Generative intelligent optimization method and device

Abstract

The invention provides a generative intelligent optimization method and device, and relates to the technical field of generative artificial intelligence optimization computation.The method comprises the steps that a mathematical model containing decision variables, an objective function and constraint conditions is established for a target optimization problem, and scene parameters are pre-calculated through a traditional optimization algorithm to obtain a theoretical optimal solution; constructing a training data set containing scene parameters and a theoretical optimal solution; training the data set through a conditional generation type artificial intelligence model, minimizing the difference between a generation decision and a theoretical optimal solution, and establishing a mapping relation from scene parameters to an optimal decision, so that the mapping relation implicitly learns a constraint condition satisfaction mode; current scene parameters are collected in real time and input into the trained model, a near-optimal decision scheme is generated through reverse denoising or hidden variable decoding, and the near-optimal decision scheme is applied to real-time scenes such as industrial control. According to the method, a high-quality solution close to theoretical optimum is realized, the online solution time consumption is remarkably reduced, and the real-time requirement is met.

Description

Technical Field

[0001] This invention relates to the field of generative artificial intelligence optimization computing technology, and in particular to a generative intelligent optimization method and apparatus. Background Technology

[0002] Generative artificial intelligence, as a significant technological breakthrough in the field of optimization computing, is widely applied in real-time decision-making scenarios such as industrial control, logistics scheduling, and financial transactions. Among related technologies, a complete technical system has been constructed from mathematical modeling to decision generation through the collaborative operation of traditional optimization algorithms and machine learning models. Specifically, this system covers the entire process from problem abstraction and modeling to algorithm solution, including key steps such as defining decision variables, constructing objective functions, and handling constraints. Traditional numerical methods such as linear programming and integer programming achieve theoretically optimal solutions through precise mathematical modeling, while data-driven methods such as reinforcement learning approximate optimal decisions through policy iteration.

[0003] However, in existing optimization methods, when directly using traditional numerical algorithms to solve complex problems, the computational complexity increases exponentially with the problem size, making it difficult to meet real-time requirements. Therefore, existing technologies typically employ data-driven methods such as reinforcement learning, but these have significant limitations: First, the training process relies on a large amount of simulation data, and convergence is difficult to guarantee; second, the output results are prone to violating physical constraints, for example, in power system scheduling, they may generate decision schemes exceeding the generator output range; third, the lack of a theoretical optimality guidance mechanism leads to a negative correlation between solution quality and solution time. While traditional methods can guarantee the theoretical optimality of the solution, their computational time increases non-linearly with the problem size. Specifically, in logistics route planning scenarios, when the number of nodes exceeds 100, the solution time may extend from minutes to hours, severely restricting practical applications. Summary of the Invention

[0004] The present invention aims to at least partially solve one of the technical problems in the related art.

[0005] Therefore, the first objective of this invention is to propose a generative intelligent optimization method.

[0006] The second objective of this invention is to propose a generative intelligent optimization device.

[0007] The third objective of this invention is to provide an electronic device.

[0008] The fourth objective of this invention is to provide a computer-readable storage medium.

[0009] The fifth objective of this invention is to provide a computer program product.

[0010] To achieve the above objectives, a first aspect of the present invention proposes a generative intelligent optimization method, comprising: S1. A mathematical model containing decision variables, objective function and constraints is established for the target optimization problem, and the theoretical optimal solution is obtained by pre-calculating the scenario parameters based on traditional optimization algorithms. S2, Construct a training dataset containing scene parameters and theoretical optimal solutions; S3 uses a conditional generative artificial intelligence model to train the training dataset. By minimizing the difference between the generated decision and the theoretical optimal solution, a mapping relationship between scene parameters and the optimal decision is established, so that the model implicitly learns how to satisfy the constraints. S4 collects parameters of the current scene in real time and inputs them into the trained generative model. Through the reverse denoising process or latent variable decoding, it generates a near-optimal decision scheme, which can be directly applied to real-time decision-making scenarios.

[0011] Optionally, the step of establishing a mathematical model for the target optimization problem, including decision variables, objective function, and constraints, and obtaining the theoretical optimal solution by pre-calculating the scenario parameters based on traditional optimization algorithms, further includes: S11, use KKT conditions to reduce variables in the objective optimization problem, and use the Lagrange multiplier method to transform the constraints into gradient conditions for the optimization variables; S12 uses the dual decomposition method to break down a large-scale optimization problem into multiple subproblems, and constructs the theoretical optimal solution of the overall problem by coordinating the optimal solutions of the subproblems.

[0012] Optionally, the construction of the training dataset containing scene parameters and theoretical optimal solutions further includes: S21, the specific scenario parameters include cargo volume, transportation distance and time window constraints in logistics scenarios, channel gain, interference intensity and QoS requirements in communication scenarios, and node load, line capacity and power generation cost function in power grid scenarios; S22, the theoretical optimal solution specifically includes the shortest distance sequence of transportation paths, the optimal allocation matrix of resource block lengths, and the Pareto optimal solution set of generator output.

[0013] Optionally, the step of training the training dataset using a conditional generative artificial intelligence model, and establishing a mapping relationship between scene parameters and optimal decisions by minimizing the difference between the generated decisions and the theoretical optimal solutions, thereby enabling the model to implicitly learn how to satisfy constraints, further includes: S31, the conditional generative artificial intelligence model adopts the UNet architecture combined with a multi-head attention mechanism to enhance the ability to model structured condition variables through cross-layer feature interaction; S32, minimizing the difference between the generated decision and the theoretical optimal solution is calculated using the mean squared error loss function, as shown in the following formula: , in To generate decision variables, This is to pre-calculate the optimal solution.

[0014] Optionally, the step of collecting parameters of the current scene in real time and inputting them into the trained generative model, generating a near-optimal decision scheme through a reverse denoising process or latent variable decoding, and directly applying it to the real-time decision-making scenario, further includes: S41, The reverse denoising process iteratively optimizes the initial noise vector. At every step Calculate the conditional gradient And update status ,in The learning rate; S42, latent variable decoding employs a variational inference framework, by maximizing the lower bound of evidence. Achieve the generation of near-optimal solutions.

