A method for urban multivariate spatiotemporal forecasting based on large model agents

Abstract

A method for urban multivariate spatiotemporal prediction based on a large model agent belongs to the field of smart city technology. In order to solve the problem of difficulty in predicting urban operating conditions under incomplete data, the present invention includes defining urban multivariate spatiotemporal prediction as the prediction of a discrete random process, setting the probability of the system state at each moment as the conditional independence hypothesis and the Markov hypothesis; performing state decoupling, decoupling the probability prediction of the system state into the prediction of each channel state given the previous system state; constructing a spatial region, which includes a domain agent layer, a spatiotemporal agent layer, and an individual agent layer; constructing an urban multivariate spatiotemporal prediction system based on a large model agent, including constructing an action space, a state space, a transfer function, a reward function, and a memory flow; and performing task prediction on the constructed urban multivariate spatiotemporal prediction system based on a large model agent. The present invention can improve the accuracy and efficiency of urban management.

Description

Technical Field

[0001] The present invention belongs to the field of smart city technology, and in particular relates to a large model agent-based urban multivariate spatiotemporal prediction method. Background Art

[0002] With the acceleration of urbanization and the rapid development of digital technologies, urban management faces unprecedented challenges and opportunities, such as irrational urban planning, inefficient resource utilization, and traffic congestion. As cities continue to expand in size and become increasingly complex, traditional urban planning and management often rely on empirical evidence, statistical data, and expert opinion, making it difficult to fully and accurately grasp the complexity and diversity of urban development. In recent years, with the continuous maturity of technologies such as artificial intelligence, big data analysis, and multimodal models, there is an urgent need for interpretable reasoning technology driven by both data and knowledge to enable computational applications for urban operations and management.

[0003] Patent application number 202411245264.3, entitled "Large Language Model Training Method for Urban Domains and Urban Generative Intelligence Method and Apparatus," relates to the field of artificial intelligence, and in particular, to large language model training methods and urban generative intelligence methods and apparatus for urban domains. The training method comprises: obtaining a general large language model and an initial dataset; generating a composite dataset by setting up an intelligent agent in a virtual city scene to simulate various human behaviors in a real city scene; combining the composite dataset with the initial dataset to form a pre-training dataset, and using the pre-training dataset to incrementally pre-train the general large language model to obtain a first large city model; fine-tuning the first large city model using a fine-tuning dataset constructed for the urban domain to obtain a second large city model; and performing preference alignment training on the second large city model using a human preference dataset to obtain a target large city model. This results in a target large city model that possesses urban domain expertise, universal common sense, and cognitive reasoning capabilities. This model can address general urban question-answering and city data retrieval for travel navigation, but is essentially a data retrieval interface and lacks complex knowledge analysis and data computation capabilities. Summary of the Invention

[0004] The problem to be solved by the present invention is the difficulty in predicting the urban operation status under the condition of incomplete data. A method for urban multivariate spatiotemporal prediction based on a large model agent is proposed.

[0005] To achieve the above object, the present invention is implemented through the following technical solutions:

[0006] A method for urban multivariate spatiotemporal prediction based on a large model agent includes the following steps:

[0007] S1. Define urban multivariate spatiotemporal forecasting as a discrete random process prediction. To predict the system state at each moment, assume the conditional independence and Markov assumptions for the probability of the system state at each moment, and derive the probability of the system state under these assumptions.

[0008] S2. Decoupling the probability of the system state under the assumptions obtained in step S1, decoupling the probability prediction of the system state into a prediction of each channel state given the previous system state;

[0009] S3. Based on the Markov assumption in step S1 and the state decoupling in step S2, construct a spatial region. The spatial region includes a domain agent layer, a spatiotemporal agent layer, and an individual agent layer.

[0010] S4. Based on the domain agent layer, spatiotemporal agent layer, and individual agent layer constructed in step S3, construct a large-model agent-based urban multivariate spatiotemporal prediction system, including constructing the action space, state space, transition function, reward function, and memory flow.

[0011] S5. Perform task predictions on the large-model agent-based urban multivariate spatiotemporal prediction system constructed in step S4. All decision events are recorded in the memory stream, along with feedback with reward values. Before taking the next action, each agent reads relevant information from the event stream and uses the transfer function to perform rolling predictions on the action strategy. This optimizes and updates the large-model agent-based urban multivariate spatiotemporal prediction system.

