowledge graph;In digital twin, utilize planning intent knowledge graph to drive multi-agent to carry out autonomous simulation and decision deduction, output planning decision instruction set;Planning decision instruction set is carried out executability verification and serialization, and is issued to physical city execution unit, and execution effect is fed back to digital twin and is continuously optimized.The application utilizes planning intent knowledge graph to drive multi-agent to carry out concurrent simulation and strategy evolution in digital twin, by real-time calculation decision influence on intent atom achievement degree and dynamically adjust each agent strategy, realized the collaborative optimization under multi-objective conflict and Pareto frontier convergence.

Description

Technical Field

[0001] This invention relates to the field of smart city technology, and in particular to a dynamic optimization method for urban construction planning empowered by big data. Background Technology

[0002] Currently, big data and digital twin technologies offer new technological pathways for urban planning. Conventional methods typically rely on static analysis models based on historical data and GIS (Geographic Information System), or single-objective simulations for localized optimization of specific subsystems (such as traffic flow). In recent years, accessing multi-source data through the Internet of Things (IoT), remote sensing, and other means to construct urban information models, and utilizing intelligent agents to simulate individual behaviors for complex system simulations, has become a significant trend in this field.

[0003] However, existing technologies have significant limitations. First, the data layer and the decision-making layer are often disconnected. Although multi-source heterogeneous data are collected, the lack of a high-fidelity digital twin that evolves synchronously with the physical city provides a unified decision-making context, leading to discrepancies between the spatiotemporal benchmark of simulations and the real world. Second, planning objectives are often presented in the form of natural language text or abstract indicators, making them difficult for algorithms to directly understand and quantify into computable optimization goals. Furthermore, they cannot serve as a unified co-evolutionary signal in multi-agent concurrent decision-making processes, thus failing to effectively address the dynamic conflicts and collaborative balance issues among multiple objectives (such as traffic efficiency, land development, and ecological protection). Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a big data-enabled dynamic optimization method for urban construction planning to solve the problem of difficulty in achieving multi-objective dynamic collaborative optimization and closed-loop execution verification due to the disconnect between data, models and planning intentions.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a big data-enabled dynamic optimization method for urban construction planning, comprising:

[0008] By integrating real-time data streams from multiple heterogeneous sources, a digital twin that evolves synchronously with the physical city can be constructed.

[0009] Analyze urban master plan texts to generate a knowledge graph of planning intent;

[0010] In a digital twin, a planning intent knowledge graph is used to drive multiple agents to perform autonomous simulation and decision-making, and output a set of planning decision instructions.

[0011] The feasibility of the planning decision instruction set is verified and serialized, and then distributed to the physical city execution units. The execution results are fed back to the digital twin for continuous optimization.

[0012] As a preferred embodiment of the big data-enabled dynamic optimization method for urban construction planning described in this invention, the specific steps for integrating multi-source heterogeneous real-time data streams to construct a digital twin that evolves synchronously with the physical city are as follows.

[0013] Access to multi-source heterogeneous raw data streams from IoT sensor networks, mobile signaling data, remote sensing data, and public data platforms;

[0014] By using edge computing nodes, multi-source heterogeneous raw data streams are cleaned and timestamped to generate standardized real-time data streams.

[0015] Standardized real-time data streams are dynamically integrated into the city information model platform to generate a digital twin that evolves in sync with the physical city.

[0016] As a preferred embodiment of the big data-enabled dynamic optimization method for urban construction planning described in this invention, the specific steps for parsing the urban master plan text and generating a planning intent knowledge graph are as follows:

[0017] Analyze urban planning texts and related normative documents to extract key target statements;

[0018] Deconstruct key target statements into intent atoms containing the main object, quantitative indicators, and desired state;

[0019] In a digital twin, spatial mapping and data association are performed on the intention atoms, mapping the desired state to the corresponding quantitative indicators and target values;

[0020] Based on quantitative indicators and target values, the relationships and conflicts between different intention atoms are identified, a dynamic intention network with weights and constraints is constructed, and a planning intention knowledge graph is generated.

[0021] As a preferred embodiment of the big data-enabled dynamic optimization method for urban construction planning described in this invention, the specific steps of using a planning intent knowledge graph to drive multi-agent autonomous simulation and decision-making inference are as follows:

[0022] Based on the planning intent knowledge graph, a cluster of intelligent agents representing different urban functions is instantiated in a digital twin; the cluster of intelligent agents includes traffic flow intelligent agents, land development intelligent agents, and ecological space intelligent agents.

[0023] The intelligent agent cluster learns behavioral patterns from the historical and real-time data of the digital twin, and uses the quantitative indicators and target values ​​defined in the planning intent knowledge graph as long-term reward signals for co-evolution.

[0024] Drive the intelligent agent cluster to make asynchronous and concurrent decisions and interactions in the sandbox environment of the digital twin, triggering simulation and prediction of changes in the city's state;

[0025] The system calculates the impact of changes in city status on the achievement of relevant intention atoms, generates guidance signals and dynamically adjusts the decision-making strategies of each agent, synchronously records the state prediction data of the agents before the execution of instructions, and generates a planning decision instruction set.

[0026] As a preferred embodiment of the big data-enabled dynamic optimization method for urban construction planning described in this invention, the intelligent agent cluster learns behavioral patterns from historical and real-time data of the digital twin, with the following specific steps:

[0027] A deep reinforcement learning model is constructed for traffic flow intelligent agents. The deep reinforcement learning model is trained using historical traffic flow data and real-time mobile signaling data to learn the mapping relationship between traffic light timing strategies and traffic flow status.

[0028] A graph neural network model is constructed for the land development agent. The graph neural network model is trained using historical land use change data, land price data and normative document text to learn the probability distribution of land development type conversion.

[0029] A multi-objective optimization model is constructed for the intelligent agent of ecological space. The model is trained using historical environmental monitoring data and green space data to learn a balance strategy between ecological benefits and spatial constraints.

[0030] As a preferred embodiment of the big data-enabled dynamic optimization method for urban construction planning described in this invention, the driving intelligent agent cluster performs asynchronous and concurrent decision-making and interaction in the sandbox environment of the digital twin. The specific steps are as follows.

[0031] Based on the learned traffic light timing strategy, the traffic flow agent issues adjustment instructions for the traffic light cycle and the ratio of green light time to signal cycle duration at intersections in the digital road network within the simulation cycle;

[0032] Based on the learned probability distribution, the land development agent issues proposed instructions for changing land use and adjusting intensity for land parcels that meet the development conditions within the simulation period.