[0015] Optionally, the method further includes: S5. Feature extraction is performed on the pre-calculated theoretical optimal solution to generate an enhanced feature vector containing gradient information of decision variables. ; S6 will enhance the feature vector Compared with the original scene parameters The data is concatenated and used as input conditions for the generative model. This is to improve the model's fitting accuracy to the optimal solution distribution.

[0016] To achieve the above objectives, a second aspect of the present invention provides a generative intelligent optimization device, comprising: The modeling and pre-computation module is used to build a mathematical model containing decision variables, objective functions and constraints for the target optimization problem, and to pre-compute the scene parameters based on traditional optimization algorithms to obtain the theoretical optimal solution. The dataset building module is used to build a training dataset containing scene parameters and theoretical optimal solutions; The model training module is used to train the training dataset using a conditional generative artificial intelligence model. By minimizing the difference between the generated decision and the theoretical optimal solution, a mapping relationship between scene parameters and the optimal decision is established, so that the model implicitly learns how to satisfy the constraints. The real-time decision generation module is used to collect parameters of the current scene in real time and input them into the trained generative model. It generates a near-optimal decision scheme through reverse denoising or latent variable decoding.

[0017] To achieve the above objectives, a third aspect of the present invention provides an electronic device, comprising: a processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method as described in any one of the first aspects.

[0018] To achieve the above objectives, a fourth aspect of the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of the first aspects.

[0019] To achieve the above objectives, a fifth aspect of the present invention provides a computer program product that, when executed by a processor, implements the method described in any one of the first aspects.

[0020] The present invention provides a generative intelligent optimization method, device, electronic device and storage medium that can achieve high-quality solutions close to the theoretical optimum, while reducing the online solution time to the millisecond level, significantly improving the real-time solution efficiency of complex optimization problems, and effectively avoiding the risk of constraint violation through training with constraint datasets.

[0021] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0022] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart illustrating a generative intelligent optimization method provided in an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the principle and flow of a generative intelligent optimization method provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of the structure of a generative intelligent optimization device provided in an embodiment of the present invention. Detailed Implementation

[0023] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0024] This invention provides a generative intelligent optimization method. Figure 1 This is a flowchart illustrating a generative intelligent optimization method provided in an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the principle and flow of a generative intelligent optimization method provided in an embodiment of the present invention. Figure 1 , Figure 2 As shown, the method includes the following steps: Step S1: Establish a mathematical model for the target optimization problem, including decision variables, objective function and constraints, and pre-calculate the scenario parameters based on traditional optimization algorithms to obtain the theoretical optimal solution.

[0025] Specifically, establishing a mathematical model for the target optimization problem is the fundamental step in the entire technical solution. The core of this step lies in abstracting the practical problem into a standard mathematical optimization form, including defining decision variables. objective function and the set of constraints In some implementations, the objective function can be linear, nonlinear, or mixed-integer, depending on the specific application scenario, such as minimizing transportation costs in logistics route planning or maximizing energy efficiency in power system scheduling.

[0026] In terms of implementation, the problem must first be modeled and analyzed to clarify the physical meaning and value range of the variables. For example, in the communication resource allocation problem, decision variables might include resource block length. Power distribution The objective function might be to maximize system throughput, while constraints include channel capacity limitations and power budget constraints. Subsequently, traditional optimization algorithms (such as branch and bound, interior-point methods, and genetic algorithms) are used to pre-calculate parameters for different scenarios to obtain the theoretically optimal solution. Scenario parameters typically include order demand, channel gain matrix, and grid node load, and their value range needs to be set according to actual business requirements, such as channel gain. Node load .

[0027] At the parameter level, the pre-computation process must ensure coverage of a sufficient number of scenario parameter combinations to construct a representative training dataset. Typically, scenario parameter sampling should meet uniformity requirements or a sampling strategy based on the actual distribution, with a sample size recommended to be no less than 10^4 to ensure the model's generalization ability. Simultaneously, the accuracy requirements of the objective function should match the actual system requirements; for example, in power dispatching, the objective function error needs to be controlled within a certain range. Within.

[0028] This step is widely applicable in complex optimization scenarios such as logistics, communication, and energy dispatch. For example, in power systems, pre-calculation can generate the optimal generator output combination based on historical load data and weather forecast information. In terms of technical effectiveness, this step provides high-quality ground truth data for the subsequent training of generative models, ensuring that the model can learn the distribution characteristics of the theoretically optimal solution, thereby achieving rapid and high-quality decision generation in the online phase. Furthermore, through pre-calculation results that strictly satisfy constraints, the model implicitly learns the feasible region during training, significantly reducing the risk of constraint violations in the online generation phase.

[0029] Furthermore, S1 includes: S11. The KKT conditions are used to reduce the variables in the objective optimization problem, and the Lagrange multiplier method is used to transform the constraints into gradient conditions for the optimization variables.

[0030] Specifically, by applying the Karush-Kuhn-Tucker (KKT) conditions to reduce variables in the objective optimization problem, the complexity of the optimization model is reduced, thereby improving the solution efficiency. The KKT conditions are the core theoretical foundation for constrained optimization in nonlinear programming problems. They transform the equality and inequality constraints in the original optimization problem into gradient conditions regarding the optimization variables using the Lagrange multiplier method, thus converting the constrained optimization problem into an unconstrained one. Specifically, for the objective function... With constraints and Construct the Lagrange function ,in Lagrange multipliers constrained by inequality For multipliers of equality constraints.

