Work activity planning resource allocation simulation method and system based on digital twinning

Abstract

This application provides a simulation method and system for resource allocation in event planning based on digital twins, relating to the field of digital twin technology. The method includes: establishing a three-dimensional digital twin model; deploying a dual-discriminator generative adversarial network (GAN) in the digital twin environment; performing adversarial training and implicit constraint inversion on the dual-discriminator GAN; converting the implicit constraint set into penalty terms; reconstructing the first discriminator; and, based on the three-dimensional digital twin model, performing virtual-real adversarial inference and iteration using the first discriminator, the second discriminator, and an attack generator to output the optimal resource allocation scheme, which is then converted into an executable work order instruction and sent to a physical execution device. This solves the technical problem that existing digital twin-based resource allocation optimization for event planning directly relies on preset rules and expert experience, leading to insufficient optimization of the allocation scheme. Through iterative virtual-real adversarial interaction in digital twins, the method achieves the technical effects of feasibility, optimization, and security of the resource allocation scheme.

Description

Technical Field

[0001] This application relates to the field of digital twin technology, specifically to a simulation method and system for resource allocation in event planning based on digital twins. Background Technology

[0002] In existing project planning, resource allocation involves multiple aspects. To improve planning efficiency and feasibility, digital twin technology is widely used. By establishing a three-dimensional virtual environment that maps to the actual project, resource allocation schemes can be simulated and verified. However, existing digital twin methods often rely on fixed rules or expert experience for scheme generation and evaluation. Directly applying preset rules or single scoring indicators to the virtual environment for simulation can easily overlook complex constraints and potential coupling effects between resources, leading to insufficient scheme optimization. Some potential conflicts or high-risk configurations cannot be detected in time, especially in multi-objective and multi-constraint scenarios. Furthermore, since conflicts may exist between different optimization objectives and expert preferences, a single evaluation model cannot simultaneously take into account multiple-dimensional constraints, resulting in deviations or risks in the generated resource allocation schemes during actual implementation, failing to fully meet the comprehensive optimization requirements of the activity.

[0003] In summary, existing digital twins rely directly on preset rules and expert experience to optimize resource allocation for event planning, resulting in insufficient optimization of the allocation scheme. Summary of the Invention

[0004] The purpose of this application is to provide a simulation method and system for resource allocation in event planning based on digital twins, in order to solve the technical problem that the optimization of resource allocation in event planning based on digital twins directly relies on preset rules and expert experience, resulting in insufficient optimization of the allocation scheme.

[0005] To achieve the above objectives, this application provides a simulation method and system for resource allocation in event planning based on digital twins.

[0006] Firstly, this application provides a simulation method for resource allocation in event planning based on digital twins. The method includes: establishing a three-dimensional digital twin model that maps one-to-one with the actual event venue, exhibits, personnel, and equipment; deploying a dual-discriminator generative adversarial network (GAN) in the digital twin environment, including a first discriminator, a second discriminator, and an attack generator; performing adversarial training and implicit constraint inversion on the dual-discriminator GAN to generate a set of implicit constraints; converting the set of implicit constraints into penalty terms, reconstructing the first discriminator to obtain an enhanced first discriminator; and, based on the three-dimensional digital twin model, performing virtual-real adversarial inference and iteration with the first discriminator, the second discriminator, and the attack generator to output the optimal resource allocation scheme and convert it into an executable work order instruction sent to the physical space execution device.

[0007] Optionally, the first discriminator is a pre-trained optimizer used to output a first evaluation value for the input resource allocation scheme; the second discriminator is a scoring model trained based on historical expert decision data used to output a second evaluation value for the same resource allocation scheme corresponding to the first discriminator; the attack generator is a conditional generative adversarial network, the conditional input is a basic parameter vector of the activity, including but not limited to the number of exhibits, venue area, total budget, and personnel skill list, and the output is a resource allocation tensor.

[0008] Optionally, training samples are collected, and the parameters of the first discriminator, the second discriminator, and the attack generator are trained alternately to obtain the converged dual-discriminator generative adversarial network; all resource allocation schemes that satisfy the evaluation difference greater than a preset evaluation threshold and their corresponding first and second evaluation values ​​are saved as a divergence sample set during the training process; with the deviation between the second evaluation value and the first evaluation value as the objective, each scheme in the divergence sample set is classified into decision categories to construct the implicit constraint set.

[0009] Optionally, for each scheme in the divergent sample set, a decision tree classification is run, using the feature vector of each scheme as input and the comparison result of the scores of the first discriminator and the second discriminator as the classification label, to generate each constraint rule corresponding to each scheme. Each rule contains a specific combination of feature thresholds. The confidence of each constraint rule on the validation set is calculated, and rules with confidence scores greater than a preset confidence threshold are added to the implicit constraint set.

[0010] Optionally, step one: fix the first discriminator and the second discriminator to construct an attack loss function by maximizing the evaluation difference between the first evaluation value and the second evaluation value, and train the attack generator; step two: fix the attack generator and update the first discriminator and the second discriminator three times consecutively; alternately execute steps one and two until the attack loss function converges.

[0011] Optionally, the attack loss function is set as the negative of the difference between the first evaluation value and the second evaluation value plus a gradient penalty term.

[0012] Optionally, for numerical features appearing in each constraint rule of the implicit constraint set, the variable index in the resource allocation vector is found, and the feature threshold is converted into upper and lower bound constraints of the variable; for categorical features appearing in each constraint rule of the implicit constraint set, they are converted into linear equations or inequalities of binary flag variables; the upper and lower bound constraints and the linear equations or inequalities are used as penalty terms to constrain and train the first discriminator to obtain the enhanced first discriminator.

[0013] Optionally, basic information about the submitted works is collected and input into the attack generator. The score of the enhanced first discriminator is used as the optimization target to search for a first optimization scheme that maximizes the score of the enhanced first discriminator. The enhanced first discriminator and the second discriminator are fixed, and the attack generator is trained to optimize the first optimization scheme to generate a second optimization scheme. This second optimization scheme is generated such that the evaluation value of the enhanced first discriminator for the second optimization scheme is lower than its evaluation value for the first optimization scheme, while the evaluation value of the second discriminator for the second optimization scheme is higher than its evaluation value for the first optimization scheme. The first optimization scheme and the second optimization scheme are simulated in the three-dimensional digital twin model respectively, and the actual performance scores are recorded. If the difference between the actual performance scores is less than the convergence threshold, the iteration is terminated, and the scheme with the highest actual performance score at the time of iteration termination is output as the optimal resource allocation scheme.

[0014] Optionally, if the difference in actual performance scores is greater than or equal to the convergence threshold, the scheme with the highest actual performance score among the schemes generated in this round is taken as the new starting point for iteration, and the attack generator is trained iteratively.

