An urban emergency evacuation decision system and method based on digital twinning and deep learning

Abstract

The application relates to the technical field of emergency management, and provides a city emergency evacuation decision system and method based on digital twinning and deep learning. The decision system comprises a multi-source heterogeneous data sensing and fusion layer, a high-fidelity city-level emergency evacuation digital twinning layer, a multi-agent deep reinforcement learning decision engine embedded with uncertainty, a collaborative control instruction execution and feedback layer, and a system management and man-machine interaction interface electrically connected with the aforementioned layers. The decision system can output an environment state with uncertainty measurement through credibility evaluation and conflict resolution of multi-source heterogeneous data, generate multiple sets of future evolution scenario copies relying on a high-fidelity simulation environment, complete safety verification of multi-agent collaborative decision and control instructions, and realize the invention purpose of real-time collaborative optimization of multi-modal control instructions in city emergency evacuation. In addition, the application also provides a method applied to the aforementioned decision system.

Description

Technical Field

[0001] This application relates to the technical field of emergency management, and in particular to an urban emergency evacuation decision-making system and method based on digital twins and deep learning. Background Technology

[0002] With the continuous advancement of urbanization in my country, the complexity of urban road networks and the density of population are constantly increasing, which also raises the requirements for the refinement and intelligence of emergency evacuation decision-making. Related emergency evacuation decision support technologies have been widely studied and applied in the industry. The technology has also evolved from the traditional static emergency plan management model to a dynamic auxiliary decision-making model that integrates multi-source data perception, virtual simulation and intelligent algorithms, becoming an important technical support for the construction of smart city public safety.

[0003] Existing technologies for urban emergency evacuation decision-making first acquire multi-source heterogeneous data related to disaster development, traffic operation, and population distribution through fixed sensors, mobile data collection devices, and public information platforms. Probabilistic statistical models are then used to fuse and process this multi-source data, completing a risk assessment of the on-site situation and outputting corresponding risk assessment results. Simultaneously, existing technologies construct virtual simulation environments corresponding to the evacuation scenarios, synchronizing the virtual environment with the physical site using real-time data. Based on preset evacuation rules or offline optimization algorithms, different evacuation plans are simulated and their effects evaluated to select recommended evacuation schemes. Furthermore, existing technologies incorporate machine learning algorithms, modeling traffic signal control and route planning processes in evacuation scenarios as sequential decision-making processes. Through continuous interaction and iterative optimization of control strategies between algorithms and the simulation environment, evacuation time is shortened and traffic efficiency during evacuation is improved.

[0004] The implementation process of existing technologies reveals that they operate independently across various stages, including data perception and risk assessment, evacuation plan planning, and control command execution, failing to form an organic linkage and closed loop. Environmental uncertainty information obtained during risk assessment can only serve as a reference for plan planning and cannot be deeply integrated into the core decision-making optimization process. Virtual simulation environments can only be used for offline evaluation of contingency plans and cannot provide continuous interactive support for online optimization of decision-making algorithms. Furthermore, related algorithm optimizations are mostly targeted at simplified, single evacuation scenarios, making it difficult to adapt to the collaborative optimization needs of multiple control units in large-scale urban evacuation scenarios. Based on these shortcomings, the core problem with existing technologies is their inability to achieve collaborative optimization of the entire urban emergency evacuation process, making it difficult to guarantee the real-time nature, robustness, and adaptability of emergency evacuation decisions. Summary of the Invention

[0005] In view of this, this application aims to propose an urban emergency evacuation decision-making system and method based on digital twins and deep learning to solve the above-mentioned technical problems.

[0006] To achieve the above objectives, the technical solution of this application is implemented as follows: In the first aspect, this application proposes an urban emergency evacuation decision-making system based on digital twins and deep learning, including a multi-source heterogeneous data perception and fusion layer that communicates with each other, a high-fidelity urban-level emergency evacuation digital twin layer, a multi-agent deep reinforcement learning decision engine embedded with uncertainty, a collaborative control command execution and feedback layer, and a system management and human-computer interaction interface that is electrically connected to each of the aforementioned layers. The multi-source heterogeneous data sensing and fusion layer is used to collect multi-source heterogeneous emergency-related data from the physical world, complete the data credibility assessment and conflict resolution, and output a probabilistic environmental state with uncertainty measurement. The high-fidelity city-level emergency evacuation digital twin layer is used to receive probabilistic environmental states to synchronize the virtual environment with the physical world, construct a high-fidelity emergency evacuation simulation environment, and generate multiple sets of future evolution scenario copies covering uncertainties. The embedded uncertainty multi-agent deep reinforcement learning decision engine is used to receive probabilistic environmental states and scenario copies of the digital twin layer, complete multi-agent collaborative decision-making for risk perception, and output multimodal collaborative control commands. The collaborative control command execution and feedback layer is used to receive multimodal collaborative control commands, and after completing the "sandbox" pre-execution and security verification through the digital twin layer, it sends the control commands to the execution system in the physical world and collects execution feedback data to send back to the multi-source heterogeneous data perception and fusion layer. The system management and human-computer interaction interface is used to realize global situation monitoring, decision intervention, system configuration, and visualization of decision results.

[0007] Furthermore, the multi-source heterogeneous data perception and fusion layer includes a social media semantic understanding and spatiotemporal localization submodule based on a large language model and a dynamic graph structure Bayesian network processing module. The social media semantic understanding and spatiotemporal localization submodule based on a large language model has a built-in spatiotemporal regularization layer with a gated graph neural network, which is used to extract events and locate spatiotemporally in social media texts, and output spatiotemporal event tuples with confidence and localization error. The input of the dynamic graph structure Bayesian network processing module is connected to the physical sensor acquisition device and the output of the social media semantic understanding and spatiotemporal positioning submodule, respectively. It is used to perform credibility weighting, conflict resolution and probabilistic reasoning on multi-source heterogeneous data, and output probabilistic environmental states with probability distribution.

[0008] Furthermore, the high-fidelity city-level emergency evacuation digital twin layer includes a 3D scene modeling module, a physical simulation module, a dynamic environment model learner, and a scenario generator. The 3D scene modeling module uses neural radiation fields to complete high-fidelity 3D reconstruction of key areas; The physical simulation module employs a physical information neural network for simulating the physical constraints of disaster spread and human behavior. The dynamic environment model learner is used to fine-tune the simulation model parameters online based on real-time observation data, so that the digital twin environment can closely approximate the evolution of the real world. The scenario generator, based on a model-independent meta-learning framework, is used to generate training scenarios ranging from easy to difficult, as well as multiple sets of future evolution scenario replicas covering uncertainty.