[0033] Based on the learned balancing strategy, the ecological space intelligent agent issues ecological compensation and spatial optimization constraint instructions to the proposed development plots and existing green spaces during the simulation cycle.

[0034] Collision detection and conflict resolution are performed on adjustment instructions, proposal instructions, and constraint instructions within the unified spatiotemporal framework of the digital twin to generate city status update events.

[0035] As a preferred embodiment of the big data-enabled dynamic optimization method for urban construction planning described in this invention, the specific steps for calculating the impact of changes in urban state on the achievement degree of relevant intention atoms are as follows:

[0036] Listen for city status update events and extract the affected intent atoms and quantitative indicators from the planning intent knowledge graph;

[0037] By utilizing the real-time computing capabilities of the digital twin, the instantaneous values ​​of each affected quantitative indicator are calculated after a city status update event occurs.

[0038] The instantaneous value is compared with the expected state value of the corresponding intention atom to calculate the target gap.

[0039] Based on the target gap and the weight relationship between intention atoms, a multi-objective optimization algorithm is used to calculate the reward and punishment value, generate a guidance signal, and distribute it to the corresponding agent.

[0040] As a preferred embodiment of the big data-enabled dynamic optimization method for urban construction planning described in this invention, the specific steps for dynamically adjusting the decision-making strategies of each intelligent agent are as follows:

[0041] Each agent receives a guidance signal and combines it with its own inherent reward to form a new reward function;

[0042] Based on the new reward function, and combined with the updated environmental state of the digital twin triggered by the city state update event, the policy gradient is updated and the action value is reassessed.

[0043] The updated decision-making strategy is applied to the next simulation cycle, forming a strategy iterative optimization loop that converges to the Pareto front and outputs a planning decision instruction set.

[0044] As a preferred embodiment of the big data-enabled dynamic optimization method for urban construction planning described in this invention, the specific steps for verifying and serializing the executability of the planning decision instruction set are as follows:

[0045] Load the planning and decision-making instruction set and access the real-time constraint information database in the digital twin. Verify the physical feasibility, economic cost compliance and immediate risk level of each instruction, adjust conflicting instructions, and form an instruction sequence.

[0046] The verified and adjusted instruction sequence is transformed into standardized control instructions that can be recognized by the execution unit and securely issued to the corresponding physical execution unit.

[0047] As a preferred embodiment of the big data-enabled dynamic optimization method for urban construction planning described in this invention, the specific steps for feeding back the execution results to the digital twin for continuous optimization are as follows:

[0048] Start the effect tracking thread to continuously receive physical city real-time status monitoring data from the IoT sensor network;

[0049] The real-state monitoring data and the state prediction data of the agent in the digital twin before the execution of the instructions are spatiotemporally aligned and compared to calculate the prediction-reality deviation matrix.

[0050] The prediction-reality deviation matrix is ​​fed back to the digital twin to calibrate the parameters of the internal simulation model and to perform online incremental training on multiple agents, thereby achieving continuous closed-loop optimization.

[0051] The beneficial effects of this invention are as follows: By integrating multi-source heterogeneous real-time data streams to construct a synchronously evolving digital twin, a unified spatiotemporal benchmark and decision sandbox with high fidelity and millisecond-level updates aligned with the physical city is provided for the entire optimization process, solving the problem of the simulation environment being disconnected from the real world; by parsing planning text to generate a structured planning intent knowledge graph, abstract planning goals are deconstructed and mapped into quantifiable indicators and target values ​​attached to specific elements of the digital twin, and the relationships, conflicts, and weights among them are clarified, thus providing accurate and computable multi-objective guidance for subsequent optimization; the planning intent knowledge graph is used to drive multiple agents to conduct concurrent simulation and strategy evolution in the digital twin, and the impact of decisions on the degree of achievement of intent atoms is calculated in real time and the strategies of each agent are dynamically adjusted, realizing collaborative optimization and Pareto front convergence under multi-objective conflicts. Attached Figure Description

[0052] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 A flowchart of a dynamic optimization method for urban construction planning empowered by big data.

[0054] Figure 2 A flowchart for building and merging data into a digital twin.

[0055] Figure 3 This is a flowchart for multi-agent simulation and decision-making inference.

[0056] Figure 4 Flowchart for instruction verification and closed-loop optimization. Detailed Implementation

[0057] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0058] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0059] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0060] Reference Figures 1-4 As one embodiment of the present invention, this embodiment provides a big data-enabled dynamic optimization method for urban construction planning, comprising the following steps:

[0061] S1. Integrate multi-source heterogeneous real-time data streams to construct a digital twin that evolves synchronously with the physical city.

[0062] S1.1. The operational status data from the Internet of Things sensor network, including building energy consumption data and environmental monitoring data; population flow trajectory data from traffic checkpoints and mobile signaling, used to generate traffic flow data and population density data; multispectral geospatial data collected from remote sensing satellites and drones; and planning texts and geological exploration data from public data platforms are uniformly received as multi-source heterogeneous raw data streams and sent to the edge computing node.

[0063] Edge computing nodes perform data cleaning operations on multi-source heterogeneous raw data streams, including handling missing values, removing outliers, and unifying data formats. At the same time, they perform timestamp alignment operations to convert the timestamps of each data point in the multi-source heterogeneous raw data stream to a unified Coordinated Universal Time (UTC) time base, and output a standardized real-time data stream.

[0064] S1.2. Standardized real-time data streams are input into a cloud-native city information modeling platform. This platform incorporates a city-level 3D geographic information system (GIS) containing semantic information. This system statically integrates building information models, underground pipeline models, and 3D geological models. The platform uses a synchronization engine to parse the standardized real-time data streams, dynamically mapping and updating the real-time indicators such as traffic flow, building energy consumption, environmental monitoring data, and pedestrian density to the corresponding attributes of buildings, road segments, monitoring points, and areas within the 3D GIS. This achieves synchronized updates between the 3D scene and data attributes, generating a digital twin that evolves synchronously with the physical city.

[0065] S2. Analyze the urban master plan text and generate a knowledge graph of planning intent.

[0066] S2.1. Analyze urban planning texts (guiding documents that stipulate the city's medium- and long-term development goals, scale, layout, and functions) and related normative documents (e.g., annual construction land supply plans). Specifically, use dependency parsing to analyze sentence structure, identify subjects, predicates, objects, and modifiers, and perform semantic role labeling to identify the semantic roles associated with predicates in sentences, such as agent, patient, time, place, and measurement. Through analysis, select declarative sentences containing clear quantitative indicators, target values, and spatial or temporal categories, and identify and extract key target statements that express clear development goals. Key target statements are, for example, statements such as "By 2035, the public transportation modal share in the central urban area will increase to 50%" and "During the planning period, the per capita green space area will increase by 2 square meters in the eastern new area."