[0031] Furthermore, by solving the condition that the gradient of the Lagrange function is zero, i.e. And combine the original constraints with the dual feasibility conditions (such as...) This process yields candidate optimal solutions that satisfy the KKT conditions. In one embodiment of the invention, this step is used to reduce variables in the original optimization problem. Specifically, by identifying variable relationships that satisfy the KKT conditions, some variables are expressed as functions of other variables, thereby reducing the dimensionality of the optimization variables. For example, in the power system dispatching problem, if there is a strong coupling relationship between generator output and load demand, the output variable can be expressed as a function of load using the KKT conditions, thus requiring only the processing of key variables in subsequent optimization.

[0032] Optionally, the accuracy of variable reduction can be improved by setting a gradient error threshold. To control, usually set to arrive To ensure that the reduced variables retain the optimization properties of the original problem, the KKT solver can be used to improve computational efficiency. This solver supports parallel processing of large-scale problems and sparse matrix optimization.

[0033] This step plays a crucial role in this invention. Through variable reduction guided by theoretical optimality, it provides a structured, low-dimensional decision variable space for the subsequent training of the generative model, thereby significantly improving the model's convergence speed and generalization ability. Simultaneously, the variable reduction process strictly adheres to the constraints of the original problem, ensuring that the generative model learns the decision distribution that satisfies the constraints during the training phase, thus guaranteeing rapid and reliable decision-making in the online phase.

[0034] S12 uses the dual decomposition method to break down a large-scale optimization problem into multiple subproblems, and constructs the theoretical optimal solution of the overall problem by coordinating the optimal solutions of the subproblems.

[0035] Specifically, this invention employs the Dual Decomposition Method to structurally decompose a large-scale optimization problem, thereby transforming the original problem into multiple subproblems that can be solved in parallel. This method is based on Lagrange duality theory, and by introducing Lagrange multipliers to relax coupling constraints, it decomposes the original problem into several local optimization problems and a global coordination problem. Specifically, let the original optimization problem be:

[0036] in As decision variables, Let be the objective function. Let be the constraint function. In dual decomposition, if a separable structure exists, meaning the objective function and constraint functions can be partitioned into subsystems, then the problem can be rewritten as:

[0037] in For the first Decision variables for each sub-problem Let be the coupling matrix. This is the global constraint vector. Lagrange multipliers are introduced... The original problem is transformed into the dual problem:

[0038] In terms of implementation, each sub-problem These problems can be solved independently, typically using numerical optimization algorithms such as gradient descent, subgradient method, or interior point method. Global coordination problems, on the other hand, are solved by updating... This process is used to aggregate solutions to subproblems and progressively satisfy constraints. During iteration, solutions to subproblems are fed back to the coordinator via a communication link. The coordinator determines convergence based on the dual gap, which is typically set as the dual gap being less than a preset threshold. ,For example or Specifically, it depends on the precision requirements of the problem.

[0039] This step plays a crucial role in this invention, especially in the pre-optimization process. Dual decomposition significantly reduces the solution complexity of large-scale problems and improves pre-computation efficiency. For example, in power system dispatching, if the power grid is divided into multiple regions, the generator output optimization in each region can be treated as a sub-problem. By coordinating global power balance constraints, the theoretically optimal solution can be generated quickly. Furthermore, this method can be combined with KKT conditions to further reduce the dimensionality of variables and improve computational stability and convergence speed.

[0040] Step S2: Construct a training dataset containing scene parameters and theoretical optimal solutions.

[0041] Specifically, this step aims to construct a high-quality dataset for training generative intelligent optimization models. Its core lies in systematically pairing scene parameters from real-world optimization problems with their corresponding theoretical optimal solutions, thereby providing precise supervision signals for subsequent model training. In some implementations, the construction of this dataset includes sampling scene parameters, solving the optimization model, and storing and structuring the optimal solutions.

[0042] From a technical implementation perspective, the selection of scenario parameters needs to cover the key input variables of the target optimization problem, such as the quantity of goods in logistics scheduling. In communication systems, this is called channel gain. In power grid dispatching, it refers to the load at the node. These parameters typically exhibit continuous or discrete distribution characteristics, requiring appropriate sampling based on the actual business scenario. Optionally, Latin Hypercube Sampling (LHS) or Monte Carlo Sampling methods can be employed to ensure uniform coverage of the parameter space. The number of samples is generally set to... This is to meet the modeling requirements of generative models for data distribution.

[0043] After the parameters are determined, traditional optimization algorithms (such as branch and bound, interior point method, heuristic algorithms, etc.) need to be called to solve for each sampling point to obtain the corresponding theoretical optimal solution. For example, in logistics route optimization, the output is the transportation route. In communication resource allocation, the output is the resource block length. In power grid dispatching, the output is the generator output. To ensure the global optimality of the solution, the solution accuracy of the optimization algorithm must meet the following requirements. Furthermore, all constraints must be strictly met during the solution process, including resource limitations, time constraints, and physical feasibility.

[0044] Furthermore, the constructed dataset should be stored in a structured manner, typically using NumPy arrays or HDF5 file format, to facilitate efficient reading and processing during subsequent model training. Each sample consists of input condition variables (scene parameters) and output target variables (optimal solution), forming a mapping relationship. ,in Represents the scene parameter vector. This represents the optimal decision vector.

[0045] This step plays a fundamental role in the entire technical solution. The dataset it constructs not only provides ground truth samples for training the generative model, but also ensures that the model has high-quality decision output capabilities during the generation process by being guided by theoretically optimal solutions. Furthermore, because the training data strictly satisfies the constraints, the model can implicitly learn the feasible region during the deployment phase, thereby significantly reducing the risk of constraint violations.

[0046] Furthermore, S2 includes: S21, the specific scenario parameters include cargo volume, transportation distance and time window constraints in logistics scenarios, channel gain, interference intensity and QoS requirements in communication scenarios, and node load, line capacity and power generation cost function in power grid scenarios.

[0047] Specifically, this step involves defining and modeling key parameters for different application scenarios, which is a fundamental step in building the training dataset in the entire generative intelligent optimization method. In some implementations, scenario parameters are divided into three typical scenarios: logistics, communication, and power grid. Each scenario contains several variables with physical meaning and business constraints to describe the system's operating state and optimization objectives.