[0009] Furthermore, the multi-agent deep reinforcement learning decision engine with embedded uncertainty adopts a centralized training and distributed execution paradigm, including a hierarchical multi-agent execution unit, an uncertainty embedding unit, an action generation unit based on a diffusion model, and a multi-agent communication and coordination unit. The hierarchical multi-agent execution unit includes an upper-layer regional collaborative agent and a lower-layer local execution agent. The output of the upper-layer regional collaborative agent is connected to the input of the lower-layer local execution agent to realize hierarchical collaborative decision-making. The uncertainty embedding unit is used to generate a multidimensional uncertainty state vector and inject it into the policy network and value function. The value function has a built-in conditional risk value risk sensitivity measure. The diffusion-based action generation unit employs a time-conditional U-Net denoising architecture to generate high-dimensional continuous control commands that conform to physical constraints. The multi-agent communication and collaboration unit adopts a Transformer communication encoder and has a built-in counterfactual multi-agent policy gradient credit allocation module to realize information interaction and collaborative optimization among agents.

[0010] Secondly, this application proposes a method for applying the aforementioned urban emergency evacuation decision-making system based on digital twins and deep learning, comprising the following steps: S1, Emergency Event Triggering and Situation Awareness: After an emergency event is triggered, real-time multi-source heterogeneous data is processed through the multi-source heterogeneous data perception and fusion layer, and a probabilistic environmental state St with uncertainty measurement is output. S2, digital twin environment synchronization and multi-scenario generation, receives probabilistic environmental state St through a high-fidelity city-level emergency evacuation digital twin layer to complete state synchronization and generate N copies of future evolution scenarios covering uncertainties; S3, multimodal collaborative decision generation, is a decision engine that embeds a multi-agent deep reinforcement learning engine with embedded uncertainty. It completes collaborative decision-making based on the probabilistic environment state St and the multidimensional uncertain state vector Ut. After parallel inference through multiple scenario replicas, it outputs multimodal collaborative control command At. S4, Command Security Verification and Execution: The multimodal collaborative control command At is pre-executed and security verified in a "sandbox" through the digital twin layer. After verification, it is sent to the physical world execution system. S5, closed-loop feedback and strategy optimization, collects real-world feedback data after instruction execution, constructs experience tuples and stores them in the experience replay buffer, and completes online fine-tuning and offline deep training of the decision engine.

[0011] Furthermore, in step S3, the generation and injection of the multidimensional uncertain state vector Ut specifically includes: The scalar uncertainty features in the probabilistic environment state St are aggregated by a graph attention network to generate a hierarchical multidimensional uncertainty state vector Ut. The multidimensional uncertain state vector Ut is concatenated with the probabilistic environment state St and then input into the policy network. By introducing conditional risk value (CVaR) into the value function, the agent's optimization objective is set to maximize CVaRα(R) under a preset risk confidence level, where R is the cumulative reward, thus realizing the deep injection of uncertainty into policy optimization.

[0012] Furthermore, step S1 specifically includes: S11, through the social media semantic understanding and spatiotemporal localization submodule based on a large language model, performs named entity recognition, event extraction and spatiotemporal regularization processing on social media text, and outputs spatiotemporal event tuples with confidence and localization error; S12 inputs the spatiotemporal event tuples and multi-source data collected by physical sensors into a dynamic graph structure Bayesian network, and completes the confidence weighting and conflict resolution of multi-source data through the belief propagation algorithm. S13 outputs the probabilistic environment state St with probability distribution for each state variable.

[0013] Furthermore, the "sandbox" pre-execution and security verification in step S4 specifically includes: S41, load the multimodal cooperative control command At into the digital twin environment, and use a traffic flow prediction model based on graph convolutional networks to deduce the traffic and risk status at the next T time steps after the command is executed; S42, verify whether the risk indicators of the simulation results are within the preset safety threshold. If all meet the requirements, the verification is passed; if they exceed the threshold, the verification is rejected or a correction instruction is given. S43 standardizes and converts verified instructions, then distributes them to traffic signal control systems, information dissemination platforms, and emergency command platforms.

[0014] Furthermore, step S5 specifically includes: S51: Collect new real-world observation data after the command is executed, combine it with the preset target to generate a reward signal rt and the new environmental state St+1, construct an experience tuple (St,At,rt,St+1) and store it in the experience replay buffer. S52, in the online fine-tuning stage, a soft update strategy is adopted to iteratively update the policy network and value network parameters of the decision engine to achieve real-time policy adaptation. S53, the offline deep training phase, during non-emergency periods, utilizes historical data and simulation data generated by digital twins to complete deep training of the model using distributed priority experience replay and importance sampling, and optimizes the policy library.

[0015] Furthermore, it also includes steps to enhance decision interpretability, specifically: During the decision-making process, similar historical scenarios and decision-making experiences from the knowledge graph of the retrieval strategy are embedded through TransE graph similarity calculation as decision priors; For key decision points, the Monte Carlo tree search module is invoked to complete online deep planning, supplementing the neural network strategy; Through the counterfactual explanation and comparative analysis module, counterfactual deductions of different decision-making schemes are completed based on structural causal models and do-calculus, and the differences of key indicators are visualized to enhance the transparency of decision-making.

[0016] Compared with existing technologies, the urban emergency evacuation decision-making system and method based on digital twins and deep learning proposed in this application have the following advantages: (1) This application achieves the effect of orderly linkage and collaborative operation of the entire process of urban emergency evacuation from situational awareness, environmental simulation, decision generation to command execution and feedback by setting up a multi-source heterogeneous data perception and fusion layer, a high-fidelity city-level emergency evacuation digital twin layer, a multi-agent deep reinforcement learning decision engine with embedded uncertainty, a collaborative control command execution and feedback layer, and a system management and human-computer interaction interface that are electrically connected to the aforementioned layers.

[0017] (2) This application sets up a social media semantic understanding and spatiotemporal positioning submodule based on a large language model and a dynamic graph structure Bayesian network processing module in the multi-source heterogeneous data perception and fusion layer, sets up a three-dimensional scene modeling module, a physical simulation module, a dynamic environment model learner and a scenario generator in the high-fidelity city-level emergency evacuation digital twin layer, and sets up a hierarchical multi-agent execution unit, an uncertainty embedding unit, an action generation unit based on a diffusion model and a multi-agent communication and coordination unit in the multi-agent deep reinforcement learning decision engine with embedded uncertainty. This achieves the effects of effective fusion and uncertainty quantification of multi-source heterogeneous data, stable construction of a high-fidelity emergency evacuation simulation environment and efficient execution of multi-agent collaborative decision-making.