[0067] Each extracted key objective statement is deconstructed into an intent atom, which consists of three elements: the subject identified from the statement, the quantifiable indicator, and the desired state. For example, the intent atom deconstructed from the statement "increase the public transportation share in the central urban area to 50%" has the subject being the public transportation share in the central urban area, the quantifiable indicator being the public transportation share, and the desired state being 50%.

[0068] S2.2. Input each intent atom into the digital twin. In the context of the digital twin's city-level 3D geographic information system and real-time data streams (traffic flow data, building energy consumption data, environmental monitoring data, and population density data, etc.), map the subject of the intent atom to specific spatial elements or data dimensions. For example, map the subject of public transport share to the road and public transport network spatial data defined by the urban basic geographic information data contained in the digital twin's city-level 3D geographic information system, as well as the population flow data represented by mobile signaling data. At the same time, explicitly map the desired state of the intent atom to specific numerical target values ​​corresponding to quantitative indicators.

[0069] S2.3. Based on the quantitative indicators and target values ​​associated with all intent atoms, analyze the logical and spatial relationships between different intent atoms, and identify the association and conflict relationships between intent atoms. For example, identify that the intent atom for increasing per capita green space area and the intent atom for ensuring the supply of construction land have a competitive conflict in terms of spatial occupation. Specifically, in the city-level three-dimensional geographic information system of the digital twin, compare the spatial element range or data dimension associated with each intent atom. If there is geographic spatial overlap or resource type intersection, it is determined that there is a logical association. Through spatial analysis, such as overlay analysis, check whether the spatial layout or resource allocation required by the target value of the intent atom is mutually exclusive or competitive in the digital twin simulation environment, thereby identifying spatial conflicts.

[0070] After identifying competitive conflicts, the specific intention pairs involved in the conflict, the conflict type, and the spatial location or resource category where the conflict occurs are recorded. Weights are assigned to the identified relationships, and constraints in terms of space and resources are recorded. A dynamic intention network with weights and constraints is constructed to generate a planning intention knowledge graph.

[0071] It should be noted that the weights are represented by normalized weight coefficients, with a value range of [0,1]. The weights within the same evaluation set (the same set of intent atoms at the same level or the same set of conflict / associations) are normalized so that the sum of all weights within the same evaluation set is 1.

[0072] The weight allocation is based on the explicit provisions of the priority of objectives in the urban planning document. For example, if the document explicitly marks the public transportation modal share as a binding indicator and the per capita green space area as an expected indicator, then the intention atom corresponding to the binding indicator will be assigned a significantly higher weight value than the expected indicator in the knowledge graph (for example, the former is set to 0.7 and the latter to 0.3). If there are no explicit provisions, the importance of each objective is compared pairwise using the analytic hierarchy process, a judgment matrix is ​​constructed, and an initial weight vector is calculated. The initial weight vector is then normalized and mapped to the interval [0,1] and the sum of the weights is 1. This is then written into the planning intention knowledge graph as the final weight value.

[0073] To further explain, the importance of each objective is compared pairwise using the analytic hierarchy process (AHP), a judgment matrix is ​​constructed, and an initial weight vector is calculated, as follows:

[0074] (1) From the extracted intention atoms, select intention atoms of the same planning level, the same evaluation caliber and which can simultaneously constrain decision-making to form a comparison set, and bind the quantitative indicator name, target value and spatial / resource constraint label of each intention atom as a unified description benchmark for the comparison object;

[0075] (2) In accordance with the principle of prioritizing those with stronger constraints on planning, greater impact on public safety / people's livelihood, higher irreversible resource occupation, and stronger mandatory nature of higher-level planning, an importance scale is used to assign values ​​to the relative importance of any two intention atoms. The importance scale is a discrete level, for example, divided into five levels: "equally important, slightly important, obviously important, strongly important, and extremely important", and each level corresponds to a fixed numerical code. When intention atom A is more important than intention atom B, the level value of A to B is recorded; when B is more important, the reverse level value is recorded; when the two are equally important, the equal level value is recorded.

[0076] (3) Using the intention atoms in the comparison set as row and column indices, fill in the relative importance level values ​​one pair at a time to obtain the complete judgment matrix; the main diagonal of the judgment matrix is ​​filled with equal importance, and the matrix satisfies the mutual reversal and consistency filling constraint (that is, after the importance of A relative to B is determined, B takes the corresponding symmetrical reverse level relative to A).

[0077] (4) After normalizing each column of the judgment matrix, calculate the average value of each row and use the row average value as the initial weight of the corresponding intention atom;

[0078] (5) Obtain consistency indicators based on pairwise comparison results (read from the pre-set consistency indicator comparison table according to the judgment matrix dimension), compare the consistency indicators with the consistency threshold (the distribution of consistency indicators calculated by pairwise comparison data of historical / trial projects is calibrated, and “mean + 2 times standard deviation” or “95th percentile” (that is, to ensure that about 95% of the existing samples meet the consistency requirements) is taken as the threshold, and the upper bound of abnormal inconsistency in a statistical sense is used to control the impact of cyclic contradictions on the stability of weights). When the consistency indicators do not meet the consistency threshold requirements, locate the comparison pair that caused the contradiction (for example, the cyclic contradiction of “A is significantly more important than B, B is significantly more important than C, but C is significantly more important than A”) and prompt to revert to step (2) to reassign the relevant comparison pair until the consistency test is passed; normalize the initial weights that have passed the consistency test so that each weight value falls into the [0,1] interval and the weight sum is 1, and write it into the planning intent knowledge graph as the final weight value.

[0079] Constraints refer to external limiting factors extracted from digital twins or planning documents that restrict the realization of the intended purpose. Specifically, they include land engineering geological condition data extracted from the three-dimensional geological model of the digital twin, permanent basic farmland protection red line and ecological protection red line data extracted from the attached drawings or database of the planning document, and road design traffic capacity data extracted from real-time data streams.

[0080] S3. In the digital twin, the planning intent knowledge graph is used to drive multiple agents to perform autonomous simulation and decision inference, and output a planning decision instruction set.

[0081] S3.1. Based on the planning intent knowledge graph, instantiate intelligent agent clusters representing different urban functions in the digital twin. The intelligent agent clusters include traffic flow intelligent agents, land development intelligent agents, and ecological space intelligent agents.