[0048] In logistics scenarios, specific parameters include cargo volume. Transportation distance and time window constraints Among them, cargo volume Transport distance is usually expressed in units of quantity or volume. Time window constraints, measured in kilometers or between nodes. Then it is defined as a time interval These parameters are used to limit the earliest and latest delivery times of goods. Together, they determine the optimization space for decision variables such as route planning and scheduling order.

[0049] In communication scenarios, scenario parameters include channel gain. Interference intensity and QoS requirements Channel gain Interference intensity is usually expressed in dB or linear values, reflecting the transmission quality of the channel; Measured in power units (e.g., dBm), it is used to assess signal interference levels in multi-user environments; QoS requirements. These include metrics such as throughput, latency, and bit error rate, which are usually expressed in numerical or range form and are used to constrain the performance boundaries of communication resource allocation.

[0050] In the power grid scenario, scenario parameters include node load. Line capacity and power generation cost function Node load The real-time power demand at each node is expressed in MW; line capacity. The physical upper limit for power transmission, measured in MW or A; the power generation cost function. This is a nonlinear function, and a quadratic cost model is typically used. ,in These are the economic parameters of the generator, used to evaluate the operating costs at different output levels.

[0051] By using these scenario parameters as conditional variables, this step provides structured input for the subsequent training of the generative model, ensuring that the model can learn the mapping relationship for generating optimal decisions under specific conditions. In practical applications, these parameters can be collected in real time or preset by sensors, scheduling systems, or historical data, thereby supporting the rapid deployment and generalization of the model in different business scenarios. This implementation not only improves the model's input representation capability but also significantly enhances the feasibility and stability of the generated results through guidance from the theoretically optimal solution, providing a high-quality conditional input foundation for subsequent online decision-making.

[0052] S22, the theoretical optimal solution specifically includes the shortest distance sequence of transportation paths, the optimal allocation matrix of resource block lengths, and the Pareto optimal solution set of generator output.

[0053] Specifically, this step requires establishing corresponding optimization models for different application scenarios (such as logistics route planning, communication resource allocation, and power system dispatching), and clarifying the decision variables, objective functions, and constraints. For example, in a logistics scenario, the decision variable is the sequence of nodes in the transportation route, and the objective function is the total length of the route. ,in Indicates the first The distance of the path segment is constrained by conditions including vehicle capacity and time window limits; in a power system, the objective function is the total generation cost. ,in Indicates the first The output of the generator To correspond to the cost coefficient, the constraints include power balance, upper and lower limits of generator output, etc.

[0054] In terms of implementation, this step systematically traverses the scenario parameters using branch and bound, interior-point methods, or heuristic algorithms (such as genetic algorithms and simulated annealing) to find the theoretically optimal solution under the corresponding conditions. For example, in communication resource allocation, the optimal allocation matrix for resource block length... ,in For the number of users, The total number of resource blocks, matrix elements Indicates user In resource block The allocation ratio must meet the following requirements. To ensure that resources are not over-allocated.

[0055] Furthermore, the Pareto optimal solution set generated in this step can be used for multi-objective optimization scenarios, such as simultaneously optimizing power generation costs and carbon emissions in a power system. By solving the problem using a multi-objective optimization algorithm (such as NSGA-II), a set of non-dominated solutions can be obtained, and their distribution characteristics will serve as a reference for training the generative model.

[0056] This step plays a crucial role in the overall technical solution, as its output constitutes the training dataset for the generative model, ensuring that the model can learn the distribution characteristics of the theoretically optimal solution. Through pre-calculated results that strictly satisfy the constraints, the model can implicitly learn the feasible region during the online deployment phase, thereby significantly reducing the risk of constraint violations and improving the stability and practicality of the solution.

[0057] Step S3: The training dataset is trained using a conditional generative artificial intelligence model. By minimizing the difference between the generated decision and the theoretical optimal solution, a mapping relationship between scene parameters and the optimal decision is established, so that the model implicitly learns how to satisfy the constraints.

[0058] Specifically, during model training, a conditional generative artificial intelligence model is used to train the constructed training dataset. Its core objective is to establish a mapping relationship from scene parameters to optimal decisions by minimizing the difference between the generated decisions and the theoretical optimal solution, enabling the model to implicitly learn and satisfy various constraints in the optimization problem. In some implementations, this step is modeled based on a denoised diffusion probability model (DDPM), but depending on the characteristics of the problem, generative model architectures such as conditional variational autoencoders (CVAE) or normalizing flow models can be selected.

[0059] Specifically, the model's input is a scene parameter vector. ,in The dimensions representing the scenario parameters, such as channel gain, interference intensity, and user distribution in a communication system, are used as inputs; the output is the corresponding optimal decision variable. Examples include resource block allocation and power control strategies. The training objective is to minimize the generation decision. With the theoretical optimal solution The loss function between these parameters typically uses the mean squared error (MSE) as the optimization objective, i.e.:

[0060] In terms of model structure design, if DDPM is adopted, its core component is the UNet architecture, combined with an attention mechanism to enhance the modeling ability for structured conditional inputs. The model progressively adds Gaussian noise to the decision variables through a forward diffusion process, and then recovers the optimal decision from the noise through a reverse denoising process. Each denoising operation depends on the current noise state and scene parameters. The combined inputs are used to generate conditions.

[0061] In terms of parameter settings, diffusion models typically set the total number of diffusion steps. Noise scheduling employs linear or cosine decay methods to control the smoothness of noise injection. During training, the model uses the Adam optimizer to ensure full convergence.

[0062] This step plays a crucial role in the overall technical solution. By transforming the traditional optimization problem into a condition generation task, the model can learn how to satisfy constraints during the training phase. This allows for the generation of highly feasible and high-quality decision solutions during online deployment without the need for explicit constraint checks. Its technical value lies in significantly improving online decision-making efficiency, reducing computational resource consumption, and enhancing the model's generalization ability to unseen scenarios.