[0012] Furthermore, the specific implementation method of step S1 includes the following steps:

[0013] S1.1. Define urban multivariate spatiotemporal prediction as a discrete random process prediction. The goal of the urban multivariate spatiotemporal prediction task is to predict the system state at each moment. Define is a random process, ,in, yes The sample space of yes The state space of is the set of time steps, , It is a moment;

[0014] For the The system state random variable at the moment is The value in ,in, for The number of channels in for The dimension of each channel in ;

[0015] The system state at each moment is modeled as a The joint distribution of the future values of the random variables of the system state at time and before is expressed as:

[0016] ;

[0017] in, For is the probability density function of the parameter, for The N Channels:

[0018] S1.2. Set the conditional independence assumption to assume that given the historical state of the system, the next state of a channel is independent of the other channels, that is:

[0019] ;

[0020] in, is the probability density function, for The n channels, n for N Any one of;

[0021] S1.3. Assuming the Markov assumption, the complete conditional probability is simplified to the transition probability between the last two states, which is expressed as:

[0022] .

[0023] Furthermore, the specific implementation method of step S2 is to decouple the probability prediction of the system state into the prediction of each channel state given the previous system state under the assumption that:

[0024] .

[0025] Furthermore, the specific implementation method of step S3 includes the following steps:

[0026] S3.1. Constructing the domain agent layer: In the initial stage of the urban spatiotemporal state prediction task, set up the first i The action space of the domain agent is ;

[0027] Define all spatial regions It is a vector space composed of directed unweighted adjacency matrices of nodes, which is used to represent the relationship between the spatial regions of domain agents. For the j A spatial area, ;

[0028] S3.2. Constructing the individual agent layer: For j spatial regions Each individual agent in the and The status of k The action space of each individual agent is , is the number of dimensions of individual decision making, The number of discretization intervals for individual decisions;

[0029] S3.3. Constructing the spatiotemporal proxy layer: For j spatial regions The corresponding space-time agent observes and The decision results of all individual agents in the j The action space of a spatiotemporal agent is , is the number of predictor variables.

[0030] Furthermore, the specific implementation method of step S4 includes the following steps:

[0031] S4.1. Constructing the Action Space: Action Space is the possible action combination of all agents in a large model agent-based urban multivariate spatiotemporal prediction system within one time step, expressed as:

[0032] ;

[0033] in, 、 、 Represent the action space composed of all domain agents, spatiotemporal agents and individual agents, 、 、 Represent the total number of domain agents, spatiotemporal agents, and individual agents, respectively, where each spatial region corresponds to a spatiotemporal agent, i.e. ;

[0034] S4.2. Constructing the State Space: State Space Represents integrated information about the urban environment, including individual locations, quantities, and overall layout of urban areas, given by images, vectors, and natural language descriptions;

[0035] Set the At this moment The space-time state The expression is:

[0036] ;

[0037] ;

[0038] in, For the The moment j The facility layout map corresponding to each spatial area is represented by a multi-channel image; For the The moment j A vector of statistical variables corresponding to each spatial region, including regional size, regional center location, number of residents, number of employed people, energy consumption, GDP, and number of enterprises; For the The moment j The text description corresponding to the spatial region, Represents the set of all strings, Represents a character;

[0039] Therefore, the The system state at a moment is expressed as , all Constructing the state space , all agents operate in a state space and take different actions at different times according to different environments;

[0040] S4.3. Constructing the transfer function:

[0041] Transfer function , according to the state space and action space, the action is converted into state; for the individual agent action space and the spatiotemporal agent action space , Statistics for all regions The individual changes in the Add together to get the new system state ;

[0042] S4.4. Constructing the reward function:

[0043] S4.4.1. The goal of setting individual agents is to make the strategy more favorable to the individual agent after the system state changes. Considering the cost of living, commuting costs, and the convenience of facilities, the reward function of the individual agent is constructed. , the expression is:

[0044] ;

[0045] in, For living costs, For commuting costs, For the cost of facilities, 、 、 The incentive coefficients for living cost, commuting cost, and facility convenience cost respectively;