[0015] Secondly, this application also provides a simulation system for resource allocation in event planning based on digital twins. The system includes: a twin model building module for building a three-dimensional digital twin model that maps one-to-one with the actual event venue, exhibits, personnel, and equipment; an adversarial network deployment module for deploying a dual-discriminator generative adversarial network in the digital twin environment, including a first discriminator, a second discriminator, and an attack generator; a latent constraint generation module for performing adversarial training and latent constraint inversion on the dual-discriminator generative adversarial network to generate a set of latent constraints; and a configuration scheme output module for converting the set of latent constraints into penalty terms, reconstructing the first discriminator to obtain an enhanced first discriminator, and performing virtual-real adversarial inference and iteration based on the three-dimensional digital twin model, using the first discriminator, the second discriminator, and the attack generator to output the optimal resource allocation scheme and convert it into an executable work order instruction to be sent to the physical space execution device.

[0016] One or more technical solutions provided in this application have at least the following technical effects or advantages: By establishing a 3D digital twin model that maps one-to-one with the actual event venue, exhibits, personnel, and equipment; deploying a dual-discriminator generative adversarial network (GAN) within the digital twin environment, including a first discriminator, a second discriminator, and an attack generator; performing adversarial training and implicit constraint inversion on the dual-discriminator GAN to generate a set of implicit constraints; converting the set of implicit constraints into penalty terms to reconstruct the first discriminator, resulting in an enhanced first discriminator; and performing virtual-real adversarial inference and iteration based on the 3D digital twin model using the first discriminator, the second discriminator, and the attack generator to output an optimal resource allocation scheme, which is then converted into executable work order instructions and sent to the physical execution device. Ultimately, this achieves the technical effect of realizing the feasibility, optimization, and security of resource allocation schemes through iterative virtual-real adversarial interaction using a digital twin.

[0017] The above description is merely an overview of the technical solution of this application. To enable a clearer understanding of the technical means of this application and to facilitate its implementation according to the description, and to make the above and other objects, features, and advantages of this application more apparent, specific embodiments of this application are described below. It should be understood that the content described in this section is not intended to identify key or important features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent through the following description. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the simulation method for resource allocation in event planning based on digital twins, as described in this application.

[0020] Figure 2 This is a schematic diagram of the structure of the digital twin-based simulation system for planning resources and allocating creative activities, as described in this application.

[0021] Figure labeling: Twin model establishment module 11, adversarial network deployment module 12, implicit constraint generation module 13, configuration scheme output module 14. Detailed Implementation

[0022] This application provides a simulation method and system for resource allocation in event planning based on digital twins. It addresses the technical problem that existing digital twin-based resource allocation optimization relies directly on preset rules and expert experience, leading to insufficient optimization of allocation schemes. It achieves the technical effect of improving the feasibility, optimization, and security of resource allocation schemes through iterative virtual-real adversarial interaction between digital twins.

[0023] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. It should be understood that this application is not limited to the exemplary embodiments described herein. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application. It should also be noted that, for ease of description, only the parts related to this application are shown in the accompanying drawings, not all of them.

[0024] Example 1, please refer to the appendix. Figure 1 This application provides a simulation method for resource allocation in event planning based on digital twins, wherein the simulation method specifically includes: Establish a three-dimensional digital twin model that maps one-to-one with the actual event venue, exhibits, personnel, and equipment.

[0025] Specifically, a 3D digital twin model refers to a model that, in a virtual computer environment, digitally reproduces the spatial structure, state attributes, and interactive relationships of a real-world event venue and its related elements. First, a high-precision 3D geometric reconstruction of the actual event venue is performed. A lidar point cloud scanning device is used to scan the venue from all angles to ensure that the geometric details of structural elements such as walls, floors, ceilings, doors, windows, and load-bearing columns are fully captured. Simultaneously, based on the venue's architectural design drawings and on-site measurement data, semantic annotation of the internal spatial topology is performed, i.e., spatial zoning annotation—dividing the venue into functional areas such as exhibition areas, audience passageways, backstage operation areas, and equipment storage areas. Each area is assigned a unique spatial ID, such as ZoneA01 and ZoneB03, and attribute parameters such as area, floor height, and maximum capacity are recorded for each area.

[0026] Then, each exhibit is digitized through 3D scanning or by importing BIM files. 3D scanning uses optical or laser scanning to acquire the exhibit's geometric dimensions, shape, surface features, and structural details, ensuring consistency between the virtual model and the physical object. The BIM model provides the exhibit's three-dimensional geometry and attribute information, including weight, material properties, mobility, and fixed status, for subsequent resource scheduling and constraint simulation. After scanning or importing, the exhibit models are unified into the three-dimensional coordinate system of the event venue, and dynamic attributes, such as mobility, need for fixation, or specific operational requirements, are labeled to facilitate behavioral simulation and spatial layout optimization within the digital twin environment.

[0027] Based on job data, skill tags, and personnel statistics, virtual personnel entities are constructed for digital twin simulation. By utilizing historical shift records, motion trajectories collected by wearable sensors, and task execution data, each virtual personnel's movement path constraints, operable equipment permissions, and workload attributes are obtained. Movement path constraints include accessible areas and access restrictions; operable equipment permissions include lighting and security equipment; and workload attributes include the number of tasks and time allocation. Personnel information is then bound to a 3D site model, ensuring that each entity moves, operates equipment, and completes tasks according to physical constraints and job requirements during simulation. This provides a precise and dynamic behavioral basis for resource allocation optimization.

[0028] The various equipment required for the event, such as projectors, lighting, sound systems, and transport vehicles, are virtually modeled by collecting their specifications, power, operating range, capacity, and list of executable tasks. Each piece of equipment is abstracted in the model as a controllable virtual object, including its location coordinates, operational status, load capacity, and task constraints. Data collection methods include manufacturer technical manuals, equipment sensor data, historical usage records, and on-site measurements. By binding various equipment attributes to the digital twin environment, the state changes of the equipment during the event can be simulated, such as switching operations, position adjustments, or task execution, providing controllable objects for resource scheduling optimization.

[0029] Data on the actual event venue, exhibits, personnel, and equipment are uniformly imported into a digital twin platform, such as Unity, and spatially mapped under a unified coordinate system to ensure precise correspondence between the location, size, and interactive relationships of the venue, exhibits, personnel, and equipment. This one-to-one mapping means that every entity in the physical space—such as every wall, exhibit, person, and device—has one and only one corresponding virtual object in the digital twin model, and the spatial coordinates, geometric dimensions, and attribute parameters of this virtual object are numerically consistent with the physical entity. The digital twin platform maintains real-time data synchronization with the IoT gateway in the physical space via OPC UA or RESTful API at a frequency of 1Hz, updating dynamic data such as device status and personnel location every second. This ensures that the virtual model always reflects the current state of the physical world, providing a realistic and reliable virtual experimental environment for subsequent simulation and deduction.