[0018] (3) This application achieves the effects of real-time adaptive adjustment, robustness improvement and decision-making process enhancement in emergency evacuation decision-making by executing the steps of emergency event triggering and situational awareness, digital twin environment synchronization and multi-scenario generation, multimodal collaborative decision generation, instruction security verification and issuance execution, execution feedback and strategy optimization. At the same time, it completes the generation and injection of uncertain state vectors, sandbox pre-execution and security verification of control instructions, online fine-tuning and offline training of policy networks, and decision interpretability enhancement processing during the decision-making process. Attached Figure Description

[0019] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of the workflow of the urban emergency evacuation decision-making system based on digital twins and deep learning described in the embodiments of this application; Figure 2 This is a schematic diagram of the architecture of the urban emergency evacuation decision-making system based on digital twins and deep learning described in an embodiment of this application; Figure 3 This is a structural diagram of the UI-MARL decision engine described in the embodiments of this application; Figure 4 A comparison chart of different schemes in terms of comprehensive operational indicators; Figure 5 A comparison chart of different solutions in terms of safety-related indicators; Figure 6 A comparison chart of average travel times under different event types. Detailed Implementation

[0020] To make the technical solution and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0021] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0022] Furthermore, it should be noted that in the description of this application, if terms such as "upper," "lower," "inner," or "outer" appear, indicating orientation or positional relationship, these are based on the orientation or positional relationship shown in the accompanying drawings and are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation on this application. In addition, if terms such as "first" or "second" appear, they are also used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0023] Furthermore, in the description of this application, unless otherwise expressly defined, the terms "installation," "connection," "joining," and "connector" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this application in light of the specific circumstances.

[0024] In this application, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to a specific feature, structure, material, or characteristic described in connection with that embodiment or example, which is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0025] The present application will now be described in detail through exemplary embodiments. However, it should be understood that, without further description, elements, structures, and features in one embodiment may be advantageously incorporated into other embodiments.

[0026] This application relates to the field of smart city and emergency management technology, specifically to a city emergency evacuation decision-making system and method based on digital twins and deep learning.

[0027] Urban emergency evacuation is a key challenge in the field of public safety, and its decision-making process involves the rapid perception, assessment, and response to multi-source information such as disaster dynamics, transportation networks, and crowd behavior. Traditional methods are mostly based on static plans and rule-driven approaches, which are difficult to cope with the spatiotemporal evolution of disasters and the complex uncertainties of crowd behavior. With the development of technologies such as digital twins and artificial intelligence, related fields have begun to explore more intelligent solutions, aiming to build an emergency response system that coordinates perception, decision-making, and control. However, current technologies still have many limitations. Existing emergency evacuation schemes based on digital twins are mostly limited to simulation evaluation in closed scenarios, lacking the ability to coordinate and optimize with multi-modal control commands such as traffic signals, guidance information, and zone control in urban road networks in real time, making it difficult to achieve adaptive dynamic control for rapid response. Although multi-source information fusion methods based on Bayesian networks can quantify uncertainty, their output is usually isolated from downstream dynamic path planning and collaborative decision-making modules, failing to organically inject uncertainty measurement into and constrain the final optimization decision-making process, resulting in insufficient overall robustness of the decision. Existing research on traffic control based on deep reinforcement learning mostly focuses on simplified scenarios or single intelligent agents, which is difficult to cope with the high-dimensional, continuous action space collaborative optimization problem composed of multiple groups and departments in large-scale evacuation. In addition, there are challenges in engineering implementation such as low sample efficiency, difficulty in policy interpretability and implementation of safety constraints.

[0028] Example 1 To achieve real-time, robust, and adaptive collaborative optimization of multimodal control commands during urban emergency evacuation, the urban emergency evacuation decision-making system based on digital twins and deep learning in this embodiment is configured to include a multi-source heterogeneous data perception and fusion layer interconnected with each other, a high-fidelity urban-level emergency evacuation digital twin layer, a multi-agent deep reinforcement learning decision engine embedding uncertainty, a collaborative control command execution and feedback layer, and a system management and human-computer interaction interface electrically connected to each of the aforementioned layers. (Refer to...) Figure 2 The architecture of the urban emergency evacuation decision-making system based on digital twins and deep learning described in this application embodiment is divided into five layers from top to bottom. Each layer interacts with the others through standardized data interfaces and communication protocols to form an organic whole.

[0029] Based on the invention's objective of effectively collecting and processing multi-source heterogeneous emergency-related data, in this embodiment, a multi-source heterogeneous data perception and fusion layer is used to collect multi-source heterogeneous emergency-related data from the physical world, complete data credibility assessment and conflict resolution, and output a probabilistic environmental state with uncertainty measurement. The multi-source heterogeneous data perception and fusion layer includes a social media semantic understanding and spatiotemporal localization submodule based on a large language model and a dynamic graph structure Bayesian network processing module. The social media semantic understanding and spatiotemporal localization submodule based on a large language model incorporates a spatiotemporal regularization layer with a gated graph neural network, used for event extraction and spatiotemporal localization of social media text, outputting spatiotemporal event tuples with confidence and localization error.

[0030] To this end, this embodiment designs a spatiotemporal regularization layer that associates ambiguous location descriptions in the text (such as "near the east gate of the stadium") with precise coordinates in the urban geographic information system (GIS), and outputs a spatiotemporal event tuple (event, locationlat, lon, time, c, ε) with confidence level c and location error range ε. This output serves as evidence input to the aforementioned DGS-BN module. Specifically, the spatiotemporal regularization layer contains a trainable graph neural network (GNN) for modeling the spatial topological relationships between urban geographic entities (such as roads, buildings, and points of interest). This GNN adopts a gated graph neural network (GGNN) structure, and its node update formula is:

[0031] Where hv(t) is the hidden state of node v at time step t, N(v) is the set of neighbors of node v, W is the learnable weight matrix, and GRU is a gated recurrent unit. Through this GNN, the location descriptions extracted from the text are mapped to graph nodes and accurately located by combining contextual information.

[0032] The input of the dynamic graph structure Bayesian network processing module is connected to both the physical sensor acquisition device and the output of the social media semantic understanding and spatiotemporal localization submodule. It is used to perform credibility weighting, conflict resolution, and probabilistic inference on multi-source heterogeneous data, outputting a probabilistic environmental state with a probability distribution. The social media semantic understanding and spatiotemporal localization submodule, based on a large language model, employs a finely tuned BERT or GPT architecture to perform named entity recognition, sentiment analysis, and event extraction on social media text. The spatiotemporal regularization layer contains a trainable gated graph neural network used to model the spatial topological relationships between urban geographic entities. It maps location descriptions extracted from the text to corresponding geographic nodes and combines contextual information to achieve precise localization. The update of the hidden state of nodes in the gated graph neural network at each time step is completed by a gated recurrent unit combining the hidden states of neighboring nodes. The hidden state of a node is calculated by inputting the hidden state of the node at the previous time step and the hidden states of neighboring nodes after transformation by a learnable weight matrix into the gated recurrent unit, thereby achieving effective modeling of the spatial topological relationships of urban geographic entities.