[0046] S4.4.2. The goal of setting up a spatiotemporal agent is to make the predicted state at each time step as close as possible to the actual state, so as to achieve the purpose of accurate spatiotemporal prediction. Consider the local state vector to construct the reward function of the spatiotemporal agent , the expression is:

[0047] ;

[0048] in, and Represent the predicted value and actual value of the local state vector at a time step, is the reward coefficient of local fitness, is the 2-norm;

[0049] Each dimension of the local state vector includes the number of residents, the number of employed people, the amount of energy consumption, the amount of GDP, and the number of enterprises;

[0050] S4.4.3. The goal of setting the domain agent is to limit the sparsity of the adjacency matrix while maintaining the global fit, and obtain the reward function of the domain agent , the expression is:

[0051] ;

[0052] in, is the 0 norm, is the total predicted value of the relevant variables under the domain agent, is the total actual value of the relevant variables under the domain agent, is the adjacency matrix, is the reward coefficient for sparsity, is the reward coefficient of global fitness;

[0053] S4.5. Construct a memory stream, including an event stream and a perception stream. The event stream records events that occur over time, including the agent's actions, external intervention events, and environmental changes. The perception stream records the agent's tracking and understanding of events in the event stream. Each node in the perception stream is linked to one or more nodes in the event stream.

[0054] Furthermore, the specific implementation method of step S5 includes the following steps:

[0055] S5.1. Given the current system state First, each domain agent gives its current action by observing the state. The action set of the domain agent layer is , similarly, the action sets of the individual agent layer and the spatiotemporal agent layer are and ;

[0056] S5.2. State transition function Convert the action set of the domain agent layer, the action set of the individual agent layer, and the action set of the spatiotemporal agent layer in step S5.1 into the new system state at the next moment ;

[0057] S5.3. All agents are fed from the memory stream according to Get their respective rewards at every moment, adjust their action strategy at the next moment, and thus Make new actions to complete the rolling prediction.

[0058] Beneficial effects of the present invention:

[0059] The large-scale model agent-based multivariate spatiotemporal urban forecasting method described in this invention can improve the accuracy and efficiency of urban management. By integrating and analyzing multi-source data on urban transportation, environment, population, and economy, the invention can monitor urban operations in real time, promptly identify potential problems, and provide accurate decision support. For example, in traffic management, the model can predict traffic flow, optimize signal timing, reduce traffic congestion, and improve traffic efficiency.

[0060] The present invention describes a multivariate spatiotemporal urban forecasting method based on a large-scale model agent, enabling optimized resource allocation and planning. The model analyzes various urban regions, identifies imbalances in resource distribution, and assists in developing rational resource allocation plans. In urban planning, this method can simulate the impact of different planning options on urban development, helping decision makers select the optimal solution and enhancing the sustainability of urban development.

[0061] The large-scale model agent-based multivariate spatiotemporal urban forecasting method described in this invention can enhance emergency response and risk management. When an emergency or natural disaster occurs, the model can quickly assess the impact, predict potential subsequent developments, assist in the development of emergency plans, and improve the speed and effectiveness of emergency response. For example, in flood warnings, this method can predict the spread of floodwaters, allowing for early evacuation and minimizing losses.

[0062] The large-scale agent-based multivariate spatiotemporal urban forecasting method described in this invention can promote the development of smart cities, providing data support and decision-making basis for such projects. By deeply mining various urban data, this invention can provide technical support for smart transportation, smart healthcare, smart security, and other fields, thereby enhancing the intelligent level of urban services.

[0063] The large-scale model agent-based multivariate spatiotemporal urban forecasting method described in this paper can support policy formulation and evaluation. It can simulate the impact of different policies on urban development, assisting policymakers in evaluating policy effectiveness and optimizing policy design. For example, when formulating environmental protection policies, the model can predict changes in air quality after policy implementation and assess policy effectiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 This is a structural diagram of a large model agent-based urban multivariate spatiotemporal prediction method according to the present invention;

[0065] Figure 2 This is a flow chart of a large model agent-based urban multivariate spatiotemporal prediction method described in the present invention. DETAILED DESCRIPTION

[0066] In order to make the objectives, technical solutions, and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain the present invention and are not intended to limit the present invention. That is, the specific embodiments described herein are only some embodiments of the present invention, not all embodiments. Generally, the components of the specific embodiments of the present invention described and illustrated in the drawings herein can be arranged and designed in various different configurations, and the present invention can also have other embodiments.