[0030] By establishing a three-dimensional digital twin model, we can accurately map real-world scenarios and resource information, providing a reliable virtual testing environment for subsequent dual-discriminator adversarial training and resource allocation optimization. This not only verifies the feasibility of the solution in a virtual environment, but also identifies potential conflicts and optimization space in advance by simulating various configurations and constraints, providing a reliable basis for resource allocation.

[0031] Deploy a dual-discriminator generative adversarial network in a digital twin environment, comprising a first discriminator, a second discriminator, and an attack generator.

[0032] Furthermore, the first discriminator is a pre-trained optimizer used to output a first evaluation value for the input resource allocation scheme; the second discriminator is a scoring model trained based on historical expert decision data used to output a second evaluation value for the same resource allocation scheme corresponding to the first discriminator; the attack generator is a conditional generative adversarial network, the conditional input is the basic parameter vector of the activity, including but not limited to the number of exhibits, venue area, total budget, and personnel skill list, and the output is a resource allocation tensor.

[0033] Specifically, a dual-discriminator generative adversarial network (GAN) is deployed in a digital twin environment to associate virtual activity scenarios with an intelligent optimization model, enabling simulation, scoring, and iterative optimization of resource allocation schemes. First, a 3D digital twin model is used as the operating platform, containing the geometric and attribute information of all resources, including the venue, exhibits, personnel, and equipment. Then, a dual-discriminator GAN module is loaded into this digital twin environment, allowing the first discriminator D1, the second discriminator D2, and the attack generator G to directly access the resource allocation vectors and dynamic state information in the digital twin. Resource allocation schemes generated in the digital twin environment are input into the first and second discriminators D1 and D2 for scoring via APIs or data interfaces. Simultaneously, the attack generator G generates new schemes based on conditional parameters, achieving real-time interaction between the virtual environment and the intelligent optimization network. During deployment, data consistency and coordinate uniformity among modules must be ensured. A unified scene manager or data bus can be used to transmit the spatial location, state attributes, and task information of virtual entities to the network input, while the optimization results are fed back to the digital twin for simulation verification.

[0034] The dual-discriminator generative adversarial network includes a first discriminator D1, a second discriminator D2, and an attack generator G. The first discriminator D1 is a pre-trained optimizer, which can be constructed using a multilayer perceptron network. It outputs a first evaluation value for the input resource allocation scheme. The first discriminator D1 includes: an input layer that receives a resource allocation scheme vector, the vector dimension of which is determined by all resource features in the activity, such as exhibit location coordinates, personnel task allocation matrix, equipment location and status; the total dimension can be automatically calculated based on the number of resources. Following the input layer are three hidden layers, containing 128, 64, and 32 nodes respectively. The layers are fully connected, and each hidden layer is followed by a ReLU activation function to increase non-linear expressive power. The output layer generates a first evaluation value, ranging from 0 to 1, after normalization using a Sigmoid activation function.

[0035] The first discriminator D1 employs a mean squared error loss function, using performance indicators simulated in the digital twin environment as the true labels. These indicators include space utilization, personnel load balancing, equipment utilization efficiency, and safety constraint satisfaction. In the digital twin environment, space utilization is calculated by determining the proportion of the area occupied by exhibits to the total available exhibition area. This can be further combined with aisle or safety distance constraints to calculate the effective utilization rate. Personnel load balancing is achieved by statistically analyzing the number of tasks undertaken and the duration of each virtual worker during the event, calculating the standard deviation or variance. A smaller value indicates a more balanced load. Equipment utilization efficiency is determined by analyzing the proportion of time each device is actually scheduled and operated during the event to the total available time, combined with the task completion rate for a comprehensive score. Safety constraint satisfaction is achieved by detecting whether the minimum distance between all exhibits and personnel movement paths, the safe distance for equipment operation, and the site's carrying capacity in the digital twin model meet preset thresholds. If the conditions are met, it is recorded as 1; otherwise, it is recorded as 0. A weighted average of all constraints is then calculated to obtain the overall safety constraint satisfaction. The weight values ​​can be set according to actual needs and expert experience. Finally, these indicators are normalized to between 0 and 1 and used as training labels for the first discriminator D1, enabling the multilayer sensing network to quantitatively evaluate the overall advantages and disadvantages of resource allocation schemes in terms of spatial layout, personnel allocation, equipment utilization, and security.

[0036] During the training of the multilayer perceptron, the execution results of multiple resource allocation schemes in the digital twin are first collected, and the corresponding performance indicators are calculated as labels. The resource allocation vector is then input into D1 for forward propagation to calculate the predicted score. The mean squared error between the predicted score and the actual performance indicator is calculated, and the Adam gradient descent optimization algorithm is used to update the network weights in reverse. This process is repeated until training converges. The convergence condition is that the mean squared error of the validation set is lower than a preset threshold of 0.01 or the error change is less than 0.001 for several consecutive rounds. After training is completed, D1 is obtained and used to evaluate the newly generated resource allocation schemes. The first evaluation value reflecting the merits of the resource allocation schemes is output and used for subsequent generator optimization and adversarial training.

[0037] The second discriminator D2 employs a scoring model trained on historical expert decision data. It is used to output a second evaluation value for the same resource allocation scheme corresponding to the first discriminator. For example, a lightweight Transformer encoder architecture is selected, which includes two multi-head self-attention layers with eight heads and a hidden dimension of 256 as the backbone. Its training data comes from the scoring records of the expert review panel on the resource allocation scheme in no less than 500 similar exhibitions in the past. The input of each record is a one-dimensional feature vector after the resource allocation tensor is flattened, and the output is the discrete score (0-1 score) given by the expert. In the data preprocessing stage, the flattened feature vector is subjected to Layer Normalization as training labels.

[0038] The training process uses the mean squared error loss function, employs the Adam optimizer, and sets the learning rate to 1×10⁻⁶. -4 Training is conducted for approximately 200 epochs, with the training and validation sets split in an 8:2 ratio. Training stops and the model is saved when the root mean square error on the validation set falls below 0.01, resulting in a second discriminator, D2. The second discriminator D2 analyzes the same resource allocation scheme corresponding to the first discriminator and outputs a second evaluation value. This simulates the decision-making experience of human experts, providing a satisfaction score reflecting the business rationality of the input resource allocation scheme. A higher score indicates that the scheme is closer to the configuration level recognized by experts.