[0063] Furthermore, S3 includes: S31, the conditional generative artificial intelligence model adopts the UNet architecture combined with a multi-head attention mechanism to enhance the ability to model structured condition variables through cross-layer feature interaction.

[0064] Specifically, the conditional generative artificial intelligence model employed in this invention is based on the UNet architecture and incorporates a multi-head attention mechanism to enhance the model's ability to model structured condition variables. UNet is an encoder-decoder structure widely used in image generation and segmentation tasks. It achieves multi-scale feature fusion through skip connections, thereby improving the structural consistency and detail fidelity of the generated results. In this invention, UNet is used to model the complex mapping relationship between decision variables and condition variables, where condition variables (such as cargo quantity in logistics, channel gain in communication, and node load in power grids) are embedded into various levels of the model to guide the generation process.

[0065] Furthermore, a multi-head attention mechanism is introduced into each decoding stage of UNet to capture the long-range dependencies between conditional and decision variables. Specifically, in each attention module, the model treats the conditional variable as a query vector and the features of the current decoding layer as key and value vectors. By calculating attention weights, cross-layer feature interaction and information transfer are achieved. This mechanism can be represented as:

[0066] in, , , These are query, key, and value matrices, respectively. Let be the dimension of the key vector. This refers to the sequence length or feature map size. Multi-head attention improves the model's expressive power and robustness by computing multiple attention heads in parallel and then linearly combining them.

[0067] Regarding parameter settings, UNet's encoder and decoder typically use ResNet-34 or ResNet-50 as the backbone network to extract deep semantic features of the condition variables. The number of heads in the attention module is generally set to 8 or 16, and the embedding dimension is 256 or 512, depending on the complexity of the condition variables and the dimension of the decision variables. Furthermore, the mean squared error (MSE) is used as the loss function during model training, i.e.:

[0068] in, Decision variables generated for the model, For the pre-calculated theoretical optimal solution, This represents the number of training samples.

[0069] This step plays a crucial role in the overall technical solution. By modeling structured condition variables, the generative model can accurately perceive the constraints and objectives of the input scenario, thereby rapidly generating high-quality decision solutions with low constraint violation risk in the online phase. Its technical value lies in significantly improving the generative model's adaptability to complex optimization problems, while ensuring the feasibility and stability of the generated results.

[0070] S32, minimizing the difference between the generated decision and the theoretical optimal solution uses the mean squared error loss function. ,in To generate decision variables, This is to pre-calculate the optimal solution.

[0071] Specifically, in one embodiment of the present invention, the training objective of the generative model is to ensure that the decision variables output by the model statistically approximate the optimal solution by minimizing the difference between the generated decisions and the theoretical optimal solution. To this end, Mean Squared Error (MSE) is used as the loss function, and its mathematical expression is:

[0072] in, The model represents the first time. Decision variables generated on a sample This represents the pre-calculated theoretical optimal solution corresponding to this sample. This represents the dimension of the decision variable in a single sample. The loss function quantifies the deviation between the model output and the ideal solution by calculating the average of the squared Euclidean distances between the generated decision and the optimal solution, thereby guiding the model towards the optimal solution distribution during training.

[0073] At the technical implementation level, this loss function is embedded in the training process of the generative model as the gradient signal for updating the model parameters. Taking the Denoising Diffusion Probability Model (DDPM) as an example, the model gradually generates decision variables during the reverse denoising process. The output of each step is compared with the corresponding optimal solution, and the model parameters are updated through a gradient descent algorithm (such as the Adam optimizer). In the early stages of training, the model output may deviate significantly from the optimal solution, but as the training iterations proceed, the loss value gradually converges, and the model's generative ability tends to stabilize.

[0074] Regarding parameter settings, the loss function needs to be calculated in each training batch and is usually consistent with the model's output dimension. In actual training, a loss threshold can be set. ,when At this point, the model is considered to have sufficient fitting ability. Furthermore, to improve training stability, a weighted loss mechanism can be introduced, assigning higher weights to key decision variables. For example, in power dispatching scenarios, larger coefficients can be set for variables that directly affect system stability, such as generator output.

[0075] This step is central to the entire technical solution. By using the theoretically optimal solution as a monitoring signal, the model not only learns the distribution characteristics of the decision variables but also implicitly masters how the constraints are satisfied. In practical applications, such as logistics route optimization and communication resource allocation, minimizing this loss function directly corresponds to improved system performance, such as reduced transportation costs or increased spectrum utilization. Simultaneously, because the training data strictly satisfies the constraints, the probability of constraint violations during model generation is significantly reduced, thereby improving the system's security and reliability.

[0076] Step S4: Collect parameters of the current scene in real time and input them into the trained generative model. Generate a near-optimal decision scheme through reverse denoising process or latent variable decoding, which can be directly applied to real-time decision-making scenarios such as industrial control, logistics scheduling or power systems.

[0077] Specifically, the system collects parameters (i.e., condition variables) of the current scene in real time and inputs them into a trained generative model to quickly generate near-optimal decision solutions. This step is a key application link in the framework of this invention. Its technical implementation is based on the reasoning capability of the conditional generative model, combined with a reverse denoising process or a latent variable decoding mechanism, thereby achieving efficient and high-quality decision output in real-time decision-making scenarios such as industrial control, logistics scheduling, or power systems.

[0078] In some implementations, real-time collected scenario parameters include, but are not limited to, cargo volume, order distribution, and transportation time windows in logistics scenarios; channel gain, interference level, and resource block status in communication systems; and node load, line capacity, and generator status in power systems. These parameters are typically represented in vector or tensor form; for example, in a power system, the input could be... ,in Indicates the first Real-time power requirements of each node Indicates the first The current load status of each line. The sampling frequency can be set according to the system response requirements, such as tens or even hundreds of times per second in industrial control.