[0082] The intelligent agent swarm learns behavioral patterns from the historical and real-time data of digital twins, and uses the quantitative indicators and target values ​​defined in the planning intent knowledge graph as long-term reward signals for co-evolution, specifically:

[0083] A deep reinforcement learning model is constructed for traffic flow intelligent agents. The deep reinforcement learning model is trained using historical traffic flow data and real-time mobile signaling data stored in the digital twin to learn the mapping relationship between traffic light timing strategies and traffic flow status.

[0084] It should be noted that the deep reinforcement learning model is constructed using a deep deterministic policy gradient algorithm. The input state space of the deep reinforcement learning model is defined as the traffic flow, queue length, average vehicle speed, and current signal timing parameters of each intersection in the digital road network within the digital twin. The output action space is defined as the adjustment amount of the signal light cycle, the ratio of green light time to signal cycle duration at the intersection. The training process utilizes historical traffic flow data and real-time mobile signaling data stored in the digital twin, with the reward being maximizing traffic efficiency (defined as the total number of vehicles passing through) and minimizing average delay time. Through interaction with the environment (i.e., the traffic simulation in the digital twin), the strategy is iterated to learn the mapping relationship between signal timing strategy and traffic flow situation.

[0085] A graph neural network model is constructed for the land development agent. The graph neural network model is trained using historical land use change data, land price data and text features parsed from relevant normative documents stored in the digital twin to learn the probability distribution of land development type conversion.

[0086] It should be noted that the graph neural network model is built on graph convolutional networks. The graph neural network model discretizes urban space into a graph of land parcel units. Node features include the current use, plot ratio, land price, accessibility of surrounding facilities, and text feature vectors parsed from relevant regulatory documents. Edge features represent the spatial adjacency and functional connection between land parcels. The training process utilizes historical land use change data and land price data stored in the digital twin to predict the probability distribution of land parcel use conversion types (e.g., conversion from industrial land to commercial land) in the next planning cycle. Through supervised learning, the cross-entropy loss function is used for backpropagation to update network parameters and learn the probability distribution of land parcel development type conversion.

[0087] A multi-objective optimization model is constructed for the intelligent agent of ecological space. The model is trained using historical environmental monitoring data and green space data stored in the digital twin to learn a balance strategy between ecological benefits and spatial constraints.

[0088] It should be noted that the multi-objective optimization model is constructed using a non-dominated sorting genetic algorithm. The decision variables of the multi-objective optimization model are green space layout (location and area of ​​newly added green space) and ecological compensation measures (such as vegetation type). The objective function includes maximizing ecological benefits (quantified by indicators such as vegetation carbon sequestration and cooling effect) and minimizing space occupation costs. The constraints include ecological protection red lines, land engineering geological conditions, and available budget. The training process utilizes historical environmental monitoring data and green space data stored in the digital twin to generate Pareto optimal solution sets through algorithm iteration, and learns the balance strategy between ecological benefits and spatial constraints.

[0089] Ideally, by instantiating intelligent agent clusters with specific learning capabilities for different urban functions, abstract planning intentions are transformed into executable and evolvable continuous optimization mechanisms. Traffic flow intelligent agents adaptively optimize signal timing in simulations through deep reinforcement learning, directly improving road network traffic efficiency. Land development intelligent agents accurately predict the probability of land use conversion using graph neural networks, providing quantitative basis for development decisions and reducing planning conflicts. Ecological space intelligent agents use multi-objective optimization to solve the Pareto optimal solution for green space layout under multiple constraints, achieving a balance between ecology and construction. The three types of intelligent agents use the planning intention knowledge graph as a unified reward signal to conduct concurrent inference and strategy co-evolution in a digital twin, ultimately outputting a planning decision instruction set that conforms to multiple objectives and is physically executable, realizing a paradigm shift from static planning to dynamic continuous optimization.

[0090] S3.2. Drive the intelligent agent cluster to make asynchronous and concurrent decisions and interactions within the sandbox environment of the digital twin, triggering simulated and predicted changes in the city's state. The sandbox environment is a simulation space within the digital twin that is isolated from the real-time data stream, specifically as follows:

[0091] Within each simulation cycle, the traffic flow agent, based on the learned traffic light timing strategy, issues adjustment instructions for the traffic light cycle and the ratio of green light time to signal cycle duration at intersections within the digital road network of the digital twin city-level 3D geographic information system.

[0092] Based on the learned probability distribution, the land development agent issues proposed instructions for land use change and intensity adjustment for land parcels that meet the development conditions (marked as land parcels to be developed or updatable, and not located within the permanent basic farmland protection red line or ecological protection red line) during the simulation period.

[0093] Based on the learned balancing strategy, the ecological space intelligent agent issues ecological compensation and spatial optimization constraint instructions to the proposed development plots and existing green spaces during the simulation cycle.

[0094] Collision detection and conflict resolution are performed on adjustment instructions, proposal instructions, and constraint instructions within the unified spatiotemporal framework of the digital twin.

[0095] Collision detection refers to checking whether the overlap of different instructions in space or time leads to conflicts. Conflicts include, but are not limited to: the same plot of land being proposed for two mutually exclusive uses in the same simulation cycle; traffic light adjustment instructions causing cycle mismatch at associated intersections; and ecological constraint instructions and high-density development proposals being physically impossible to satisfy simultaneously.

[0096] Conflict resolution refers to prioritizing or modifying conflicting instructions based on the constraints and weights recorded in the planning intent knowledge graph, thereby generating city status update events.

[0097] The priority ordering rule is: instructions associated with intent atoms with higher weights have higher priority.

[0098] Modification refers to adjusting the parameters of low-priority instructions (such as reducing the floor area ratio or changing the purpose) or postponing their effective time until the conflict is resolved.

[0099] Ideally, by instantiating a multi-agent cluster and endowing it with the ability to learn and co-evolve based on specific models, static and textual planning objectives can be effectively transformed into dynamic, computable, and sustainably optimized decision-making processes. By utilizing models such as deep reinforcement learning, graph neural networks, and multi-objective optimization, traffic timing, land development probability prediction, and ecological layout balancing can be accurately optimized respectively. Furthermore, a planning intent knowledge graph can be used to uniformly guide the concurrent inference of multiple agents in a digital twin, ultimately generating a planning decision instruction set that meets multiple objectives and can be directly executed. This fundamentally changes the traditional planning model that relies on human experience and static solutions, and realizes a paradigm leap to data-driven, real-time simulation, and closed-loop optimization, significantly improving the overall scientific nature of urban construction and management, inter-departmental collaboration, and adaptability to dynamic environments.