[0039] The attack generator G employs a conditional generative adversarial network (cGAN) architecture, aiming to generate executable resource allocation schemes given basic activity parameters. During construction, the input consists of a conditional vector and a random noise vector. The conditional vector includes, for example, the number of exhibits, venue area, total budget, and personnel skill list. The random noise vector typically has a dimension of 50-100, sampled from a standard normal distribution to increase the diversity of resource allocation schemes. The conditional vector is first mapped to a 64-dimensional high-dimensional vector through an embedding layer to ensure the conditional information is fully expressed in the network. It is then concatenated with the noise vector as the input to the hidden layer. The generator contains three fully connected hidden layers with 256, 128, and 64 nodes respectively. Each layer is followed by a LeakyReLU activation function to enhance non-linear expressiveness. The output layer has the same number of nodes as the resource allocation vector dimension, including exhibit locations, personnel task allocation matrices, and equipment state coordinates. Each output element is normalized to between 0 and 1 using a Sigmoid activation function for easy mapping to actual numerical values ​​later.

[0040] By deploying a dual-discriminator generative adversarial network, intelligent generation, evaluation, and optimization of resource allocation schemes in a virtual environment are achieved. D1 provides data-driven optimization directions, D2 provides expert experience constraints, and G explores between the two, iteratively generating feasible and high-quality resource allocation schemes. This provides foundational data for subsequent implicit constraint extraction and digital twin simulation, realizing a closed loop of virtual and real-world optimization.

[0041] Adversarial training and implicit constraint inversion are performed on the dual discriminator generative adversarial network to generate a set of implicit constraints.

[0042] Furthermore, the dual-discriminator generative adversarial network is subjected to adversarial training and implicit constraint inversion to generate an implicit constraint set, including: collecting training samples, alternately adversarially training the parameters of the first discriminator, the second discriminator, and the attack generator to obtain the converged dual-discriminator generative adversarial network; saving all resource allocation schemes that satisfy the evaluation difference greater than a preset evaluation threshold and their corresponding first and second evaluation values ​​as a divergence sample set during the training process; and classifying each scheme in the divergence sample set based on the deviation between the second and first evaluation values ​​to construct the implicit constraint set.

[0043] Furthermore, training samples are collected, and the parameters of the first discriminator, the second discriminator, and the attack generator are trained alternately to obtain the converged dual-discriminator generative adversarial network, including: Step 1: Fix the first discriminator and the second discriminator, construct an attack loss function to maximize the evaluation difference between the first evaluation value and the second evaluation value, and train the attack generator; Step 2: Fix the attack generator, and update the first discriminator and the second discriminator three times consecutively; Step 1 to Step 2 are executed alternately until the attack loss function converges.

[0044] Furthermore, the attack loss function is set as the negative of the difference between the first evaluation value and the second evaluation value plus a gradient penalty term.

[0045] Specifically, training samples are first collected from the digital twin environment. These samples include multiple resource configuration schemes and their performance metrics after simulated execution in the digital twin, such as space utilization, personnel load balancing, equipment efficiency, and security constraint satisfaction. Each resource configuration scheme outputs a first evaluation value and a second evaluation value through a first discriminator D1 and a second discriminator D2, serving as feedback signals for adversarial learning during training. The training employs an alternating adversarial training strategy: the first discriminator D1 and the second discriminator D2 are fixed, and all network layer parameters of D1 and D2 are set to an untrainable state, i.e., requirements are set. grad =False, prevents gradient backpropagation to update weights, and uses the parameters of the attack generator G as the optimization object, updating the parameters of G with the attack loss function as the target.

[0046] The attack loss function is constructed by maximizing the negative value of the difference between the first and second evaluation values ​​and adding a gradient penalty term. This represents the attack generator G creating a difference in scores between the first discriminator D1 and the second discriminator D2 when generating resource allocation schemes, thereby exposing potential constraints or optimizable boundaries. The gradient penalty term is used to constrain the gradient difference of the scores in order to ensure gradient stability and avoid the discriminator gradient explosion caused by the generated resource allocation scheme under extreme conditions. This can be achieved by calculating the gradient of the output score of the generated scheme in the first discriminator D1 with respect to the input resource allocation vector, then calculating the squared deviation of its L2 norm from 1, and multiplying it by the penalty coefficient λ. For example, taking 10 and adding it to the attack loss function ensures that the generator update does not cause the score to fluctuate drastically in the gradient space, while constraining the generated scheme to be fine-tuned in a physically feasible continuous space, thereby improving training stability.

[0047] Gradient descent, such as the Adam optimizer, iteratively updates the parameters of the attack generator G, enabling it to generate resource allocation schemes that meet the conditional constraints but may violate implicit constraints. In other words, through the game between the generator and the bi-discriminator, it uncovers divergent schemes that are physically feasible but considered unreasonable by experts, or that are considered reasonable by experts but have potential physical vulnerabilities.

[0048] Then, proceed to step two: fix the attack generator G, and continuously update the parameters of the first discriminator D1 and the second discriminator D2 three times. Updating three times means that in each round of step two, the parameters of D1 and D2 are updated three times each, rather than just once. This allows the discriminator to adjust its decision boundary more fully when facing adversarial examples from the attack generator, preventing the attack generator from exploiting the weakness of insufficient discriminator updates to deceive the discriminator. This enables the discriminator to more accurately evaluate the quality and potential constraints of the generated scheme.

[0049] Steps one and two are executed repeatedly until the attack loss function converges, that is, the score difference between the resource configuration scheme generated by the fixed attack generator G and the first discriminator D1 and the second discriminator D2 is stable, such as being less than 0.001 for 10 consecutive times and not being able to increase further, thus obtaining the dual discriminator generative adversarial network that has been trained and converged.

[0050] During training, each resource allocation scheme generated in each round is scored using D1 and D2 methods, and the absolute value of the score difference is calculated. When the score difference of a resource allocation scheme exceeds a preset evaluation threshold, it indicates that the scheme has potential discrepancies, meaning it is physically feasible but receives a low expert score, or receives a high expert score but has potential physical vulnerabilities. In this case, the resource allocation scheme, such as exhibit location, personnel task allocation, equipment status, and corresponding first and second evaluation values, is saved to a discrepancy sample set. The saving method can be a structured data table or a database, for example, storing the feature vector of one scheme and two score values ​​in each row.

[0051] The preset evaluation threshold is used to determine whether a resource allocation scheme belongs to the divergent sample. It can be based on the score difference statistics in the early stage of training. For example, the mean μ and standard deviation σ of the score difference of all generated schemes in the training set can be calculated, and the threshold can be set as μ + k × σ, where k can be 1 to 2, to ensure that only samples with obvious divergence are selected. For example, if the mean of the score difference between D1 and D2 of 1000 schemes generated in one training session is 0.08 and the standard deviation is 0.04, then the threshold can be set as 0.08 + 1 × 0.04 = 0.12. At this time, all resource allocation schemes with a score difference greater than 0.12 are added to the divergent sample set. For example, if resource allocation scheme A has a D1 score of 0.85 and a D2 score of 0.70, and the difference 0.15 > 0.12, then resource allocation scheme A and its D1 and D2 scores are saved in the divergent sample set for subsequent implicit constraint analysis to deduce the constraint missing in D1. The discrepancy sample set is analyzed, and the feature vectors of each solution, such as exhibit location, personnel task allocation, equipment status, etc., and the deviation from the score difference are used as decision targets. Decision trees are used to classify the features and extract rules that can distinguish between high and low scores, forming a set of implicit constraints.