[0033] The Dynamic Graph Bayesian Network (DGBN) processing module models urban road networks, sensor nodes, and information sources as a graph structure. Nodes represent state variables, and edges represent causal relationships or information dependencies. The network structure can be dynamically adjusted based on sudden information events. For conflicting information from different sources, the module performs weighted fusion and conflict resolution by calculating the historical credibility weights of each information source and the difference between the current information and the posterior probability inferred by the network. The inference process of the DGBN is implemented using a belief propagation algorithm. During the iteration process, information interaction between nodes is achieved through message passing. The edge probability distribution of nodes is iteratively updated on the dynamic graph, and finally, the probability distribution corresponding to each state variable is output. The probability distribution includes the mean and variance of the state variable, with the variance representing the degree of uncertainty of the corresponding information. The number of iterations of the belief propagation algorithm is set to 3 to 10. This range ensures the convergence accuracy of the node edge probability distribution on the graph structure while meeting the real-time requirements of emergency scenarios, avoiding response delays caused by excessive iterations.

[0034] To achieve state synchronization between the virtual environment and the physical world and construct a high-fidelity emergency evacuation simulation environment, this embodiment uses a high-fidelity city-level emergency evacuation digital twin layer to receive probabilistic environmental states, synchronize the virtual environment with the physical world, construct a high-fidelity emergency evacuation simulation environment, and generate multiple sets of future evolution scenario replicas covering uncertainties. The high-fidelity city-level emergency evacuation digital twin layer includes a 3D scene modeling module, a physical simulation module, a dynamic environment model learner, and a scenario generator.

[0035] The 3D scene modeling module uses neural radiation fields to complete high-fidelity 3D reconstruction of key areas. The neural radiation field model establishes a mapping relationship through a multilayer perceptron, taking spatial position and viewing direction as inputs and mapping outputs as corresponding colors and volume densities. Then, through the volume rendering process, it synthesizes scene images from the corresponding perspective, thereby achieving ultra-high-fidelity 3D geometry and appearance reconstruction of key areas.

[0036] For key areas such as transportation hubs and large venues, the reconstruction resolution of the 3D scene modeling module is set to a range of 1080P to 4K. This range ensures the accuracy of scene geometry and appearance restoration while controlling the computational load of the model, meeting the requirements of real-time rendering. The physics simulation module employs a physical information neural network (PIN) for simulating the physical constraints of disaster spread and crowd behavior. The loss function of the PIN consists of a data fitting error term and a physical constraint residual term. The overall loss is the weighted sum of the data fitting error and the physical constraint residual term using a weighted coefficient. The physical constraint residual term is the square of the difference between the physical field predicted by the differential operator and the source term. By embedding this residual term in the loss function, it ensures that the simulation results both meet the fitting requirements of the observed data and adhere to the corresponding physical control laws. The weighted coefficients for the data fitting error and the physical constraint residual term in the loss function are set to a range of 0.1 to 10. This range can be flexibly adjusted according to the richness of observed data for different disaster scenarios, increasing the fitting weight when data is abundant and strengthening the physical constraint when data is sparse.

[0037] Specifically, the NeRF model maps spatial location x and observation direction d to color c and volume density σ using a multilayer perceptron (MLP) FΘ: (x, d) → (c, σ), and synthesizes images through volume rendering. Simultaneously, for disaster physics processes (such as fluid diffusion), a Physical Information Neural Network (PINN) is introduced. PINN ensures that the simulation results conform to both the data and physical laws by embedding the residual term Lphysics = |N(u) - f|² of the governing equations (such as the Navier-Stokes equations) into the loss function, where N is the differential operator, u is the physical field predicted by the network (such as velocity, concentration), and f is the source term. PINN's network architecture typically employs a fully connected network with an adaptive activation function, and its loss function is:

[0038] Where Ldata is the fitting error with the observed data, and λ is the weighting coefficient.

[0039] A dynamic environment model learner is used to fine-tune simulation model parameters online based on real-time observation data, making the digital twin environment closely approximate the evolution of the real world. For crowd behavior models, the dynamic environment model learner employs a meta-model based on a long short-term memory network to predict an individual's movement decisions in a given environmental state. The hidden state of the meta-model is calculated and updated jointly by the individual's perceived local environmental state, the previous action, and the model parameters. The model parameters are iteratively optimized through online gradient descent, thereby achieving continuous adaptation to changes in real-world crowd behavior. The online learning rate of the model parameters is set to a range of 0.0001 to 0.001, which ensures smooth parameter updates, avoids model oscillations caused by real-time data fluctuations, and enables rapid adaptation to changes in the real world.

[0040] Specifically, for the crowd behavior model, we use a meta-model based on a Long Short-Term Memory (LSTM) network to predict an individual's movement decision in a given environmental state, and its hidden state update is as follows:

[0041] Where st represents the local environmental state perceived by the individual, at-1 represents the action at the previous time step, and θ represents the model parameters, which are updated through online gradient descent.

[0042] The scenario generator, based on a model-independent meta-learning framework, generates training scenarios ranging from easy to difficult, along with multiple sets of future evolution scenario replicas covering uncertainty. Built upon this framework, the inner loop rapidly adapts model parameters for each historical disaster task through one or more steps of gradient descent. The outer loop calculates the meta-loss based on the adapted parameters from the inner loop, updating the meta-model parameters. This allows the meta-model to quickly learn adaptation strategies from a limited amount of new disaster scenario data, generating simulation scenarios for corresponding new disasters. Simultaneously, it automatically constructs training scenario sequences from easy to difficult, improving the training efficiency of multi-agent deep reinforcement learning. The number of future evolution scenario replicas generated by the scenario generator is set to 3 to 20 sets. This range effectively covers the various possible futures brought about by environmental uncertainty while controlling the computational load of parallel inference, ensuring the responsiveness of the decision-making process.

[0043] Based on the invention's objective of achieving risk-aware multi-agent collaborative decision-making and outputting multimodal collaborative control commands, this embodiment employs a multi-agent deep reinforcement learning decision engine embedded with uncertainty. This engine receives probabilistic environmental states and scenario copies from a digital twin layer, completes risk-aware multi-agent collaborative decision-making, and outputs multimodal collaborative control commands. The multi-agent deep reinforcement learning decision engine embedded with uncertainty adopts a centralized training and distributed execution paradigm, including a hierarchical multi-agent execution unit, an uncertainty embedding unit, an action generation unit based on a diffusion model, and a multi-agent communication and coordination unit.

[0044] Reference Figure 3 The UI-MARL decision engine structure diagram described in this application clearly illustrates the internal structure and data flow of the decision engine, realizing a complete processing flow from state input to collaborative control command output. The hierarchical multi-agent execution unit includes an upper-level regional collaborative agent and a lower-level local execution agent. The output of the upper-level regional collaborative agent is connected to the input of the lower-level local execution agent to achieve hierarchical collaborative decision-making. The upper-level regional collaborative agents are divided according to administrative regions or functional areas, with each agent responsible for generating macro-level evacuation commands for one administrative region or functional area. The lower-level local execution agents correspond to specific control units, including intersection traffic lights and variable message signs, used to generate specific high-dimensional continuous control commands. The uncertainty embedding unit is used to generate multi-dimensional uncertainty state vectors and inject them into the policy network and value function. The value function incorporates conditional risk value and risk sensitivity metrics.