[0100] S3.3. Calculate the impact of city state update events on the achievement degree of relevant intention atoms, generate guiding signals, and dynamically adjust the decision-making strategies of each agent, specifically:

[0101] Listen for city status update events and extract the intent atoms affected by the city status update events and their corresponding quantitative indicators from the planning intent knowledge graph.

[0102] By leveraging the real-time computing capabilities of the digital twin, and based on the latest state of the digital twin after the city status update event, the instantaneous values ​​of each affected quantitative indicator are calculated. Taking the public transport modal share as an example, the expression is:

[0103] ;

[0104] in, Indicates the first The instantaneous value of the quantitative indicator corresponding to each intention atom. This indicates the number of trips made using public transportation. This represents the total number of trips made using all modes of transportation.

[0105] The instantaneous value is compared with the expected state value of the corresponding intention atom to calculate the target gap, expressed as:

[0106] ;

[0107] in, Indicates the first The target gap degree of each intentional atom Indicates the first The desired state value (i.e., the target value) of an intention atom.

[0108] Based on the target gap and the weighted relationship between intention atoms, a multi-objective optimization algorithm (such as linear weighting) is used to calculate the reward and penalty values, generate guidance signals, and distribute them to the corresponding agents. The expression is as follows:

[0109] ;

[0110] in, Indicates distribution to intelligent agents The guiding signal (reward / penalty value). Representation and intelligent agent The set of intention atoms related to decision-making behavior Indicates the first The weight of each intention atom recorded in the planning intention knowledge graph. Indicates the first The target gap degree of each intentional atom Indicates the first The negative operation of the difference between the intention atoms (the smaller the difference, the greater the reward or penalty value).

[0111] Each agent receives a guidance signal and combines it with its own inherent reward to form a new reward function, expressed as:

[0112] ;

[0113] in, This represents the total reward of the agent. This represents the inherent reward coefficient (a preset constant, such as 0.1). This represents the agent's inherent reward (e.g., the inherent reward for a traffic flow agent is an increase in the average speed of the road segment).

[0114] Based on the new reward function, and combined with the updated environmental state of the digital twin triggered by the city state update event, policy gradient updates and action value reassessments are performed, specifically as follows:

[0115] For deep reinforcement learning models, the policy gradient method is used to update their policy network parameters. The updated formula is:

[0116] ;

[0117] in, For learning rate, Represents the gradient operator, For the policy function, For action, The state is...

[0118] For graph neural network models, the network parameters are updated using the backpropagation algorithm with the goal of minimizing prediction errors (such as cross-entropy loss).

[0119] For multi-objective optimization models, the weight coefficients of each sub-objective in the objective function are adjusted according to the objective preferences reflected by the new reward function.

[0120] The updated decision strategy is applied to the next simulation cycle, forming a strategy iterative optimization loop. After multiple simulation cycles of strategy iterative optimization, the decision strategy of the agent cluster gradually converges to the Pareto front. Convergence to the Pareto front means that it is no longer possible to improve the target difference of any intention atom without harming the state of other intention atoms. The state prediction data of the agent before instruction execution is recorded synchronously to generate a planning decision instruction set. The planning decision instruction set contains a series of time-series decision instructions. Each instruction includes: instruction type, execution location, execution time, expected effect (i.e., state prediction data), and corresponding intention atom number.

[0121] Ideally, by calculating the quantitative impact of changes in urban state on the achievement of planning intentions in real time, a closed-loop feedback and dynamic adjustment mechanism driving multi-agent collaborative optimization is constructed. By monitoring urban state update events and calculating the target gap, abstract planning objectives are transformed into quantifiable reward and punishment signals. A linear weighting method is used to synthesize multi-objective weights to generate guiding signals, enabling agents to align with global planning intentions while pursuing individual rewards. Furthermore, real-time evolution of multi-agent decision-making strategies is achieved through policy gradient updates, backpropagation, and target weight adjustments. This process forms a continuous optimization loop from state perception and intention evaluation to policy iteration, ensuring that the collaborative decision-making of multi-agents can dynamically converge to the Pareto front of multi-objective equilibrium. Ultimately, a planning decision instruction set that conforms to multiple planning objectives and has high executability is generated, significantly improving the scientific nature, collaboration, and adaptability of urban dynamic planning.

[0122] S4. Verify and serialize the feasibility of the planning decision instruction set, and distribute it to the physical city execution unit. Feed back the execution results to the digital twin for continuous optimization.

[0123] S4.1. Access the real-time constraint information database in the digital twin. The real-time constraint information database is constructed by integrating land engineering geological condition data, permanent basic farmland protection red line data, ecological protection red line data, road design traffic capacity data, and real-time engineering status and equipment health status data obtained from the urban information model platform.

[0124] Based on the real-time constraint information database, the physical realizability, economic cost compliance, and immediate risk level of each instruction are verified, specifically as follows:

[0125] Physical feasibility refers to checking whether the adjustments, proposals, or constraints described in the instruction exist in the physical world and whether there are feasible execution paths and equipment support. For example, for a traffic light timing adjustment instruction, it is necessary to verify whether the signal control hardware of the target intersection supports the new timing scheme (cycle range, phase setting) specified in the instruction; for a land development proposal instruction, it is necessary to verify whether the engineering geological conditions of the target land plot allow the construction of the building type and strength proposed in the instruction.

[0126] Economic cost compliance refers to comparing the estimated costs required to execute an instruction with the annual departmental budget database and project funding limits associated with the digital twin to ensure that the instruction is economically feasible and complies with budget management regulations.

[0127] The immediate risk level refers to the immediate risks that an assessment order may cause within the intended execution time window. For example, assessing the risk level of traffic congestion caused by executing road construction orders during peak hours, or assessing the immediate impact of development and construction in sensitive areas on the ecological environment. The risk level is determined based on industry safety standards and historical accident data.

[0128] S4.2. Adjust instructions that conflict with the real-time constraint information database during the verification process. The principle of adjustment is to modify the parameters of the instructions (such as adjusting the specific values ​​of traffic light timing or reducing development intensity) or postpone the execution time of the instructions, while ensuring that the deviation from the core goal of the planning intent knowledge graph is minimized, so as to eliminate conflicts at the physical, economic or risk level. All instructions that have been verified and adjusted are arranged in the order of execution time to form an executable instruction sequence.