[0052] Adversarial training enables the attack generator to actively explore potential constraints and boundary conditions, and the analysis of divergent samples extracts implicit constraints not explicitly stated by experts, providing constraint guidance for the enhanced training of the first discriminator, realizing a closed loop of virtual and real-world optimization, and improving the security, feasibility, and optimization potential of resource allocation schemes.

[0053] Furthermore, taking the deviation between the second evaluation value and the first evaluation value as the target, decision classification is performed on each scheme in the divergence sample set to construct the implicit constraint set. This includes: for each scheme in the divergence sample set, running decision tree classification, using the feature vector of each scheme as input, and using the comparison result of the scores of the first discriminator and the second discriminator as the classification label, generating each constraint rule corresponding to each scheme, with each rule containing a specific combination of feature thresholds; calculating the confidence of each constraint rule on the validation set, and adding rules with confidence scores greater than a preset confidence threshold to the implicit constraint set.

[0054] Specifically, after training is completed and a divergence sample set is generated, the conditions under which the resource allocation features lead to inconsistent scores between the two discriminators are analyzed based on the deviation between the second and first evaluation values. Each sample in the divergence sample set contains a complete resource allocation scheme and its D1 and D2 scores. The feature vector consists of exhibit location coordinates, personnel task allocation matrix, equipment status, etc., where the personnel task allocation matrix includes the number of tasks and task type for each person. The scoring deviation is defined as Δ = D1 - D2. For decision classification, the classification label of each sample is set to binary: if the D1 score is higher than the D2 score, the classification label is 1, indicating that there is a discriminator scoring deviation and the D1 evaluation is better; if the D2 score is higher than the D1 score, the classification label is 0.

[0055] A decision tree is constructed using the CART algorithm. CART is a binary tree classification method that minimizes the impurity within a node by recursively splitting the feature space. The decision tree structure is designed as follows: the input layer receives sample feature vectors, with each feature corresponding to an input node, such as the X-coordinate of exhibit A, the Y-coordinate of exhibit A, the number of tasks performed by personnel 1, and the usage status of equipment 2. Each split node selects a feature and its threshold to divide the node samples into left and right subtrees, maximizing the label purity of the left and right subtrees. Label purity refers to the proportion of samples of the same class within a given node. A commonly used quantification metric is the Gini index. Maximizing purity means that the selected feature and threshold minimize the weighted average of the Gini indices of the left and right subtrees, i.e., each subtree after splitting should contain as many samples as possible of a single class. For example, traverse all features of the current node, such as the X coordinate of exhibits, the number of tasks for personnel 1, and the status of equipment 2. For numerical features, such as the X coordinate of exhibits, take the unique value or equally spaced candidate values ​​of this feature for all samples within the node, such as 10 meters, 11 meters, 12 meters...15 meters as thresholds. For each candidate threshold, place samples in the left subtree if the feature value is less than or equal to the threshold, and in the right subtree if the feature value is greater than the threshold. Then calculate the Gini index and weighted average of the left and right subtrees. Select the feature and threshold that minimizes the weighted Gini index as the final splitting scheme. For example, if the X coordinate sample value of exhibit A in the node ranges from 8 to 20 meters, and the traversal calculation shows that the threshold of 12 meters results in 90% of the label 1 in the left subtree and 85% of the label 0 in the right subtree, with the lowest weighted average Gini index, then the node selects exhibit A's X coordinate being less than or equal to 12 meters as the splitting condition. For categorical features, such as equipment status (on or off), directly divide the samples of different categories into the left and right subtrees and calculate the Gini index to select the optimal split. The leaf node outputs a classification result of 1 or 0 and records the corresponding feature interval of the sample, thus generating constraint rules.

[0056] During training, the minimum number of samples per node is set to 5, and the maximum depth of the tree is set to 5 layers. Splitting stops when either condition is met to avoid overfitting. Each generated leaf node path is a constraint rule, and each constraint rule contains a specific combination of feature thresholds. For example, if the X coordinate of exhibit A is within 12 meters and the number of personnel task 1 is greater than 3, the D1 score is higher than the D2 score, indicating that the scores of the two discriminators are inconsistent under this feature combination.

[0057] To verify the reliability of the rules, each constraint rule is applied to an independent validation set, and the confidence score is calculated. This confidence score is calculated by dividing the number of samples in the validation set that meet the rule's conditions by the total number of samples in the validation set that meet the condition across all feature intervals. For example, if a rule has 50 samples in the validation set that satisfy the condition that exhibit A's X-coordinate is 12 meters and personnel task 1 is greater than 3, and 42 of these samples have a D1 score that is indeed higher than their D2 score, then the confidence score for this rule is 42 / 50 = 0.84, which exceeds the preset confidence threshold of 0.8. Therefore, this rule is added to the final implicit constraint set.

[0058] By extracting implicit constraints not explicitly specified by experts from divergent samples, a quantifiable and executable set of rules is formed. The first discriminator is then reconstructed, enabling the adversarial generator to generate resource allocation schemes that not only meet the constraints but also avoid physically infeasible or potentially risky configurations, thereby further improving the security, feasibility, and optimization efficiency of resource allocation schemes.

[0059] The implicit constraint set is converted into a penalty term, and the first discriminator is reconstructed to obtain an enhanced first discriminator. Based on the three-dimensional digital twin model, the first discriminator, the second discriminator, and the attack generator are used to perform virtual-real adversarial simulation and iteration, output the optimal resource allocation scheme, and convert it into an executable work order instruction to be sent to the physical space execution device.

[0060] Furthermore, the implicit constraint set is converted into penalty terms, and the first discriminator is reconstructed to obtain an enhanced first discriminator. This includes: for numerical features appearing in each constraint rule of the implicit constraint set, finding the variable index in the resource allocation vector, and converting the feature threshold into upper and lower bound constraints of the variable; for categorical features appearing in each constraint rule of the implicit constraint set, converting them into linear equations or inequalities of binary flag variables; and using the upper and lower bound constraints and the linear equations or inequalities as penalty terms to constrain and train the first discriminator to obtain the enhanced first discriminator.