[0079] The model inference process varies depending on the chosen generative model. If a Denoising Diffusion Probability Model (DDPM) is used, the inverse denoising process is employed to deduce the initial noise. Stepwise generation of decision variables Each iteration uses conditional input. The denoising process is guided to ultimately output a decision scheme that satisfies the constraints. If a conditional variational autoencoder (CVAE) is used, latent variables are extracted through the encoder. The decoder then generates the decision variables. In actual deployments, the computation time for this process is typically controlled within 10ms, meeting real-time requirements.

[0080] The technical value of this step lies in its ability to compress the traditional optimization process, which originally required seconds or even minutes, to the millisecond level through the rapid inference capability of generative models, while maintaining the quality of the solution close to the theoretical optimum. Because the training data strictly satisfies the constraints, the decision schemes output by the model during inference also have an extremely low probability of constraint violation, thus significantly reducing the system's operational risk. Furthermore, this method possesses good out-of-distribution generalization ability, capable of handling scenario parameter combinations not present in the pre-computation stage, improving the system's adaptability and robustness.

[0081] Furthermore, S4 includes: S41, The reverse denoising process iteratively optimizes the initial noise vector. At every step Calculate the conditional gradient And update status ,in This is the learning rate.

[0082] Specifically, in one embodiment of the present invention, the inverse denoising process is the core mechanism for the diffusion model to generate high-quality decision variables. This process iteratively optimizes the initial noise vector. At every step Calculate the conditional gradient And update the state based on gradient information. ,in The learning rate is used. This step essentially uses gradient descent to gradually reduce the noise vector. Push towards current scene parameters The relevant optimal decision distribution.

[0083] From a technical implementation perspective, the reverse denoising process is based on a pre-trained diffusion model, typically employing the UNet architecture combined with an attention mechanism to enhance its ability to model structured conditional inputs. In each iteration, the model receives the current noise state. With scene parameters and output gradient direction The gradient is determined by the loss function. right The partial derivatives of the formula are used to measure the difference between the current state and the optimal solution. Learning rate. The settings are usually in Within the scope, dynamic adjustments can be made based on the convergence characteristics of the problem, or adaptive optimization algorithms (such as Adam) can be used for control.

[0084] At the parameter level, the key parameters of this process include the number of iterations. Noise scheduling strategies (such as linear or cosine noise attenuation), gradient update methods (such as full gradient or stochastic gradient descent), etc. Typically, The range of values ​​is within The noise level is between 1000 and 1000 to balance generation quality and computational efficiency. The noise scheduling strategy determines the magnitude of denoising at each step, which directly affects the stability and accuracy of the generated results.

[0085] In application scenarios, this step is widely applicable to situations requiring the real-time generation of feasible and high-quality decisions, such as power system dispatching, communication resource allocation, and logistics route planning. In these scenarios, the system receives real-time input of current scenario parameters. The model quickly generates near-optimal decision variables through a reverse denoising process, meeting the millisecond-level response requirement.

[0086] The technical advantage of this step lies in the fact that, through iterative optimization, the model can gradually approach the theoretical optimal solution. Furthermore, because the training data strictly satisfies the constraints, the generated decision variables have an extremely low probability of constraint violation in actual deployment. In addition, the computational complexity of this process is linearly related to the problem size, significantly outperforming the exponential complexity of traditional optimization methods, thus achieving a dual improvement in computational efficiency and solution quality.

[0087] S42, latent variable decoding employs a variational inference framework, by maximizing the lower bound of evidence. Achieve the generation of near-optimal solutions.

[0088] Specifically, this step involves using a variational inference framework within generative intelligent optimization methods to decode latent variables in order to generate near-optimal solutions. The core technology lies in maximizing the Evidence Lower Bound (ELBO). Thus, in a given scene parameter Under the given conditions, generate decision variables that satisfy the constraints and are close to the theoretical optimum. .

[0089] At the technical implementation level, this step is based on the structure of a Variational Autoencoder (VAE), modeling the optimization problem as a condition generation task. Specifically, latent variables... The latent representation of decision variables, and It is an approximate posterior distribution of the encoder network output, usually using a Gaussian distribution. Decoder network The model is responsible for reconstructing the original input or generating the target decision variable from the latent variables. During training, ELBO is maximized to achieve a balance between the model's reconstruction ability and the prior matching of the latent variable distribution.

[0090] Regarding the parameters and metrics, ELBO's first item Represents reconstruction likelihood, measuring the model's performance in a given context. Time The fitting ability is usually achieved using mean squared error (MSE) or cross-entropy loss as specific implementations. (Second term) This is the KL divergence, used to measure the difference between the approximate posterior and prior distributions, where... It is usually set to a standard normal distribution. KL weight Optionally, it can be introduced to adjust the relative importance of the two items, forming -VAE framework, thereby enhancing the model's ability to control the structure of latent variables.

[0091] In application scenarios, this step is suitable for model training, especially when dealing with optimization problems with complex constraints, such as power system scheduling and communication resource allocation. By using the optimal solution obtained by traditional optimization algorithms as training data, the model implicitly learns the distribution characteristics of the feasible region during training, thereby generating high-quality decisions that satisfy the constraints during the online deployment phase.

[0092] In terms of technical effectiveness, this step effectively reduces the uncertainty of latent variables through a variational inference framework, improving the stability and interpretability of the generated decisions. Simultaneously, because the training data strictly satisfies the constraints, the model's output solution has an extremely low probability of constraint violation in actual deployment, significantly improving the system's robustness and security. Furthermore, by introducing latent variables, this method enhances the model's responsiveness to input conditions, providing a favorable initial state for subsequent diffusion or decoding processes, thereby accelerating online generation and improving the quality of the solutions.

[0093] Step S5: Extract features from the pre-calculated theoretical optimal solution to generate an enhanced feature vector containing gradient information of the decision variables. .

[0094] Specifically, this step involves extracting features from the pre-calculated theoretical optimal solution to generate an enhanced feature vector containing gradient information of the decision variables. The core objective of this step is to extract gradient information of the optimal solution in the decision variable space, thereby providing richer feature representations for the subsequent training of generative models and improving the model's ability to fit the distribution of the optimal solution.