[0129] Each abstract instruction in the instruction sequence is transformed into a standardized control instruction that can be recognized and executed by a specific physical execution unit. For example, traffic light timing adjustment instructions are transformed into standardized messages that conform to the data communication protocol between the traffic signal controller and the host computer; land development supervision instructions are transformed into electronic forms that conform to the data exchange standards of the online planning approval system; and standardized control instructions are securely distributed to the corresponding physical execution units through the city's private network using encrypted transmission and identity authentication mechanisms. The physical execution units include intelligent traffic signal controllers, the online planning approval system, and public facility monitoring platforms.

[0130] S4.3. Start an independent effect tracking thread to continuously receive physical city real-time status monitoring data from the IoT sensor network as effect feedback after command execution.

[0131] The effect tracking thread performs spatiotemporal alignment and comparison between the real state monitoring data and the state prediction data generated by the agent before the instruction is executed. Spatiotemporal alignment refers to matching the timestamp and geographical location of the monitoring data with the spatiotemporal coordinates corresponding to the prediction data.

[0132] After a successful match, for each executed instruction and each quantitative indicator it affects (such as the average vehicle speed on a specific road segment or the PM2.5 concentration in a specific area), the deviation between the actual measured value and the predicted value is calculated. The expression is as follows:

[0133] ;

[0134] in, Indicates the first Prediction-to-reality deviation of each indicator This indicates the first [data point] obtained from the real-state monitoring data. The measured values ​​of each indicator, The first term extracted from the agent's state prediction data represents the... The predicted values ​​of each indicator.

[0135] The deviations of all indicators are summarized to form a prediction-actual deviation matrix.

[0136] S4.4. Using the prediction-reality deviation matrix, the parameters of the internal computational models in the digital twin, such as traffic simulation and environmental diffusion, are calibrated through parameter estimation methods (such as least squares method). Specifically:

[0137] Taking a traffic simulation model as an example, multiple sets of real-state monitoring data (such as road segment flow and speed) over a historical period are used as observations, and the predicted output of the traffic simulation model at the corresponding time is used as the predicted value. An error sum of squares function is constructed with traffic simulation model parameters (such as road capacity coefficient and driver reaction time) as independent variables. The least squares method is used to solve for the parameter estimate that minimizes the function value, thereby updating the traffic simulation model parameters and reducing the systematic bias of future predictions. At the same time, feedback data containing real results (reward and penalty values) and prediction biases are used as new training samples to perform online incremental training on the deep reinforcement learning model, graph neural network model, and multi-objective optimization model in step S3.1, so that the strategy and prediction capabilities can adapt to the dynamic changes of the physical city. Through iterative execution, a continuous closed loop is achieved from digital space decision-making, physical space execution to effect feedback and model optimization, enabling the entire urban construction planning dynamic optimization method to have the ability to continuously learn and self-evolve.

[0138] This embodiment also provides a computer device applicable to the dynamic optimization method of urban construction planning empowered by big data, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the dynamic optimization method of urban construction planning empowered by big data as proposed in the above embodiment.

[0139] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0140] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the dynamic optimization method for urban construction planning empowered by big data as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0141] In summary, this invention constructs a synchronously evolving digital twin by integrating multi-source heterogeneous real-time data streams. This provides a high-fidelity, millisecond-level updated unified spatiotemporal benchmark and decision sandbox aligned with the physical city for the entire optimization process, solving the problem of the simulation environment being disconnected from the real world. By parsing planning text to generate a structured planning intent knowledge graph, abstract planning goals are deconstructed and mapped into quantifiable indicators and target values ​​attached to specific elements of the digital twin, clarifying the relationships, conflicts, and weights among them. This provides precise and computable multi-objective guidance for subsequent optimization. The planning intent knowledge graph drives concurrent simulation and strategy evolution of multiple agents within the digital twin. By calculating the impact of decisions on the degree of achievement of intent atoms in real time and dynamically adjusting the strategies of each agent, collaborative optimization and Pareto front convergence under multi-objective conflicts are achieved.

[0142] Example 2, referring to Table 1, is the second embodiment of the present invention. To further verify the technical solution of the present invention, experimental simulation data of the dynamic optimization method for urban construction planning empowered by big data are provided.

[0143] Using the Eastern New Area, a planned urban new district, as the experimental subject, this study verifies the effectiveness of the proposed dynamic optimization method for urban construction planning driven by multi-agent collaborative optimization based on digital twins and planning intent knowledge graphs. The experiment aims to compare the performance of the "dynamic optimization scheme" generated based on this method with the "baseline scheme" from traditional manual static planning simulation in achieving multiple planning objectives.

[0144] The experimental preparation included: First, deploying an IoT sensor network that connected energy consumption monitoring systems for 50 key buildings in the new district, 100 environmental monitoring stations (monitoring PM2.5, NO2, temperature, and humidity), video streams from 500 traffic checkpoints, and anonymized mobile signaling data from telecom operators. Simultaneously, it integrated 0.5-meter resolution remote sensing imagery, oblique photogrammetry 3D models, geological exploration reports, and the "Eastern New District 2035 Master Plan" text from the new district. This multi-source heterogeneous data underwent real-time cleaning, outlier removal, and UTC timestamp alignment via an edge computing gateway (deployed in the new district's data center) to form a standardized real-time data stream. Subsequently, the standardized data stream was injected into a cloud-native urban information modeling platform. This platform has statically integrated a city-level 3D geographic information system containing all planned building BIM models, underground pipe networks, and 3D geological models. The synchronization engine dynamically maps and binds real-time traffic flow, building energy consumption, environmental indicators, and pedestrian heat maps based on signaling inversion to the corresponding roads, buildings, monitoring points, and grid areas of the 3D model, constructing a digital twin of the Eastern New Area with a synchronization rate of over 95% with the physical world.

[0145] Next, the 2035 master plan text for the Eastern New Area was analyzed. Through dependency parsing and semantic role labeling, five key target statements were extracted as intent atoms, including: A1. "By 2035, the public transportation modal share in the central urban area will increase to 50%", A2. "During the planning period, the per capita green space area in the Eastern New Area will increase by 2 square meters", A3. "By the end of the planning period, the annual average PM2.5 concentration in the New Area will be reduced to below 35 micrograms per cubic meter", A4. "The average vehicle speed during peak hours in the core business district will not be less than 30 kilometers per hour", and A5. "During the planning period, a supply of 500 hectares of new construction land will be guaranteed". These intent atoms were deconstructed into three elements: subject, indicator, and target value, and then mapped in a digital twin. For example, A1's "public transportation modal share" is mapped to road network data and mobile signaling data; A2's "per capita green space area" is mapped to green space and population distribution data in a three-dimensional geographic information system. Spatial overlay analysis revealed a strong spatial competition between intention atoms A2 (new green space) and A5 (new construction land). Based on the planning text, A1 and A3 were labeled as "binding" indicators, while A2, A4, and A5 were labeled as "expected" indicators. Using the analytic hierarchy process (AHP), combined with explicit textual priorities, weights were assigned to the five intention atoms, and constraints such as geological conditions and ecological protection red lines were integrated to construct a weighted planning intention knowledge graph.