[0045] The uncertainty embedding unit aggregates scalar uncertainty features in probabilistic environmental states using a graph attention network, generating hierarchical, multidimensional uncertainty state vectors. During aggregation, for the uncertainty vector of a corresponding region, the uncertainty features of each road segment within that region are transformed using a learnable weight matrix. The transformed features are then weighted and summed using attention coefficients, which are obtained by concatenating the overall uncertainty features of the region with those of each road segment, followed by linear transformation and activation function calculation, and then normalized. Simultaneously, a conditional risk value is introduced into the value function as a risk sensitivity measure. The conditional risk value is the expected value of the accumulated reward within the range less than or equal to the risk value. The agent's optimization objective is set to maximize the conditional risk value at a preset risk confidence level, thereby forcing the agent to proactively avoid potential losses caused by high uncertainty during decision-making. The risk confidence level is set to a range of 0.9 to 0.99, which can adapt to different levels of emergency scenarios, improving the confidence level in high-risk disaster scenarios and strengthening the strategy's ability to avoid extreme risks. The action generation unit based on the diffusion model employs a temporal conditional U-Net denoising architecture to generate high-dimensional continuous control commands that conform to physical constraints.

[0046] Specifically, for example, for segment-level uncertainty, it is aggregated using a graph attention network (GAT): For region R, its uncertainty vector UR is calculated as follows:

[0047] Where u l The uncertainty characteristics of road segment l are represented by W, which is the weight matrix, and α is the weight matrix. R,l It is the attention coefficient, expressed as α R,l =(exp(LeakyReLU(a T [W u R || Wu l ]))) / (∑ k ∈ R) exp(LeakyReLU(aT[W uR || W uk])) Calculate where 'a' is the attention vector, and || denotes concatenation. U t With the original state S t After being spliced ​​together, they together serve as the policy network π(\( a t \) | \( s t \), u t The input is γ. Simultaneously, a risk-sensitive metric, such as Conditional Value at Risk (CVaR), is introduced into the value function design. The agent's optimization objective is to maximize the expected cumulative reward E[∑γ]. t \( r t This transforms into maximizing the expected value of the worst-case (tail) return under a certain risk confidence level α, i.e., maximizing CVaR. α (R), where R is the cumulative reward, CVaR α (R) = E[R | R ≤ VaR α (R)],VaR α It is the value at risk. This forces intelligent agents to proactively avoid the potentially huge losses caused by high uncertainty when making decisions.

[0048] In the reverse diffusion process of the action generation unit based on the diffusion model, each step of the action is calculated by combining the action from the previous diffusion step, the noise predicted by the denoising network, and random noise. The denoising network takes the embedding of the diffusion step, the concatenation result of the environmental state and the uncertain state vector as input, and outputs the predicted noise. Through a multi-step denoising diffusion process, the initial noisy action is gradually optimized into the final control action that conforms to physical constraints, thereby achieving effective exploration of the high-dimensional continuous action space and high-quality action generation. The total number of steps in the diffusion process is set to 50 to 200. This range can control the computation time of the denoising process while ensuring the quality of action generation, meeting the real-time requirements of emergency decision-making.

[0049] Specifically, the agent's policy network first outputs a noisy action, and then refines it into a final action through a denoising diffusion process. This process effectively explores the action space and generates high-quality actions. The core steps of its reverse diffusion process can be represented as follows:

[0050] Among them, a t k This is the diffusion action at step k, ε θ This is the trained denoising network, where z is noise and α is noise. k , β k , α k , σ k These are diffusion scheduling parameters. Denoising network ε θ The architecture employs a time-conditional U-Net, whose inputs are the embeddings of diffusion step k and the state s. t and uncertainty u t The concatenation of these elements produces the predicted noise output.

[0051] The multi-agent communication and coordination unit employs a Transformer communication encoder, which incorporates a counterfactual multi-agent policy gradient credit allocation module to achieve information exchange and collaborative optimization among agents. Using the Transformer communication encoder, each agent concatenates its observation code with an uncertainty vector, performs a linear transformation to obtain three sets of vectors: query, key, and value. Then, a multi-head attention mechanism is used to calculate the relevance weights with other agents. The value vectors are then weighted and summed to obtain the corresponding agent's communication message, thus achieving effective information exchange among agents.

[0052] Specifically, for agent i, its communication message mi is calculated as follows:

[0053] Q i = W Q [h i || u i ], K j = W K [h j || u j ], V j = W V [h j || u j ], h i It is the observation code of agent i, u i It is its uncertainty vector, d kThis refers to the key dimension. During centralized training, methods such as Counterfactual Multi-Agent Policy Gradient (COMA) are used for credit allocation to accurately evaluate the contribution of each agent's action. COMA maintains a counterfactual advantage function A for each agent i. i (s, a) = Q(s, a) - ∑ a' i π i (a' i |τ i ) Q(s, (a -i , a' i )), where a -i This represents the actions of other agents, thereby isolating the contribution of agent i's actions.

[0054] Simultaneously, a counterfactual multi-agent policy gradient credit allocation module is built-in. This module maintains a counterfactual advantage function for each agent. The counterfactual advantage function is obtained by subtracting the expected action value function of the agent when taking other actions from the action value function corresponding to the global state and joint actions. This isolates the contribution of individual agent actions to the global effect, achieving accurate credit allocation. The Transformer communication encoder has 4 to 16 attention heads. This range ensures sufficient information interaction between agents while controlling the model's computational complexity, adapting to the collaborative needs of large-scale agents at the city level.

[0055] To ensure secure verification and reliable execution of control commands, the collaborative control command execution and feedback layer in this embodiment receives multimodal collaborative control commands. After sandbox pre-execution and security verification through a digital twin layer, it issues the control commands to the execution system in the physical world and collects execution feedback data, which is then transmitted back to the multi-source heterogeneous data perception and fusion layer. The collaborative control command execution and feedback layer incorporates a traffic flow prediction model based on graph convolutional networks for rapid deduction during command pre-execution. The inter-layer propagation of the graph convolutional network-based traffic flow prediction model is achieved by multiplying the normalized results of the adjacency matrix and degree matrix (including self-connected ones) with the features of the upper-layer nodes and the trainable weight matrix, followed by activation function calculation. This enables rapid and accurate prediction of traffic flow states, completing the rapid deduction of the command pre-execution process. The deduction time domain of the traffic flow prediction model is set to 5 to 30 time steps, with each time step lasting 5 to 30 seconds. This range covers the evolution of traffic flow after command execution, accurately assessing potential risks while ensuring efficient rapid deduction.