[0095] From a technical implementation perspective, this feature extraction process is based on a mathematical model of an optimization problem, and it performs pre-computation on each scene parameter. Solve the corresponding objective function And calculate its relation to the decision variable. gradient Gradient information reflects the sensitivity of the objective function to input parameters at the optimal solution, helping the model understand the impact mechanism of parameters on optimal decision-making in different scenarios. In practical implementation, gradient calculation can be performed using automatic differentiation tools (such as PyTorch and TensorFlow) or numerical differentiation methods (such as the central difference method). For large-scale optimization problems, sparse gradient calculation strategies can be combined to reduce memory and computational overhead.

[0096] At the parameter level, the dimension of the gradient vector is consistent with the number of decision variables, that is... The range of values ​​for each gradient component depends on the properties of the objective function. For example, in logistics scheduling problems, the gradient might represent the marginal impact of a unit of cargo quantity on transportation costs, and its value is typically in the range of [-10^3, 10^3]. Furthermore, the update frequency of gradient information is related to the pre-computed sampling density, usually calculated offline at a frequency of 10^2 to 10^3 times per second to ensure coverage of gradient changes in typical scenarios.

[0097] At the application level, this step is widely applicable to complex optimization scenarios requiring real-time decision-making, such as power system dispatching, communication resource allocation, and supply chain route planning. In these scenarios, the system needs to quickly generate optimal decisions that satisfy constraints based on conditional variables such as current load, channel status, or order demand. By using gradient information as an enhancing feature, the model can more accurately capture the nonlinear mapping relationship between input and output, thereby achieving efficient and high-quality decision generation in the online phase.

[0098] The technical advantage of this step lies in its significant improvement of the generative model's ability to model the distribution of optimal solutions by introducing gradient features. Gradient information, as a sensitivity indicator of decision variables, helps the model better align with the local structure of the optimal solution during generation, thereby reducing solution bias and improving convergence speed. Furthermore, this enhanced feature vector provides theoretical optimality guidance for the model, ensuring that the generated results, while satisfying constraints, are as close as possible to the global optimum, providing solid technical support for intelligent optimization in high real-time and high-reliability scenarios.

[0099] Step S6, enhance the feature vector Compared with the original scene parameters The data is concatenated and used as input conditions for the generative model. This is to improve the model's fitting accuracy to the optimal solution distribution.

[0100] Specifically, enhance the feature vector Compared with the original scene parameters The data is concatenated and used as input conditions for the generative model. This step is one of the key steps in model training and online deployment in this invention. It aims to improve the fitting accuracy of the generative model to the optimal solution distribution by fusing original scene information with enhanced features, thereby achieving higher-quality decision generation in the online phase.

[0101] From a technical implementation perspective, the original scene parameters These typically include key input variables from the problem instance, such as the quantity of goods in logistics scheduling, the channel gain matrix in a communication system, and the node load in power dispatching. These parameters directly determine the constraint boundary of the optimization problem and the input space of the objective function. (Enhanced feature vectors) This is achieved through feature extraction modules (such as convolutional neural networks, graph neural networks, or attention mechanisms) The high-dimensional feature representation obtained after nonlinear transformation aims to capture the structured information or latent patterns hidden in the original parameters. (This refers to the splicing operation.) Typically, a vector-level concatenation method is used, that is... and The features are concatenated along the feature dimension to form a unified input condition vector. Its dimensions are ,in For the original parameter dimensions, To enhance feature dimensions.

[0102] At the parameter level, the concatenated input conditions The input specifications of the generative model must be met. For example, if the DDPM model is used, its input is usually a vector or tensor of fixed dimensions, therefore... The dimensions need to match the structure of the model's input layer. Furthermore, to ensure the stability of model training, the dimensions need to be adjusted before splicing. and Normalize the values ​​to fit within a uniform range (e.g., [0,1] or [-1,1]), and preprocess them using standardization methods (e.g., Z-score or Min-Max).

[0103] At the application level, this step is widely applicable to optimization problems with complex constraints and high-dimensional inputs. For example, in power system dispatching, It can represent the real-time load of each node. This can represent the power grid topology features extracted through a graph neural network, and the resulting concatenated topology features. As input to the diffusion model, it is used to generate generator output allocation schemes. In communication resource allocation, It can be Channel State Information (CSI). User behavior features extracted through the attention mechanism can be concatenated and used to generate resource block allocation strategies.

[0104] The technical effect of this step is that it introduces enhanced features. The model can more comprehensively understand the underlying structure of scene parameters, thus more accurately approximating the distribution of the theoretical optimal solution during the generation process. Experiments show that this splicing method can reduce the mean squared error (MSE) of the generated model by more than 15%, significantly improving the quality and stability of the solution. Meanwhile, because... The introduction of this feature enhances the model's ability to perceive constraints, significantly reducing the probability of constraint violations in the generated solution and further improving the system's security and reliability.

[0105] To achieve the above embodiments, the present invention also proposes a generative intelligent optimization device. Figure 3 This is a schematic diagram of a generative intelligent optimization device provided in an embodiment of the present invention. Figure 3 As shown, the device includes: The modeling and pre-computation module 100 is used to establish a mathematical model containing decision variables, objective function and constraints for the target optimization problem, and to pre-compute the scene parameters based on traditional optimization algorithms to obtain the theoretical optimal solution. Dataset building module 200 is used to build a training dataset containing scene parameters and theoretical optimal solutions; The model training module 300 is used to train the training dataset using a conditional generative artificial intelligence model. By minimizing the difference between the generated decision and the theoretical optimal solution, a mapping relationship between scene parameters and the optimal decision is established, so that the model implicitly learns how to satisfy the constraints. The real-time decision generation module 400 is used to collect parameters of the current scene in real time and input them into the trained generative model, and generate a near-optimal decision scheme through reverse denoising process or latent variable decoding.