[0061] Specifically, for numerical features involved in the implicit constraint set, such as budget allocations for each time period, number of personnel, and total equipment power, the corresponding variable index is found in the resource allocation vector, transforming the feature thresholds in the constraint rules into upper and lower bounds for the variables. The resource allocation vector is a vector that digitally represents all configurable resources in an activity, with each element corresponding to a specific resource attribute or state. For example, in an exhibition activity, the vector might include the three-dimensional coordinates of exhibits, the number of personnel in each exhibition area, the personnel task allocation matrix, equipment status, and equipment power or capacity. The vector length is determined by the total number of all resource features and maintains a fixed order so that specific variables can be indexed and located during network training or optimization. The variable index refers to the position of each element in the resource allocation vector; for example, the 7th element corresponds to the number of personnel in exhibition area A, and the 12th element corresponds to the total equipment power in exhibition area B. When it is necessary to apply implicit constraint rules to specific variables, the rules can be transformed into upper and lower bounds or linear constraints for the corresponding elements in the vector by looking up the index. For example, if an implicit constraint rule is: the number of people in exhibition area A shall not exceed 15, then look up the index of the number of people in exhibition area A in the resource allocation vector, set its upper bound to 15, and keep the lower bound at its original setting or adjust it according to the actual scenario.

[0062] For each categorical feature appearing in the implicit constraint set, such as exhibition area type labels and personnel skill tags, the constraint rules are transformed into linear equations or inequalities. For example, for the rule: if the exhibition area number is A and the manpower density is greater than 0.6, the scores are inconsistent, this is transformed into the constraint: the manpower density variable minus 0.6 multiplied by the exhibition area A label variable is less than or equal to 0. This constraint is stored in the Aeq or Aineq matrix of the mixed-integer linear programming model in the form of a coefficient matrix row, ensuring that the constraint is satisfied during the solution process.

[0063] Repeat the above process for all implicit constraint rules to obtain the complete upper and lower bound constraint matrix and the linear equality or inequality constraint matrix. Introduce the upper and lower bound constraints and the linear equality or inequality as penalty terms into the loss function of the first discriminator D1, that is, add a penalty term to the original mean squared error loss, L=L MSE +λΣP violation Where λ is the penalty weight coefficient, which can be set to 0.2~0.5 and optimized using the validation set. ΣP violationTo represent the total penalty term, i.e. the degree of violation of implicit constraints, in the current D1 prediction resource allocation scheme output, if the numerical features exceed the upper and lower bound constraints, the degree of violation is calculated by subtracting the feature from the nearest boundary value and taking the absolute value. When the categorical features in the output resource allocation scheme do not satisfy their corresponding linear equality or inequality constraints, the degree of violation can be quantified by calculating the difference between the left-hand expression of the constraint and the right-hand threshold. The degree of violation is calculated, and all degree of violation is summed as the total penalty term.

[0064] During training, when the resource allocation scheme output by D1 violates any implicit constraint, the penalty term increases the loss value, thereby guiding the optimizer to reduce the behavior of violating implicit constraints during parameter updates. Through constraint training, an enhanced first discriminator is obtained, which ensures that when generating the first evaluation value, the enhanced first discriminator D1 takes into account the implicit constraint rules, so that its output score reflects both the merits and demerits of the resource allocation schemes and avoids potentially infeasible or high-risk schemes.

[0065] By transforming implicit constraint rules into numerical upper and lower bound constraints, categorical linear equality or inequality constraints, and introducing them as penalty terms into the loss function of the first discriminator for constraint training, the enhanced first discriminator can simultaneously consider explicit performance indicators and potential implicit constraints when scoring resource allocation schemes, thereby improving the overall feasibility and security of resource allocation schemes and ensuring that the generated resource allocation schemes are both optimized and reliable.

[0066] Furthermore, based on the three-dimensional digital twin model, the first discriminator, the second discriminator, and the attack generator perform virtual-real adversarial simulation iterations to output the optimal resource allocation scheme. This includes: collecting basic information about the activity works, inputting it into the attack generator, using the score of the enhanced first discriminator as the optimization target, and searching for a first optimization scheme that maximizes the score of the enhanced first discriminator; fixing the enhanced first discriminator and the second discriminator, training the attack generator to optimize the first optimization scheme, generating a second optimization scheme such that the evaluation value of the enhanced first discriminator for the second optimization scheme is lower than its evaluation value for the first optimization scheme, while the evaluation value of the second discriminator for the second optimization scheme is higher than its evaluation value for the first optimization scheme; simulating the first optimization scheme and the second optimization scheme respectively in the three-dimensional digital twin model and recording the actual performance scores; if the difference in actual performance scores is less than a convergence threshold, the iteration is terminated, and the scheme with the highest actual performance score at the time of iteration termination is output as the optimal resource allocation scheme.

[0067] Furthermore, if the difference in actual performance scores is greater than or equal to the convergence threshold, the scheme with the highest actual performance score among the schemes generated in this round is taken as the new starting point for iteration, and the attack generator is trained iteratively.

[0068] Specifically, basic information about the artworks and activities is acquired through 3D scanning, BIM modeling, or on-site measurement. This includes the geometric dimensions, weight, material properties, quantity of exhibits, and spatial distribution information of the exhibition area. This information is encoded using a unified resource allocation vector format, with each element corresponding to a specific resource attribute or status, such as the number of exhibits, site area, total budget, and personnel skill list. This basic information is then input into the attack generator G to generate an initial resource allocation scheme. An enhanced first discriminator (EDB) scores the generated initial resource allocation scheme, using its score as the optimization objective—that is, searching for the first optimal scheme that maximizes the EDB's evaluation value. To achieve scheme optimization, the attack generator G uses forward propagation to generate a resource allocation tensor. The loss function uses the negative of the EDB score as the optimization objective. The attack generator G calculates the gradient of the loss function with respect to network parameters through backpropagation and uses optimization algorithms, such as Adam, to iteratively update the parameters. After each forward generation, a penalty term is applied to features that violate constraints, ensuring that the generated scheme is physically feasible, such as ensuring that the number of personnel in the exhibition area does not exceed capacity limits and that equipment power does not exceed the total budget.

[0069] After the first optimized scheme is generated, the parameters of the enhanced first discriminator and the second discriminator are fixed, and the attack generator is trained to optimize the first optimized scheme and generate a second optimized scheme. This makes the score of the enhanced first discriminator for the second optimized scheme lower than the score for the first optimized scheme, while the score of the second discriminator for the second optimized scheme is higher than the score for the first optimized scheme. The difference in scores between the two discriminators guides the generator to explore potential implicit constraint boundaries and optimization space, thereby simulating the contradiction and balance between expert decision-making and system model optimization.