[0056] Specifically, the rapid extrapolation uses a simplified traffic flow prediction model based on graph convolutional networks (GCNs), with the following inter-layer propagation rules:

[0057] Where à = A + I is the adjacency matrix with self-connection, D is its degree matrix, and H... (l) It is the node feature of layer l, W (l) σ is the trainable weight, and σ is the activation function.

[0058] Based on the invention's objective of achieving global situational awareness monitoring, decision intervention, system configuration, and visualization of decision results, this embodiment electrically connects the system management and human-computer interaction interface to the multi-source heterogeneous data perception and fusion layer, the high-fidelity city-level emergency evacuation digital twin layer, the multi-agent deep reinforcement learning decision engine embedded with uncertainty, and the collaborative control command execution and feedback layer. This interconnection enables global situational awareness monitoring, decision intervention, system configuration, and visualization of decision results. The system management and human-computer interaction interface includes a built-in counterfactual explanation and comparative analysis module. This module, based on structural causal models and DO calculus, performs counterfactual deductions of different decision-making schemes, quantifies the differences in key performance indicators under different decisions, and then visualizes these differences through a comparative dashboard, thereby enhancing the transparency of the decision-making process and the human-computer collaboration capability.

[0059] Reference Figure 4 The comparison chart of different schemes in terms of comprehensive operational indicators shows the comparison results between the system of this application and the existing technical solutions in terms of comprehensive operational indicators such as average travel time, residual queue length, and evacuation completion rate, verifying the optimization effect of the system of this application in terms of evacuation efficiency. (Refer to...) Figure 5 The comparison charts of different solutions on safety-related indicators show the results of comparing the proposed system with existing technologies in terms of risk exposure, proportion of unsafe paths, and average number of safety violations, verifying the proposed system's ability to control safety risks during evacuation. (Refer to...) Figure 6The comparison chart of average travel time under different event types shows the performance of the proposed system in terms of average travel time under different event types such as floods, smoke, and compound disasters, verifying the adaptability and optimization effect of the proposed system in various emergency scenarios. Through the above system structure, the multi-source heterogeneous data perception and fusion layer can effectively process multi-source heterogeneous emergency data, complete conflict resolution and uncertainty quantification, and solve the problem of the separation between perception and evaluation links in the prior art; the high-fidelity city-level emergency evacuation digital twin layer can build a high-fidelity simulation environment synchronized with the real world, providing a reliable interactive and verification basis for the decision-making process, and solving the problem of the disconnect between the simulation environment and the real control link in the prior art; the multi-agent deep reinforcement learning decision engine embedded with uncertainty can realize multi-agent collaborative decision-making in large-scale city-level scenarios, deeply injecting the uncertainty quantification results into the decision optimization process, and solving the problems of the separation between uncertainty processing and decision optimization and the difficulty in implementing high-dimensional collaborative decision-making in the prior art; the collaborative control command execution and feedback layer can complete the security verification and execution feedback of commands, ensuring the safe and reliable execution of control commands, and ultimately achieving the effect of real-time, robust, adaptive collaborative optimization of multi-modal control commands in the process of urban emergency evacuation.

[0060] Example 2 Based on the same inventive concept as the aforementioned system embodiments, this embodiment provides a city emergency evacuation decision-making method based on digital twins and deep learning. This method is applied to the city emergency evacuation decision-making system based on digital twins and deep learning described in the aforementioned embodiments. (Refer to...) Figure 1 The flowchart of the urban emergency evacuation decision-making system based on digital twins and deep learning described in this application illustrates the complete execution steps of the method.

[0061] To ensure the complete and orderly operation of the emergency evacuation decision-making process, the urban emergency evacuation decision-making method based on digital twins and deep learning in this embodiment includes the following steps: Step S1, Emergency Event Triggering and Situational Awareness: After the emergency event is triggered, real-time multi-source heterogeneous data is processed through a multi-source heterogeneous data perception and fusion layer to output a probabilistic environmental state with uncertainty measurement. Based on the invention objective of improving the accuracy of multi-source data processing and uncertainty quantification capabilities, in this embodiment, Step S1 specifically includes S11, using a social media semantic understanding and spatiotemporal positioning submodule based on a large language model to perform named entity recognition, event extraction, and spatiotemporal regularization processing on social media text, outputting spatiotemporal event tuples with confidence and positioning error; S12, inputting the spatiotemporal event tuples and multi-source data collected by physical sensors into a dynamic graph structure Bayesian network processing module, and completing the confidence weighting and conflict resolution of multi-source data through a confidence propagation algorithm; S13, outputting a probabilistic environmental state with probability distribution for each state variable. In step S1, the number of iterations of the belief propagation algorithm is set to 3 to 10. This range can ensure the convergence accuracy of probabilistic inference and meet the real-time response requirements in emergency scenarios.

[0062] Step S2 involves digital twin environment synchronization and multi-scenario generation. A high-fidelity city-level emergency evacuation digital twin layer receives probabilistic environmental states to achieve state synchronization, generating multiple sets of future evolution scenario replicas covering uncertainties. In Step S2, the number of generated future evolution scenario replicas is set to 3 to 20 sets. This range can fully cover the various evolutionary possibilities brought about by environmental uncertainties, while controlling the computational load of parallel simulations to ensure the response speed of the decision-making process.

[0063] Step S3, multimodal collaborative decision generation, utilizes a multi-agent deep reinforcement learning decision engine embedded with uncertainty to complete collaborative decision-making based on probabilistic environmental states and multidimensional uncertain state vectors. After parallel extrapolation through multiple scenario replicas, multimodal collaborative control commands are output. Based on the invention's objective of deeply injecting uncertainty into the decision optimization process to enhance decision robustness, in this embodiment, step S3 specifically includes the generation and injection of multidimensional uncertain state vectors. This involves aggregating scalar uncertainty features in the probabilistic environmental states using a graph attention network to generate hierarchical multidimensional uncertain state vectors; concatenating the multidimensional uncertain state vectors with the probabilistic environmental states and inputting them into the policy network; and introducing conditional risk value into the value function, setting the agent's optimization objective to maximize the conditional risk value at a preset risk confidence level. The conditional risk value is the expected value of the accumulated reward within a range less than or equal to the risk value, thus achieving deep injection of uncertainty into policy optimization. In step S3, the risk confidence level is set to a range of 0.9 to 0.99, which can be flexibly adjusted according to the disaster risk level, strengthening the avoidance of extreme risks in high-risk scenarios. For key decision points, step S3 can also call the Monte Carlo tree search module to complete online deep planning, supplementing the neural network strategy. The number of Monte Carlo tree search simulations is set to 100 to 1000, which can balance search depth and decision-making time, meeting the real-time requirements of emergency scenarios. During the decision-making process, similar historical scenarios and decision-making experiences from the knowledge graph of the retrieval strategy can also be embedded through the TransE graph as decision priors, improving decision-making efficiency and reliability.