[0067] Therefore, the following detailed description of the specific embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention as claimed, but is merely representative of selected specific embodiments of the present invention. All other specific embodiments obtained by those skilled in the art based on the specific embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0068] In order to further understand the content, features and effects of the present invention, the following specific embodiments are given as examples, and the attached Figure 1 -Attached Figure 2 The detailed instructions are as follows:

[0069] Example 1:

[0070] A method for urban multivariate spatiotemporal prediction based on a large model agent includes the following steps:

[0071] S1. Define urban multivariate spatiotemporal forecasting as a discrete random process prediction. To predict the system state at each moment, assume the conditional independence and Markov assumptions for the probability of the system state at each moment, and derive the probability of the system state under these assumptions.

[0072] Furthermore, the specific implementation method of step S1 includes the following steps:

[0073] S1.1. Define urban multivariate spatiotemporal prediction as a discrete random process prediction. The goal of the urban multivariate spatiotemporal prediction task is to predict the system state at each moment. Define is a random process, ,in, yes The sample space of yes The state space of is the set of time steps, , It is a moment;

[0074] For the The system state random variable at the moment is The value in ,in, for The number of channels in for The dimension of each channel in ;

[0075] The system state at each moment is modeled as a The joint distribution of the future values of the random variables of the system state at time and before is expressed as:

[0076] ;

[0077] in, For is the probability density function of the parameter, for The N Channels:

[0078] Furthermore, the distribution drift problem becomes more prominent as the dimension increases, making parameter estimation of such high-dimensional distributions infeasible in practical calculations. To address this problem, the urban multivariate spatiotemporal prediction system based on large model agents is based on the following two assumptions:

[0079] S1.2. Set the conditional independence assumption to assume that given the historical state of the system, the next state of a channel is independent of the other channels, that is:

[0080] ;

[0081] in, is the probability density function, for The n channels, n for N Any one of;

[0082] S1.3. Assuming the Markov assumption, the complete conditional probability is simplified to the transition probability between the last two states, which is expressed as:

[0083] ;

[0084] S2. Decoupling the probability of the system state under the assumptions obtained in step S1, decoupling the probability prediction of the system state into a prediction of each channel state given the previous system state;

[0085] Furthermore, the specific implementation method of step S2 is to decouple the probability prediction of the system state into the prediction of each channel state given the previous system state under the assumption that:

[0086] ;

[0087] Furthermore, the determined goal (i.e., the probabilistic prediction of the system state at each moment) is decoupled into a prediction of the state of each channel (i.e., each spatial region) given the previous system state. The urban multivariate spatiotemporal prediction system based on a large model agent in this embodiment is suitable for solving the probabilistic prediction process proposed in this inference. In other words, the prediction of the system state probability distribution (specifically, the probability density function) is decoupled into a prediction of the probability distribution of the state of each spatial region at the next moment given the previous system state.

[0088] Furthermore, the decoupled system state prediction task is transformed into a reinforcement learning task: the system state probability prediction task is modeled as a Markov decision process (MDP), that is, an interaction process between a group of agents and the system environment. These agents dynamically adjust their strategies to enhance their prediction capabilities by evaluating the difference between the current state and the desired goal. The optimization task of this process is usually called a reinforcement learning task, which is represented as a tuple , which contains the state space , action space , transfer function , and the reward function The swarm agent observes the "state" (the current state of each spatial region), takes "action" (their respective prediction) at each step, and receives "rewards" issued by the environment according to the desired goal. The environment "transfers" the state (the state of each spatial region changes) according to the agent's actions.

[0089] S3. Based on the Markov assumption in step S1 and the state decoupling in step S2, construct a spatial region. The spatial region includes a domain agent layer, a spatiotemporal agent layer, and an individual agent layer.