[0070] After generating the first and second optimized schemes, they are simulated in a 3D digital twin model to calculate their actual performance scores. Scoring indicators may include space utilization, personnel load balancing, equipment efficiency, and satisfaction of safety constraints. If the difference between the actual performance scores of the first and second optimized schemes is less than a preset convergence threshold (e.g., 0.01), the iteration terminates, and the optimized scheme with the highest actual performance score during the iteration is output as the optimal resource allocation scheme. If the difference in actual performance scores is greater than or equal to the threshold, the best-performing optimized scheme in this round is used as a new starting point to continue iteratively training the attack generator until the difference in the actual performance scores of the two optimized schemes simulated in the 3D digital twin model is less than the preset convergence threshold, at which point the optimal resource allocation scheme is output.

[0071] After generating the optimal resource configuration scheme, the specific attributes and status of each resource in the optimal resource configuration scheme, such as exhibit location coordinates, personnel task assignments, equipment on / off status, and power settings, are mapped into a standardized work order format. Each work order record includes resource identifier, action type, target parameters, and execution sequence information. Then, the work order data is sent to the corresponding physical equipment control terminal through interface protocols such as Industrial Ethernet, MQTT, or PLC control commands. After parsing the work order commands, the control terminal automatically adjusts the equipment status, assigns personnel tasks, or moves exhibit positions to ensure that the resource layout in the physical space is completely consistent with the digital twin simulation scheme. This achieves closed-loop management of the optimal resource configuration scheme from virtual optimization to physical execution, and can transmit status data back in real time during execution for further monitoring and verification.

[0072] Based on a 3D digital twin model, iterative simulations of virtual and real adversarial scenarios are performed. An attack generator generates candidate resource allocation schemes under constraints. An enhanced first discriminator performs quantitative optimization scoring on the schemes, while a second discriminator simulates expert decision-making preferences for subjective evaluation. The difference between the two scores guides the generator to explore potential implicit constraint boundaries and optimization space, thereby achieving a balance between system optimization goals and expert experience judgment. Through iterative simulation and scoring comparison, the feasibility, optimizability, and security of activity resource allocation schemes are improved, realizing intelligent and verifiable resource allocation optimization in a digital twin environment. Furthermore, the optimal resource allocation scheme is transformed into executable work order instructions and sent to physical execution devices, achieving consistency between virtual optimization schemes and actual execution in the digital twin environment, and improving the implementation accuracy and execution efficiency of resource allocation for event planning.

[0073] Example 2: Based on the same inventive concept as the digital twin-based simulation method for resource allocation in event planning, this application also provides a digital twin-based simulation system for event planning resource allocation. Please refer to the appendix. Figure 2 The digital twin-based simulation system for resource allocation in event planning includes: The twin model building module 11 is used to build a three-dimensional digital twin model that maps one-to-one with the actual activity venue, exhibits, personnel, and equipment; the adversarial network deployment module 12 is used to deploy a dual-discriminator generative adversarial network in the digital twin environment, including a first discriminator, a second discriminator, and an attack generator; the implicit constraint generation module 13 is used to perform adversarial training and implicit constraint inversion on the dual-discriminator generative adversarial network to generate a set of implicit constraints; the configuration scheme output module 14 is used to convert the set of implicit constraints into penalty terms, reconstruct the first discriminator to obtain an enhanced first discriminator, and perform virtual-real adversarial inference and iteration based on the three-dimensional digital twin model, using the first discriminator, the second discriminator, and the attack generator to output the optimal resource configuration scheme and convert it into an executable work order instruction to be sent to the physical space execution device.

[0074] Furthermore, the first discriminator is a pre-trained optimizer used to output a first evaluation value for the input resource allocation scheme; the second discriminator is a scoring model trained based on historical expert decision data used to output a second evaluation value for the same resource allocation scheme corresponding to the first discriminator; the attack generator is a conditional generative adversarial network, the conditional input is the basic parameter vector of the activity, including but not limited to the number of exhibits, venue area, total budget, and personnel skill list, and the output is a resource allocation tensor.

[0075] Furthermore, the implicit constraint generation module 13 is also used to: collect training samples, alternately adversarially train the parameters of the first discriminator, the second discriminator, and the attack generator to obtain the converged dual-discriminator generative adversarial network; save all resource allocation schemes that satisfy the evaluation difference greater than a preset evaluation threshold and their corresponding first evaluation value and second evaluation value as a divergence sample set during the training process; and perform decision classification on each scheme in the divergence sample set with the deviation between the second evaluation value and the first evaluation value as the target to construct the implicit constraint set.

[0076] Furthermore, the implicit constraint generation module 13 is also used to: run decision tree classification for each scheme in the divergence sample set, using the feature vector of each scheme as input and the comparison result of the scores of the first discriminator and the second discriminator as the classification label, to generate each constraint rule corresponding to each scheme, each rule containing a specific combination of feature thresholds; calculate the confidence of each constraint rule on the validation set, and add the rules with confidence scores greater than the preset confidence threshold to the implicit constraint set.

[0077] Furthermore, the implicit constraint generation module 13 is also used for: Step 1: fixing the first discriminator and the second discriminator to construct an attack loss function by maximizing the evaluation difference between the first evaluation value and the second evaluation value, and training the attack generator; Step 2: fixing the attack generator and continuously updating the first discriminator and the second discriminator three times; alternately executing Step 1 to Step 2 until the attack loss function converges.

[0078] Furthermore, the attack loss function is set as the negative of the difference between the first evaluation value and the second evaluation value plus a gradient penalty term.

[0079] Furthermore, the configuration scheme output module 14 is also used to: for numerical features appearing in each constraint rule of the implicit constraint set, find the variable index in the resource configuration vector, and convert the feature threshold into upper and lower bound constraints of the variable; for categorical features appearing in each constraint rule of the implicit constraint set, convert them into linear equations or inequalities of binary flag variables; and use the upper and lower bound constraints and the linear equations or inequalities as penalty terms to constrain and train the first discriminator to obtain the enhanced first discriminator.

[0080] Furthermore, the configuration scheme output module 14 is also used for: collecting basic information of the activity works, inputting it into the attack generator, taking the score of the enhanced first discriminator as the optimization target, and searching for a first optimization scheme that maximizes the score of the enhanced first discriminator; fixing the enhanced first discriminator and the second discriminator, training the attack generator to optimize the first optimization scheme, generating a second optimization scheme, such that the evaluation value of the enhanced first discriminator for the second optimization scheme is lower than the evaluation value for the first optimization scheme, while the evaluation value of the second discriminator for the second optimization scheme is higher than the evaluation value of the second discriminator for the first optimization scheme; simulating the first optimization scheme and the second optimization scheme respectively in the three-dimensional digital twin model, and recording the actual performance scores; if the difference in actual performance scores is less than the convergence threshold, terminating the iteration, and outputting the scheme with the highest actual performance score at the time of iteration termination as the optimal resource configuration scheme.