[0146] Based on this knowledge graph, three agent clusters are instantiated in the sandbox environment of the digital twin: a traffic flow agent (using the DDPG algorithm), a land development agent (using the GCN model), and an ecological space agent (using the NSGA-II algorithm). Each agent is pre-trained using historical data from the digital twin. Subsequently, the agent clusters are driven to perform asynchronous concurrent simulations for 10 simulation cycles (each cycle representing a quarter). In each cycle, the traffic flow agent issues traffic light optimization instructions based on real-time traffic conditions; the land development agent issues proposals for changing the development type of land parcels; and the ecological space agent issues green space layout and ecological compensation constraints. The instructions undergo collision detection and conflict resolution within a unified spatiotemporal framework, with priority determined by weights in the knowledge graph. For example, when a land parcel is simultaneously proposed for development as a commercial facility (associated with A5) and a park (associated with A2), the system prioritizes satisfying the intent related to the constraint indicator with higher weight (such as the environmental requirements of A3) and adjusts the plot ratio of the development proposal. After each state update, the system calculates the target difference of each intention atom and synthesizes a reward / penalty signal to feed back to the corresponding agent, driving its policy update. After multiple iterations, the agent's policy tends to stabilize, outputting a set of "dynamic optimization scheme" instructions. In contrast, the traditional "baseline scheme" is designed statically by planning experts based on the same initial data, and its execution effect is simulated over the same 10 cycles, without dynamic adjustment or multi-objective collaborative optimization mechanisms. The simulation results of the two schemes are recorded and compared.

[0147] Table 1 Comparison of Dynamic Projection Results of Planning Objectives in the Eastern New Area

[0148]

[0149] The comparison table above clearly demonstrates the significant advantages of the "dynamic optimization scheme" based on the method of this embodiment in achieving multi-objective synergy compared to the traditional "baseline scheme". Data shows that after 10 simulation cycles, the dynamic optimization scheme is closer to the target values ​​in most planning indicators and ultimately achieves a higher overall goal attainment.

[0150] First, this method demonstrates innovation in its ability to achieve dynamic balance and conflict resolution across multiple objectives. Considering the spatially competitive indicators of per capita green space increase and construction land supply, the baseline scheme, lacking a dynamic coordination mechanism, severely squeezed green space (per capita increase of only 1.0 square meter, achieving only 50% of the target) while pursuing construction land supply (reaching 600 hectares in cycle 10, exceeding the target by 20%), resulting in a trade-off. In contrast, the dynamic optimization scheme involves continuous interaction and competition between the land development agent and the ecological space agent guided by the weights of the planning intent knowledge graph. Data curves show that the growth of construction land supply slowed significantly in the later stages (after cycle 7) (finally reaching 480 hectares, close to the target), while per capita green space maintained steady growth (finally reaching 1.9 square meters, achieving 95%). This verifies the effectiveness of the agents in conflict resolution based on knowledge graph weights (with high weights for binding environmental protection objectives), achieving a balance between land development and ecological protection at the Pareto front—something the static scheme could not achieve.

[0151] Secondly, this method demonstrates novelty in its closed-loop feedback and continuous adaptive optimization capabilities. Taking public transport modal share and average vehicle speed in the core area as examples, both are strongly correlated with the decision-making of the traffic flow agent. The curves of the baseline scheme tend to flatten or even deteriorate in the mid-term (vehicle speed decreases from 26.0 km / h in cycle 5 to 24.8 km / h in cycle 10), because the static scheme cannot respond to dynamic changes in traffic flow. In contrast, in the dynamic optimization scheme, the traffic flow agent continuously receives reward and penalty signals transformed from target gaps (such as insufficient modal share and decreased vehicle speed) through deep reinforcement learning, and adjusts the signal timing strategy in real time. The effect is reflected in the continuous and stable optimization trend of both indicator curves (modal share increases from 40.2% to 49.6%, and vehicle speed increases from 25.0 km / h to 29.6 km / h), proving that the closed-loop feedback mechanism based on digital twin simulation can enable the system to continuously tend towards the optimum.

[0152] Furthermore, this method is groundbreaking in terms of global collaboration and maximizing overall benefits. The optimization of PM2.5 concentration (dynamic scheme 35.2 vs baseline scheme 37.5 μg / m³) not only benefits from the direct action of the ecological space intelligent agent, but also indirectly benefits from the traffic flow intelligent agent's improvement of traffic efficiency and reduction of congestion emissions, and the land development intelligent agent's optimization of layout to promote green travel. This cross-agent and cross-objective synergistic effect is facilitated by the core design of unifying dispersed objectives into reward signals for agent co-evolution through the planning intent knowledge graph. The systematic lead in the comprehensive indicator of overall objective achievement (0.854 vs 0.514) strongly demonstrates that this method, through multi-agent collaborative inference, can systematically approach the global optimal solution for multiple objectives, rather than achieving local optimization.

[0153] In summary, the tabular data, from three dimensions—multi-objective balancing, dynamic adaptation, and global collaboration—empirically demonstrates the creativity and superiority of the proposed method compared to traditional static planning simulation. This method achieves precise mapping by constructing a digital twin, forms a computable knowledge graph by parsing planning intentions, and ultimately generates a highly executable planning decision instruction set through continuous learning and game theory among multiple agents in simulation. This achieves a paradigm shift from "static blueprints" to "dynamic system optimization," providing an effective technical path for scientific, refined, and intelligent urban governance.

[0154] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A dynamic optimization method for urban construction planning empowered by big data, characterized in that: include, By integrating real-time data streams from multiple heterogeneous sources, a digital twin that evolves synchronously with the physical city can be constructed. Analyze urban master plan texts to generate a knowledge graph of planning intent; In a digital twin, a planning intent knowledge graph is used to drive multiple agents to perform autonomous simulation and decision-making, and output a set of planning decision instructions. The feasibility of the planning decision instruction set is verified and serialized, and then distributed to the physical city execution units. The execution results are fed back to the digital twin for continuous optimization.