[0090] Furthermore, the large-model agent-based urban multivariate spatiotemporal prediction system designed in this embodiment consists of three sets of collaborative agent models: a domain agent layer, a spatiotemporal agent layer, and an individual agent layer, each operating at a different spatial scale. The agent models in the domain agent layer act as "domain experts," while the agent models in the individual agent layer simulate the various behaviors of heterogeneous individuals (people, vehicles, corporate entities, etc.) in different environments, such as travel, consumption, social interaction, and production. The agent models in the spatiotemporal agent layer use observed states to predict the state of various variables in the region (such as energy consumption and population changes) at the next moment.

[0091] Furthermore, the specific implementation method of step S3 includes the following steps:

[0092] S3.1. Constructing the domain agent layer: In the initial stage of the urban spatiotemporal state prediction task, set up the first i The action space of the domain agent is ;

[0093] Define all spatial regions It is a vector space composed of directed unweighted adjacency matrices of nodes, which is used to represent the relationship between the spatial regions of domain agents. For the j A spatial area, ;

[0094] Furthermore, the domain agent layer predicts spatial associations; for example, for individual migration or travel behavior prediction, the decision result of the "traffic domain agent" is a certain directed unweighted adjacency matrix , for any region , available Zhongyu Other areas associated with Regions with strong OD (origin-destination) associations, where individual agents and spatiotemporal agents only observe a subset , thus greatly reducing the complexity of the state space. The number of domain agents The setting of is variable, and the present invention does not discuss its value.

[0095] S3.2. Constructing the individual agent layer: For j spatial regions Each individual agent in the and The status of k The action space of each individual agent is , is the number of dimensions of individual decision making, The number of discretization intervals for individual decisions;

[0096] Furthermore, the individual agent layer predicts the individual strategy. k The action space of an individual agent is a dimension A vector space where the decision of each dimension is discretized between 1 and between, is the number of uniform discretized intervals. For example, if the first dimension is migration to other areas, then the decision is The fourth dimension is consumption, and its consumption value is classified into 1 and Between, and so on. Discretization interval The setting of is variable, and the present invention does not discuss its value. The configuration is scalable. The decision dimensions implemented by the present invention include, but are not limited to, residence selection, workplace selection, and travel mode preference selection. Further decision dimensions under this method are also protected by this patent.

[0097] S3.3. Constructing the spatiotemporal proxy layer: For j spatial regions The corresponding space-time agent observes and The decision results of all individual agents in the j The action space of a spatiotemporal agent is , is the number of predictor variables;

[0098] Furthermore, the spatiotemporal agent layer predicts the spatiotemporal state of the system. j The action space of a spatiotemporal agent is a dimension The vector space is the total number of The statistical values of the variables, including but not limited to the number of residents, the number of employed people, the number of energy consumption, the number of GDP, and the number of enterprises. It is scalable and more predictor variables under this method are also protected by this patent.

[0099] S4. Based on the domain agent layer, spatiotemporal agent layer, and individual agent layer constructed in step S3, construct a large-model agent-based urban multivariate spatiotemporal prediction system, including constructing the action space, state space, transition function, reward function, and memory flow.

[0100] Furthermore, the specific implementation method of step S4 includes the following steps:

[0101] S4.1. Constructing the Action Space: Action Space is the possible action combination of all agents in a large model agent-based urban multivariate spatiotemporal prediction system within one time step, expressed as:

[0102] ;

[0103] in, 、 、 Represent the action space composed of all domain agents, spatiotemporal agents and individual agents, 、 、 Represent the total number of domain agents, spatiotemporal agents, and individual agents, respectively, where each spatial region corresponds to a spatiotemporal agent, i.e. ;

[0104] Furthermore, the global action space is the combination of all possible actions of all agents in one time step;

[0105] S4.2. Constructing the State Space: State Space Represents integrated information about the urban environment, including individual locations, quantities, and overall layout of urban areas, given by images, vectors, and natural language descriptions;

[0106] Set the At this moment The space-time state The expression is:

[0107] ;

[0108] ;

[0109] in, For the The moment j The facility layout map corresponding to each spatial area is represented by a multi-channel image; For the The moment jA vector of statistical variables corresponding to each spatial region, including regional size, regional center location, number of residents, number of employed people, energy consumption, GDP, and number of enterprises; For the The moment j The text description corresponding to the spatial region, Represents the set of all strings, Represents a character;

[0110] Therefore, the The system state at a moment is expressed as , all Constructing the state space , all agents operate in a state space and take different actions at different times according to different environments;

[0111] S4.3. Constructing the transfer function:

[0112] Transfer function , according to the state space and action space, the action is converted into state; for the individual agent action space and the spatiotemporal agent action space , Statistics for all regions The individual changes in the Add together to get the new system state ;

[0113] Furthermore, in the new system state, each agent can have a new prediction for the next moment, thus achieving rolling predictions for the current moment, the next moment, the next moment, and so on.