[0064] Step S4 involves the security verification and execution of the multimodal collaborative control commands. These commands are pre-executed and security verified in a sandbox using a digital twin layer. Once verified, they are sent to the physical world execution system. Based on the invention's objective of ensuring the security of control command execution, in this embodiment, the sandbox pre-execution and security verification in step S4 specifically includes: S41, loading the multimodal collaborative control commands into the digital twin environment and using a traffic flow prediction model based on graph convolutional networks to predict the traffic and risk states at multiple future time steps after command execution; S42, verifying whether the risk indicators of the prediction results are within preset safety thresholds. If all indicators meet the thresholds, the verification is passed; otherwise, the command is rejected or corrected; S43, standardizing and converting the verified commands and sending them to the traffic signal control system, information release platform, and emergency command platform. In step S4, the prediction time domain is set to 5 to 30 time steps, with each time step lasting 5 to 30 seconds. This range can fully cover the evolution of traffic flow after command execution, accurately assess potential risks, and ensure the efficiency of rapid prediction.

[0065] Step S5 involves executing feedback and policy optimization, collecting real-world feedback data after instruction execution, constructing experience tuples and storing them in the experience replay buffer, thus completing the online fine-tuning and offline deep training of the decision engine. Based on the invention's objective of achieving continuous policy optimization and iteration, in this embodiment, step S5 specifically includes: S51, collecting new real-world observation data after instruction execution, combining preset targets to generate reward signals and new environmental states, constructing experience tuples and storing them in the experience replay buffer; S52, the online fine-tuning stage, using a soft update strategy to iteratively update the policy network and value network parameters of the decision engine to achieve real-time policy adaptation; and S53, the offline deep training stage, during non-emergency periods, utilizing historical data and simulation data generated by digital twins, employing distributed priority experience replay and importance sampling to complete deep model training and optimize the policy library. In step S5, the update process of the soft update strategy is as follows: the parameters of the target network are obtained by summing the current network parameters after weighting by the update coefficient and the original target network parameters after weighting by subtracting the update coefficient. The update coefficient is set to the range of 0.001 to 0.01. This range can ensure that the target network parameters are updated smoothly, avoid oscillations during training, and achieve rapid adaptation to new experiences.

[0066] The method described in this embodiment can be applied to regional collaborative evacuation scenarios under sudden urban flooding disasters. When multiple road sections in the city center experience waterlogging due to short-term heavy rainfall, and there are conflicts and gaps in multi-source data, step S1 completes situational awareness and uncertainty state construction; step S2 completes digital twin environment synchronization and flood diffusion scenario generation; step S3 generates collaborative evacuation control commands; step S4 completes command security verification and issuance; and step S5 completes execution feedback and strategy optimization, achieving efficient and safe evacuation under flooding scenarios. The method described in this embodiment can also be applied to dynamic rerouting in combined scenarios of large-scale event dispersal and sudden fires. When a sudden fire breaks out in surrounding buildings during the dispersal of a sports event, and it is necessary to balance evacuation efficiency and personnel safety, the aforementioned steps complete fire situational awareness, smoke diffusion scenario generation, collaborative decision generation balancing safety and efficiency, command security verification, and iterative optimization, achieving adaptive evacuation control in combined disaster scenarios.

[0067] In other optional embodiments of this application, the 3D scene modeling module and physical simulation module of the high-fidelity city-level emergency evacuation digital twin layer can also adopt an agent-based modeling and computational fluid dynamics coupling method to jointly simulate the micro-behavior of the crowd and the physical field of the disaster, or use a game engine to construct a highly realistic 3D scene, and connect the physical engine with an external mathematical model to achieve high-dynamic and high-fidelity environmental simulation. The multi-agent deep reinforcement learning decision engine with embedded uncertainty can also adopt a completely decentralized training and execution architecture, and achieve global collaborative optimization of agents through a distributed optimization mechanism based on game theory or consensus algorithms. The action generation unit based on the diffusion model can also adopt alternative solutions such as a policy network based on normalized flow, a method combining proximal policy optimization and autoregressive action decomposition, and evolutionary strategies to process high-dimensional continuous action spaces and generate control commands that meet the requirements. The uncertainty embedding unit can also adopt a framework based on distributed reinforcement learning, allowing the value network to directly output the complete distribution of future cumulative returns, or generate multiple sets of environmental state samples through conditional generative adversarial networks to achieve robust optimization. System management and human-computer interaction interfaces can also employ interpretable artificial intelligence technologies such as attention-weighted visualization, locally interpretable models, and causal graph models of decision trajectories to enhance the transparency of the decision-making process.

[0068] The above descriptions are merely some embodiments of this application and are not intended to limit this application. The technical features or structures in the foregoing different embodiments can be arbitrarily combined to form other specific technical solutions as needed. For those skilled in the art, this application can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of the claims of this application.

Claims

1. A city emergency evacuation decision system based on digital twinning and deep learning, characterized in that, It includes a multi-source heterogeneous data perception and fusion layer that communicates with each other, a high-fidelity city-level emergency evacuation digital twin layer, a multi-agent deep reinforcement learning decision engine with embedded uncertainty, a collaborative control command execution and feedback layer, and a system management and human-computer interaction interface that is electrically connected to each of the aforementioned layers. The multi-source heterogeneous data sensing and fusion layer is used to collect multi-source heterogeneous emergency-related data from the physical world, complete the data credibility assessment and conflict resolution, and output a probabilistic environmental state with uncertainty measurement. The high-fidelity city-level emergency evacuation digital twin layer is used to receive probabilistic environmental states to synchronize the virtual environment with the physical world, construct a high-fidelity emergency evacuation simulation environment, and generate multiple sets of future evolution scenario copies covering uncertainties. The embedded uncertainty multi-agent deep reinforcement learning decision engine is used to receive probabilistic environmental states and scenario copies of the digital twin layer, complete multi-agent collaborative decision-making for risk perception, and output multimodal collaborative control commands. The collaborative control command execution and feedback layer is used to receive multimodal collaborative control commands, and after completing the "sandbox" pre-execution and security verification through the digital twin layer, it sends the control commands to the execution system in the physical world and collects execution feedback data to send back to the multi-source heterogeneous data perception and fusion layer. The system management and human-computer interaction interface is used to realize global situation monitoring, decision intervention, system configuration, and visualization of decision results.

2. The digital-twin and deep learning based urban emergency evacuation decision system according to claim 1, wherein, The multi-source heterogeneous data perception and fusion layer includes a social media semantic understanding and spatiotemporal localization submodule based on a large language model and a dynamic graph structure Bayesian network processing module. The social media semantic understanding and spatiotemporal localization submodule based on a large language model has a built-in spatiotemporal regularization layer with a gated graph neural network, which is used to extract events and locate spatiotemporally from social media texts, and output spatiotemporal event tuples with confidence and localization error. The input of the dynamic graph structure Bayesian network processing module is connected to the physical sensor acquisition device and the output of the social media semantic understanding and spatiotemporal positioning submodule, respectively. It is used to perform credibility weighting, conflict resolution and probabilistic reasoning on multi-source heterogeneous data, and output probabilistic environmental states with probability distribution.