2. The big data-enabled dynamic optimization method for urban construction planning as described in claim 1, characterized in that: The process of integrating multi-source heterogeneous real-time data streams to construct a digital twin that evolves synchronously with the physical city involves the following specific steps. Access to multi-source heterogeneous raw data streams from IoT sensor networks, mobile signaling data, remote sensing data, and public data platforms; By using edge computing nodes, multi-source heterogeneous raw data streams are cleaned and timestamped to generate standardized real-time data streams. Standardized real-time data streams are dynamically integrated into the city information model platform to generate a digital twin that evolves in sync with the physical city.

3. The big data-enabled dynamic optimization method for urban construction planning as described in claim 1, characterized in that: The specific steps for analyzing the urban master plan text and generating a knowledge graph of planning intent are as follows. Analyze urban planning texts and related normative documents to extract key target statements; Deconstruct key target statements into intent atoms containing the main object, quantitative indicators, and desired state; In a digital twin, spatial mapping and data association are performed on the intention atoms, mapping the desired state to the corresponding quantitative indicators and target values; Based on quantitative indicators and target values, the relationships and conflicts between different intention atoms are identified, a dynamic intention network with weights and constraints is constructed, and a planning intention knowledge graph is generated.

4. The big data-enabled dynamic optimization method for urban construction planning as described in claim 1, characterized in that: The specific steps for using a planning intent knowledge graph to drive multi-agent autonomous simulation and decision-making inference are as follows: Based on the planning intent knowledge graph, a cluster of intelligent agents representing different urban functions is instantiated in a digital twin; the cluster of intelligent agents includes traffic flow intelligent agents, land development intelligent agents, and ecological space intelligent agents. The intelligent agent cluster learns behavioral patterns from the historical and real-time data of the digital twin, and uses the quantitative indicators and target values ​​defined in the planning intent knowledge graph as long-term reward signals for co-evolution. Drive the intelligent agent cluster to make asynchronous and concurrent decisions and interactions in the sandbox environment of the digital twin, triggering simulation and prediction of changes in the city's state; The system calculates the impact of changes in city status on the achievement of relevant intention atoms, generates guidance signals and dynamically adjusts the decision-making strategies of each agent, synchronously records the state prediction data of the agents before the execution of instructions, and generates a planning decision instruction set.

5. The big data-enabled dynamic optimization method for urban construction planning as described in claim 4, characterized in that: The intelligent agent cluster learns behavioral patterns from the historical and real-time data of the digital twin, and the specific steps are as follows. A deep reinforcement learning model is constructed for traffic flow intelligent agents. The deep reinforcement learning model is trained using historical traffic flow data and real-time mobile signaling data to learn the mapping relationship between traffic light timing strategies and traffic flow status. A graph neural network model is constructed for the land development agent. The graph neural network model is trained using historical land use change data, land price data and normative document text to learn the probability distribution of land development type conversion. A multi-objective optimization model is constructed for the intelligent agent of ecological space. The model is trained using historical environmental monitoring data and green space data to learn a balance strategy between ecological benefits and spatial constraints.

6. The big data-enabled dynamic optimization method for urban construction planning as described in claim 4, characterized in that: The driving intelligent agent cluster performs asynchronous and concurrent decision-making and interaction in the sandbox environment of the digital twin. The specific steps are as follows. Based on the learned traffic light timing strategy, the traffic flow agent issues adjustment instructions for the traffic light cycle and the ratio of green light time to signal cycle duration at intersections in the digital road network within the simulation cycle; Based on the learned probability distribution, the land development agent issues proposed instructions for changing land use and adjusting intensity for land parcels that meet the development conditions within the simulation period. Based on the learned balancing strategy, the ecological space intelligent agent issues ecological compensation and spatial optimization constraint instructions to the proposed development plots and existing green spaces during the simulation cycle. Collision detection and conflict resolution are performed on adjustment instructions, proposal instructions, and constraint instructions within the unified spatiotemporal framework of the digital twin to generate city status update events.

7. The big data-enabled dynamic optimization method for urban construction planning as described in claim 4, characterized in that: The specific steps for calculating the impact of changes in city status on the achievement of relevant intent atoms are as follows. Listen for city status update events and extract the affected intent atoms and quantitative indicators from the planning intent knowledge graph; By utilizing the real-time computing capabilities of the digital twin, the instantaneous values ​​of each affected quantitative indicator are calculated after a city status update event occurs. The instantaneous value is compared with the expected state value of the corresponding intention atom to calculate the target gap. Based on the target gap and the weight relationship between intention atoms, a multi-objective optimization algorithm is used to calculate the reward and punishment value, generate a guidance signal, and distribute it to the corresponding agent.

8. The big data-enabled dynamic optimization method for urban construction planning as described in claim 4, characterized in that: The specific steps for dynamically adjusting the decision-making strategies of each agent are as follows. Each agent receives a guidance signal and combines it with its own inherent reward to form a new reward function; Based on the new reward function, and combined with the updated environmental state of the digital twin triggered by the city state update event, the policy gradient is updated and the action value is reassessed. The updated decision-making strategy is applied to the next simulation cycle, forming a strategy iterative optimization loop that converges to the Pareto front and outputs a planning decision instruction set.

9. The big data-enabled dynamic optimization method for urban construction planning as described in claim 1, characterized in that: The specific steps for performing executability verification and serialization of the planning decision instruction set are as follows. Load the planning and decision-making instruction set and access the real-time constraint information database in the digital twin. Verify the physical feasibility, economic cost compliance and immediate risk level of each instruction, adjust conflicting instructions, and form an instruction sequence. The verified and adjusted instruction sequence is transformed into standardized control instructions that can be recognized by the execution unit and securely issued to the corresponding physical execution unit.

10. The big data-enabled dynamic optimization method for urban construction planning as described in claim 1, characterized in that: The specific steps for feeding back the execution results to the digital twin for continuous optimization are as follows. Start the effect tracking thread to continuously receive physical city real-time status monitoring data from the IoT sensor network; The real-state monitoring data and the state prediction data of the agent in the digital twin before the execution of the instructions are spatiotemporally aligned and compared to calculate the prediction-reality deviation matrix. The prediction-reality deviation matrix is ​​fed back to the digital twin to calibrate the parameters of the internal simulation model and to perform online incremental training on multiple agents, thereby achieving continuous closed-loop optimization.

Images