[0114] S4.4. Constructing the reward function:

[0115] Furthermore, the reward function is the key to improving the agent's autonomous learning and optimization. In this design, three levels of metrics are designed: individual demand, local fitness, and global fitness.

[0116] S4.4.1. The goal of setting individual agents is to make the strategy more favorable to the individual agent after the system state changes. Considering the cost of living, commuting costs, and the convenience of facilities, the reward function of the individual agent is constructed. , the expression is:

[0117] ;

[0118] in, For living costs, For commuting costs, For the cost of facilities, 、 、 The incentive coefficients for living cost, commuting cost, and facility convenience cost respectively;

[0119] S4.4.2. The goal of setting up a spatiotemporal agent is to make the predicted state at each time step as close as possible to the actual state, so as to achieve the purpose of accurate spatiotemporal prediction. Consider the local state vector to construct the reward function of the spatiotemporal agent , the expression is:

[0120] ;

[0121] in, and Represent the predicted value and actual value of the local state vector at a time step, is the reward coefficient of local fitness, is the 2-norm;

[0122] Each dimension of the local state vector includes the number of residents, the number of employed people, the amount of energy consumption, the amount of GDP, and the number of enterprises;

[0123] S4.4.3. The goal of setting the domain agent is to limit the sparsity of the adjacency matrix while maintaining the global fit, and obtain the reward function of the domain agent , the expression is:

[0124] ;

[0125] in, is the 0 norm, is the total predicted value of the relevant variables under the domain agent, is the total actual value of the relevant variables under the domain agent, is the adjacency matrix, is the reward coefficient for sparsity, is the reward coefficient of global fitness;

[0126] Furthermore, for example, the variables that the “transportation sector agent” is responsible for include the number of residents, the number of employed people, and the amount of energy consumption.

[0127] S4.5. Construct a memory stream, consisting of an event stream and a perception stream. The event stream records events that occur over time, including the agent's actions, external interventions, and environmental changes. The perception stream records the agent's tracking and understanding of events in the event stream. Each node in the perception stream is linked to one or more nodes in the event stream.

[0128] Furthermore, unlike traditional reinforcement learning, the large-scale agent-based multivariate spatiotemporal urban prediction system uses a "memory stream" to record all decision events and feedback with reward values. Before taking its next action, each agent reads relevant information from the event stream to update its action strategy, ultimately optimizing the entire agent system. The memory stream tracks events and provides perception over time. It consists of two types of memory streams: the event stream and the perception stream. The event stream records events that occur over time, including the agent's actions, external interventions, and environmental changes. These events are recorded in chronological order. The perception stream records the agent's tracking and understanding of events in the event stream. Each node in the perception stream is linked to one or more nodes in the event stream, reflecting how the agent perceives or reacts to a specific event. The agent's behavior is driven by its current state, which in turn influences its decision-making process and actions taken. By leveraging the memory stream, the agent adjusts its behavior over time in a manner that reflects human cognition and expert decision-making, laying the foundation for evolutionary-based spatiotemporal urban prediction.

[0129] S5. Perform task predictions on the large-model agent-based urban multivariate spatiotemporal prediction system constructed in step S4. All decision events are recorded in the memory stream, along with feedback with reward values. Before taking the next action, each agent reads relevant information from the event stream and uses the transfer function to perform rolling predictions on the action strategy. This optimizes and updates the large-model agent-based urban multivariate spatiotemporal prediction system.