[0106] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0107] To implement the above embodiments, the present invention also proposes an electronic device, comprising: a processor, and a memory communicatively connected to the processor; the memory stores computer execution instructions; the processor executes the computer execution instructions stored in the memory to implement the method provided in the foregoing embodiments.

[0108] To implement the above embodiments, the present invention also proposes a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the methods provided in the foregoing embodiments.

[0109] To implement the above embodiments, the present invention also proposes a computer program product, including a computer program that, when executed by a processor, implements the methods provided in the foregoing embodiments.

[0110] The collection, storage, use, processing, transmission, provision, and disclosure of user personal information involved in this invention all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0111] It should be noted that personal information collected from users should be used for legitimate and reasonable purposes and should not be shared or sold outside of these legitimate uses. Furthermore, such collection / sharing should only be conducted after receiving the user's informed consent, including but not limited to notifying the user to read the user agreement / user notice and sign an agreement / authorization that includes authorization of relevant user information before the user uses the function. In addition, any necessary steps must be taken to protect and safeguard access to such personal information data and ensure that others with access to personal information data comply with their privacy policies and procedures.

[0112] This invention is intended to provide implementation schemes for users to selectively prevent the use or access to personal information data. That is, this disclosure is intended to provide hardware and / or software to prevent or block access to such personal information data. Once personal information data is no longer needed, risks can be minimized by restricting data collection and deleting data. Furthermore, where applicable, such personal information can be de-identified to protect user privacy.

[0113] In the foregoing descriptions of the embodiments, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0114] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0115] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of the invention pertain.

[0116] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0117] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented using any of the following techniques known in the art, or a combination thereof: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0118] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0119] Furthermore, the functional units in the various embodiments of the present invention can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0120] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

[0121] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0122] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A generative intelligent optimization method, characterized in that, Includes the following steps: S1. A mathematical model containing decision variables, objective function and constraints is established for the target optimization problem, and the theoretical optimal solution is obtained by pre-calculating the scenario parameters based on traditional optimization algorithms. S2, Construct a training dataset containing scene parameters and theoretical optimal solutions; S3 uses a conditional generative artificial intelligence model to train the training dataset. By minimizing the difference between the generated decision and the theoretical optimal solution, a mapping relationship between scene parameters and the optimal decision is established, so that the model implicitly learns how to satisfy the constraints. S4 collects parameters of the current scene in real time and inputs them into the trained generative model. Through the reverse denoising process or latent variable decoding, it generates a near-optimal decision scheme, which can be directly applied to real-time decision-making scenarios.

2. The method according to claim 1, characterized in that, A mathematical model is established for the objective optimization problem, including decision variables, objective function, and constraints. Based on traditional optimization algorithms, the scenario parameters are pre-calculated to obtain the theoretical optimal solution. This also includes: S11, use KKT conditions to reduce variables in the objective optimization problem, and use the Lagrange multiplier method to transform the constraints into gradient conditions for the optimization variables; S12 uses the dual decomposition method to break down a large-scale optimization problem into multiple subproblems, and constructs the theoretical optimal solution of the overall problem by coordinating the optimal solutions of the subproblems.

3. The method according to claim 1, characterized in that, Constructing a training dataset that includes scene parameters and theoretical optimal solutions also includes: S21, the specific scenario parameters include cargo volume, transportation distance and time window constraints in logistics scenarios, channel gain, interference intensity and QoS requirements in communication scenarios, and node load, line capacity and power generation cost function in power grid scenarios; S22, the theoretical optimal solution specifically includes the shortest distance sequence of transportation paths, the optimal allocation matrix of resource block lengths, and the Pareto optimal solution set of generator output.

4. The method according to claim 1, characterized in that, A conditional generative artificial intelligence model is trained on the training dataset. By minimizing the difference between the generated decisions and the theoretical optimal solution, a mapping relationship between scene parameters and the optimal decisions is established. This allows the model to implicitly learn how to satisfy constraints, and also includes: S31, the conditional generative artificial intelligence model adopts the UNet architecture combined with a multi-head attention mechanism to enhance the ability to model structured condition variables through cross-layer feature interaction; S32, minimizing the difference between the generated decision and the theoretical optimal solution is calculated using the mean squared error loss function, as shown in the following formula: ， in To generate decision variables, This is to pre-calculate the optimal solution.

5. The method according to claim 1, characterized in that, The system collects parameters of the current scene in real time and inputs them into a trained generative model. Through a reverse denoising process or latent variable decoding, it generates a near-optimal decision-making scheme, which is then directly applied to real-time decision-making scenarios. Other features include: S41, The reverse denoising process iteratively optimizes the initial noise vector. At every step Calculate the conditional gradient And update status ,in The learning rate; S42, latent variable decoding employs a variational inference framework, by maximizing the lower bound of evidence. Achieve the generation of near-optimal solutions.

6. The method according to claim 1, characterized in that, Also includes: S5. Feature extraction is performed on the pre-calculated theoretical optimal solution to generate an enhanced feature vector containing gradient information of decision variables. ; S6 will enhance the feature vector Compared with the original scene parameters The data is concatenated and used as input conditions for the generative model. This is to improve the model's fitting accuracy to the optimal solution distribution.

7. A generative intelligent optimization device, characterized in that, include: The modeling and pre-computation module is used to build a mathematical model containing decision variables, objective functions and constraints for the target optimization problem, and to pre-compute the scene parameters based on traditional optimization algorithms to obtain the theoretical optimal solution. The dataset building module is used to build a training dataset containing scene parameters and theoretical optimal solutions; The model training module is used to train the training dataset using a conditional generative artificial intelligence model. By minimizing the difference between the generated decision and the theoretical optimal solution, a mapping relationship between scene parameters and the optimal decision is established, so that the model implicitly learns how to satisfy the constraints. The real-time decision generation module is used to collect parameters of the current scene in real time and input them into the trained generative model. It generates a near-optimal decision scheme through reverse denoising or latent variable decoding.

8. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-6.

10. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method of any one of claims 1-6.

Images