3. The digital-twin and deep learning based urban emergency evacuation decision system according to claim 1, wherein, The high-fidelity city-level emergency evacuation digital twin layer includes a 3D scene modeling module, a physical simulation module, a dynamic environment model learner, and a scenario generator. The 3D scene modeling module uses neural radiation fields to complete high-fidelity 3D reconstruction of key areas; The physical simulation module employs a physical information neural network for simulating the physical constraints of disaster spread and human behavior. The dynamic environment model learner is used to fine-tune the simulation model parameters online based on real-time observation data, so that the digital twin environment can closely approximate the evolution of the real world. The scenario generator, based on a model-independent meta-learning framework, is used to generate training scenarios ranging from easy to difficult, as well as multiple sets of future evolution scenario replicas covering uncertainty.

4. The digital-twin and deep learning based urban emergency evacuation decision system according to claim 1, wherein, The multi-agent deep reinforcement learning decision engine with embedded uncertainty adopts a centralized training and distributed execution paradigm, including a hierarchical multi-agent execution unit, an uncertainty embedding unit, an action generation unit based on a diffusion model, and a multi-agent communication and coordination unit. The hierarchical multi-agent execution unit includes an upper-layer regional collaborative agent and a lower-layer local execution agent. The output of the upper-layer regional collaborative agent is connected to the input of the lower-layer local execution agent to realize hierarchical collaborative decision-making. The uncertainty embedding unit is used to generate a multidimensional uncertainty state vector and inject it into the policy network and value function. The value function has a built-in conditional risk value risk sensitivity measure. The diffusion-based action generation unit employs a time-conditional U-Net denoising architecture to generate high-dimensional continuous control commands that conform to physical constraints. The multi-agent communication and collaboration unit adopts a Transformer communication encoder and has a built-in counterfactual multi-agent policy gradient credit allocation module to realize information interaction and collaborative optimization among agents.

5. A city emergency evacuation decision-making method based on digital twinning and deep learning, applied to the city emergency evacuation decision-making system based on digital twinning and deep learning in any one of claims 1 to 4, characterized in that, Includes the following steps: S1, Emergency Event Triggering and Situation Awareness: After an emergency event is triggered, real-time multi-source heterogeneous data is processed through the multi-source heterogeneous data perception and fusion layer, and a probabilistic environmental state St with uncertainty measurement is output. S2, digital twin environment synchronization and multi-scenario generation, receives probabilistic environmental state St through a high-fidelity city-level emergency evacuation digital twin layer to complete state synchronization and generate N copies of future evolution scenarios covering uncertainties; S3, multimodal collaborative decision generation, is a decision engine that embeds uncertain multi-agent deep reinforcement learning. It completes collaborative decision-making based on probabilistic environment state St and multidimensional uncertain state vector Ut. After parallel inference through multiple scenario replicas, it outputs multimodal collaborative control command At. S4, Command Security Verification and Execution: The multimodal collaborative control command At is pre-executed and security verified in a "sandbox" through the digital twin layer. After successful verification, it is sent to the physical world execution system. S5, closed-loop feedback and strategy optimization, collects real-world feedback data after instruction execution, constructs experience tuples and stores them in the experience replay buffer, and completes online fine-tuning and offline deep training of the decision engine.

6. The city emergency evacuation decision-making method based on digital twin and deep learning according to claim 5, characterized in that, In step S3, the generation and injection of the multidimensional uncertain state vector Ut specifically includes: The scalar uncertainty features in the probabilistic environment state St are aggregated by a graph attention network to generate a hierarchical multidimensional uncertainty state vector Ut. The multidimensional uncertain state vector Ut is concatenated with the probabilistic environment state St and then input into the policy network. By introducing conditional risk value (CVaR) into the value function, the agent's optimization objective is set to maximize CVaRα(R) under a preset risk confidence level, where R is the cumulative reward, thus realizing the deep injection of uncertainty into policy optimization.

7. The urban emergency evacuation decision-making method based on digital twins and deep learning according to claim 5, characterized in that, Step S1 specifically includes: S11, through the social media semantic understanding and spatiotemporal localization submodule based on a large language model, performs named entity recognition, event extraction and spatiotemporal regularization processing on social media text, and outputs spatiotemporal event tuples with confidence and localization error; S12 inputs the spatiotemporal event tuples and multi-source data collected by physical sensors into a dynamic graph structure Bayesian network, and completes the confidence weighting and conflict resolution of multi-source data through the belief propagation algorithm. S13 outputs the probabilistic environment state St with probability distribution for each state variable.

8. The urban emergency evacuation decision-making method based on digital twins and deep learning according to claim 5, characterized in that, The "sandbox" pre-execution and security verification in step S4 specifically include: S41, load the multimodal cooperative control command At into the digital twin environment, and use a traffic flow prediction model based on graph convolutional networks to deduce the traffic and risk status at the next T time steps after the command is executed; S42, verify whether the risk indicators of the simulation results are within the preset safety threshold. If all meet the requirements, the verification is passed; if they exceed the threshold, the verification is rejected or a correction instruction is given. S43 standardizes and converts verified instructions, then distributes them to traffic signal control systems, information dissemination platforms, and emergency command platforms.

9. The urban emergency evacuation decision-making method based on digital twins and deep learning according to claim 5, characterized in that, Step S5 specifically includes: S51: Collect new real-world observation data after the command is executed, combine it with the preset target to generate a reward signal rt and the new environmental state St+1, construct an experience tuple (St,At,rt,St+1) and store it in the experience replay buffer. S52, in the online fine-tuning stage, a soft update strategy is adopted to iteratively update the policy network and value network parameters of the decision engine to achieve real-time policy adaptation. S53, the offline deep training phase, during non-emergency periods, utilizes historical data and simulation data generated by digital twins to complete deep training of the model using distributed priority experience replay and importance sampling, and optimizes the policy library.

10. The urban emergency evacuation decision-making method based on digital twins and deep learning according to claim 5, characterized in that, It also includes steps to enhance decision interpretability, specifically: During the decision-making process, similar historical scenarios and decision-making experiences from the knowledge graph of the retrieval strategy are embedded through TransE graph similarity calculation as decision priors; For key decision points, the Monte Carlo tree search module is invoked to complete online deep planning, supplementing the neural network strategy; Through the counterfactual explanation and comparative analysis module, counterfactual deductions of different decision-making schemes are completed based on structural causal models and do-calculus, and the differences of key indicators are visualized to enhance the transparency of decision-making.

Images