[0130] Furthermore, the specific implementation method of step S5 includes the following steps:

[0131] S5.1. Given the current system state First, each domain agent gives its current action by observing the state. The action set of the domain agent layer is , similarly, the action sets of the individual agent layer and the spatiotemporal agent layer are and ;

[0132] S5.2. State transition function Convert the action set of the domain agent layer, the action set of the individual agent layer, and the action set of the spatiotemporal agent layer in step S5.1 into the new system state at the next moment ;

[0133] S5.3. All agents are fed from the memory stream according to Get their respective rewards at every moment, adjust their action strategy at the next moment, and thus Make new actions to complete the rolling prediction.

[0134] It should be noted that relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus comprising the element.

[0135] Although the present application has been described above with reference to specific embodiments, various modifications may be made thereto and components may be substituted with equivalents without departing from the scope of the present application. In particular, as long as there are no structural conflicts, the various features of the embodiments disclosed herein may be combined with each other in any manner, and the omission of an exhaustive description of these combinations in this specification is solely for the sake of space and resource conservation. Therefore, the present application is not limited to the specific embodiments disclosed herein, but includes all technical solutions within the scope of the claims.

Claims

1. A method for urban multivariate spatiotemporal prediction based on a large model agent, characterized in that: The steps include: S1. Define urban multivariate spatiotemporal forecasting as the prediction of a discrete random process. To predict the system state at each moment, assume the conditional independence hypothesis and the Markov hypothesis for the probability of the system state at each moment, and obtain the probability of the system state under the assumed conditions. The specific implementation method of step S1 includes the following steps: S1.

1. Define urban multivariate spatiotemporal prediction as a discrete random process prediction. The goal of the urban multivariate spatiotemporal prediction task is to predict the system state at each moment. Define X as a random process, X: Where Ω is the sample space of X, is the state space of X, is the set of time steps, t is the tth moment; X t is the system state random variable at the tth moment, from The value in Where N is The number of channels in , D is The dimension of each channel in ; The system state at each moment is modeled as a joint distribution based on the future values of the system state random variables observed at and before the tth moment, and the expression is: Among them, p Θ is the probability density function with Θ as parameter, For X t+1 The Nth channel in ; S1.

2. Set the conditional independence assumption to assume that given the historical state of the system, the next state of a channel is independent of other channels, that is: Where p is the probability density function, For X t+1 The nth channel in , where n is any one of N; S1.

3. Assuming the Markov assumption, the complete conditional probability is simplified to the transition probability between the last two states, which is expressed as: p(X t+1 |X t ,X t-1 ,…,X1)=p(X t+1 |X t ); S2. Decouple the probability of the system state under the assumptions obtained in step S1, and decouple the probability prediction of the system state into a prediction of each channel state given the previous system state; S3. Based on the Markov assumption in step S1 and the state decoupling in step S2, a spatial region is constructed. The spatial region includes a domain agent layer, a spatiotemporal agent layer, and an individual agent layer. S4. Based on the domain agent layer, spatiotemporal agent layer, and individual agent layer constructed in step S3, a large-model agent-based urban multivariate spatiotemporal prediction system is constructed, including constructing an action space, a state space, a transfer function, a reward function, and a memory stream. S5. Perform task prediction on the large-model agent-based urban multivariate spatiotemporal prediction system constructed in step S4. Record all decision events and feedback with reward values through the memory stream. Each agent reads relevant information from the event stream before taking the next action, uses the transfer function to perform rolling predictions on the action strategy, and optimizes and updates the large-model agent-based urban multivariate spatiotemporal prediction system.

2. The urban multivariate spatiotemporal prediction method based on a large model agent according to claim 1 is characterized in that: The specific implementation method of step S2 is to decouple the probability prediction of the system state into the prediction of each channel state given the previous system state under the assumption that:

3. The urban multivariate spatiotemporal prediction method based on a large model agent according to claim 2 is characterized in that: The specific implementation method of step S3 includes the following steps: S3.

1. Constructing the domain agent layer: In the initial stage of the urban spatiotemporal state prediction task, the action space of the i-th domain agent is set to Define all spatial regions is a vector space composed of directed unweighted adjacency matrices of nodes, which is used to represent the relationship between the spatial regions of domain agents. j is the jth spatial region, S3.

2. Constructing individual agent layer: For the j-th spatial region Each individual agent in the j and The state of the kth individual agent is set to D ind is the number of dimensions of individual decision making, D seg The number of discretization intervals for individual decisions; S3.