[0081] Furthermore, if the difference in actual performance scores is greater than or equal to the convergence threshold, the scheme with the highest actual performance score among the schemes generated in this round is taken as the new starting point for iteration, and the attack generator is trained iteratively.

[0082] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0083] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application also intends to include such modifications and variations.

Claims

1. A simulation method for resource allocation of works activity planning based on digital twinning, characterized in that, include: Establish a three-dimensional digital twin model that maps one-to-one with the actual event venue, exhibits, personnel, and equipment; Deploy a dual-discriminator generative adversarial network in a digital twin environment, comprising a first discriminator, a second discriminator, and an attack generator; Adversarial training and implicit constraint inversion are performed on the dual discriminator generative adversarial network to generate a set of implicit constraints; The implicit constraint set is converted into a penalty term, and the first discriminator is reconstructed to obtain an enhanced first discriminator. Based on the three-dimensional digital twin model, the first discriminator, the second discriminator, and the attack generator are used to perform virtual-real adversarial simulation and iteration, output the optimal resource allocation scheme, and convert it into an executable work order instruction to be sent to the physical space execution device.

2. The digital-twin-based work-activity-planning resource- allocation simulation method of claim 1, wherein, The first discriminator is a pre-trained optimizer used to output a first evaluation value for the input resource allocation scheme; The second discriminator is a scoring model trained based on historical expert decision data, used to output a second evaluation value for the same resource allocation scheme corresponding to the first discriminator; The attack generator is a conditional generative adversarial network. The conditional input is a vector of basic parameters of the activity, including but not limited to the number of exhibits, venue area, total budget, and personnel skill list. The output is a resource allocation tensor. 3.The digital-twin-based work activity planning resource allocation simulation method of claim 1, wherein, Adversarial training and implicit constraint inversion are performed on the dual-discriminator generative adversarial network to generate a set of implicit constraints, including: Collect training samples and alternately adversarially train the parameters of the first discriminator, the second discriminator, and the attack generator to obtain the converged dual-discriminator generative adversarial network; Save all resource allocation schemes that satisfy the evaluation difference greater than the preset evaluation threshold during the training process, along with their corresponding first and second evaluation values, as a divergence sample set. Using the deviation between the second evaluation value and the first evaluation value as the objective, decision classification is performed on each scheme in the divergence sample set to construct the implicit constraint set.

4. The digital-twin-based work-activity-planning resource- allocation simulation method of claim 3, wherein, Using the deviation between the second evaluation value and the first evaluation value as the objective, decision classification is performed on each scheme in the divergence sample set to construct the implicit constraint set, including: For each scheme in the divergent sample set, a decision tree classification is run. The feature vector of each scheme is used as input, and the comparison result of the scores of the first discriminator and the second discriminator is used as the classification label to generate each constraint rule corresponding to each scheme. Each rule contains a specific combination of feature thresholds. Calculate the confidence level of each constraint rule on the validation set, and add the rules with confidence levels greater than a preset confidence threshold to the implicit constraint set.

5. The simulation method for resource allocation in work activity planning based on digital twins as described in claim 3, characterized in that, Collect training samples and alternately adversarially train the parameters of the first discriminator, the second discriminator, and the attack generator to obtain the converged dual-discriminator generative adversarial network, including: Step 1: Fix the first discriminator and the second discriminator, construct the attack loss function to maximize the evaluation difference between the first evaluation value and the second evaluation value, and train the attack generator; Step 2: Fix the attack generator and update the first discriminator and the second discriminator three times consecutively; Alternately execute steps one and two until the attack loss function converges.

6. The digital-twin-based work-activity-planning resource- allocation simulation method of claim 5, wherein, The attack loss function is set as the negative of the difference between the first evaluation value and the second evaluation value plus a gradient penalty term.

7. The digital-twin-based work-activity-planning resource- allocation simulation method of claim 1, wherein, The implicit constraint set is converted into penalty terms, and the first discriminator is reconstructed to obtain an enhanced first discriminator, including: For each numerical feature appearing in the implicit constraint set, find the variable index in the resource allocation vector and convert the feature threshold into upper and lower bound constraints of the variable. For each categorical feature appearing in each constraint rule in the implicit constraint set, convert it into a linear equation or inequality of binary flag variables; The first discriminator is trained with the upper and lower bound constraints and the linear equations or inequalities as penalty terms to obtain the enhanced first discriminator. 8.The digital-twin-based work-activity-planning resource-allocating simulation method of claim 1, wherein, Based on the aforementioned three-dimensional digital twin model, the first discriminator, the second discriminator, and the attack generator perform virtual-real adversarial simulation iterations to output an optimal resource allocation scheme, including: Collect basic information about the works in the activity, input it into the attack generator, take the score of the enhanced first discriminator as the optimization target, and search for the first optimization scheme that maximizes the score of the enhanced first discriminator; Fix the enhanced first discriminator and the second discriminator, train the attack generator to optimize the first optimization scheme, generate a second optimization scheme, such that the evaluation value of the enhanced first discriminator for the second optimization scheme is lower than the evaluation value for the first optimization scheme, while the evaluation value of the second discriminator for the second optimization scheme is higher than the evaluation value of the second discriminator for the first optimization scheme. The first optimization scheme and the second optimization scheme are simulated in the three-dimensional digital twin model respectively, and the actual performance scores are recorded. If the difference in actual performance scores is less than the convergence threshold, the iteration is terminated, and the scheme with the highest actual performance score at the time of iteration termination is output as the optimal resource allocation scheme. 9.The digital-twin-based work-activity-planning resource-allocating simulation method of claim 8, wherein, If the difference in actual performance scores is greater than or equal to the convergence threshold, the scheme with the highest actual performance score among the schemes generated in this round is taken as the new starting point for iteration, and the attack generator is trained iteratively.

10. A work activity planning resource allocation simulation system based on digital twinning, characterized in that, For implementing the digital twin-based simulation method for resource allocation in event planning, as described in any one of claims 1 to 9, the digital twin-based simulation system for resource allocation in event planning, as described in claims 1 to 9, comprises: The twin model creation module is used to create three-dimensional digital twin models that are mapped one-to-one with the actual event venue, exhibits, personnel, and equipment. The adversarial network deployment module is used to deploy a dual-discriminator generative adversarial network in a digital twin environment, including a first discriminator, a second discriminator, and an attack generator; The implicit constraint generation module is used to perform adversarial training and implicit constraint inversion on the dual discriminator generative adversarial network to generate a set of implicit constraints. The configuration scheme output module is used to convert the implicit constraint set into penalty terms, reconstruct the first discriminator to obtain an enhanced first discriminator, and perform virtual-real adversarial simulation iteration with the first discriminator, the second discriminator and the attack generator based on the three-dimensional digital twin model, output the optimal resource configuration scheme, and convert it into an executable work order instruction to be sent to the physical space execution device.

Images