3. Constructing spatiotemporal proxy layer: For the j-th spatial region The corresponding space-time agent observes v j and The decision results of all individual agents in the space are set as the action space of the jth spatiotemporal agent D st is the number of predictor variables.

4. The urban multivariate spatiotemporal prediction method based on a large model agent according to claim 3 is characterized in that: The specific implementation method of step S4 includes the following steps: S4.

1. Constructing the Action Space: Action Space is the possible action combination of all agents in a large model agent-based urban multivariate spatiotemporal prediction system within one time step, expressed as: in, Represents the action space composed of all domain agents, spatiotemporal agents and individual agents, N dom 、N st 、N ind Represent the total number of domain agents, spatiotemporal agents, and individual agents, respectively. Each spatial region corresponds to a spatiotemporal agent, i.e., N st =N; S4.

2. Constructing the state space: state space Represents integrated information about the urban environment, including individual locations, quantities, and overall layout of urban areas, given by images, vectors, and natural language descriptions; Set the t-th moment The space-time state The expression is: in, The facility layout map corresponding to the jth spatial region at the tth moment is represented by a multi-channel image; is a vector of statistical variables corresponding to the jth spatial region at the tth moment, including the region size, the location of the region center, the number of residents, the number of employed people, the amount of energy consumption, the amount of GDP, and the number of enterprises; is the text description corresponding to the j-th spatial region at the t-th moment, Σ * Represents the set of all strings, Represents a character; Therefore, the system state at the tth moment is expressed as All Constructing the state space All agents operate in a state space and take different actions at different times depending on the environment; S4.

3. Constructing the transfer function: Transfer function According to the state space and action space, the action is converted into state; for the individual agent action space and the spatiotemporal agent action space Statistics for all regions The individual changes in the Add together to get the new system state S4.

4. Constructing reward function: S4.4.

1. The goal of setting individual agents is to make the strategy more favorable to the individual agent after the system state changes. Considering the cost of living, commuting costs, and the convenience of facilities, the reward function of the individual agent is constructed. The expression is: Among them, Cost res Cost of living od Cost of commuting service is the facility convenience cost, r res 、r od 、r service The incentive coefficients for living cost, commuting cost, and facility convenience cost respectively; S4.4.

2. The goal of setting up a spatiotemporal agent is to make the predicted state at each time step as close as possible to the actual state, so as to achieve the purpose of accurate spatiotemporal prediction. Consider the local state vector to construct the reward function of the spatiotemporal agent The expression is: Among them, s (st) and Represent the predicted value and actual value of the local state vector at a time step, r local is the reward coefficient of local fitness, ||·||2 is the 2-norm; Each dimension of the local state vector includes the number of residents, the number of employed people, the amount of energy consumption, the amount of GDP, and the number of enterprises; S4.4.

3. The goal of setting the domain agent is to limit the sparsity of the adjacency matrix while maintaining the global fit, and obtain the reward function of the domain agent The expression is: Among them, ||·||0 is the zero norm, s (dom) is the total predicted value of the relevant variables under the domain agent, is the total actual value of the relevant variables under the domain agent, is the adjacency matrix, r sparse is the reward coefficient for sparsity, r global is the reward coefficient of global fitness; S4.

5. Construct a memory stream, including an event stream and a perception stream. The event stream records events that occur over time, including the agent's actions, external intervention events, and environmental changes; the perception stream records the agent's tracking and understanding of events in the event stream. Each node in the perception stream is linked to one or more nodes in the event stream.

5. The urban multivariate spatiotemporal prediction method based on a large model agent according to claim 4 is characterized in that: The specific implementation method of step S5 includes the following steps: S5.

1. Given the current system state s t First, each domain agent gives its current action by observing the state. The action set of the domain agent layer is Similarly, the action sets of the individual agent layer and the spatiotemporal agent layer are and S5.

2. State transition function Convert the action set of the domain agent layer, the action set of the individual agent layer, and the action set of the spatiotemporal agent layer in step S5.1 into the new system state s at the next moment t+1 ; S5.

3. All agents obtain their respective rewards from the memory stream according to time t and adjust their action strategies for the next moment, so that in the new system state s t+1 Make new actions to complete the rolling prediction.

Images