An inventory management method and system based on a robust optimization model

Abstract

This invention provides an inventory management method and system based on a robust optimization model, relating to the field of data management technology. The method includes: identifying potential risk transmission paths in a heterogeneous supply chain network diagram based on a two-layer message passing mechanism, and outputting risk-aware demand; inferring virtual extreme scenario samples based on a neural symbolic model to analyze risk transmission paths; using the influencing factors of demand and supply data as graph nodes, and combining the virtual extreme scenario samples with the shelf life and historical loss data of perishable goods for in-depth joint prediction of supply and demand forecasts; calculating the supply and demand matching degree based on the supply and demand forecast results and risk-aware demand, and dynamically optimizing the safety stock level and order lead time through a robust optimization algorithm to generate an inventory operation strategy; inputting the inventory operation strategy into a lightweight inventory optimization model for real-time inference to generate an executable inventory management solution. This invention solves the problem of failing to meet the requirements for refined, dynamic, and highly robust inventory management in heterogeneous supply chains.

Description

Technical Field

[0001] This invention relates to the field of data management technology, and more specifically, to an inventory management method and system based on a robust optimization model. Background Technology

[0002] In existing data management technologies, traditional inventory management methods mostly rely on static parameter settings, which cannot effectively identify potential risk transmission paths in the supply chain network and lack the ability to perceive risks such as demand fluctuations and supply disruptions. Meanwhile, current mainstream deep learning inventory forecasting models rely heavily on historical training data to meet the independent and identically distributed assumption. When faced with off-distribution scenarios such as supply chain disruptions, sudden public events, and geopolitical conflicts, they are unable to support emergency inventory scheduling decisions. At the same time, existing robust optimization methods are unable to adapt to dynamically changing off-distribution scenarios and emergency scheduling rules, and cannot correct prediction deviations and generate reliable scheduling strategies in real time under off-distribution scenarios. Ultimately, this leads to the inability to meet the requirements of refined, dynamic, and highly robust inventory management in heterogeneous supply chains.

[0003] Therefore, there is an urgent need for an inventory management method and system based on a robust optimization model, which solves the problem of not being able to meet the needs of refined, dynamic, and highly robust inventory management in heterogeneous supply chains. Summary of the Invention

[0004] The purpose of this invention is to provide an inventory management method and system based on a robust optimization model to improve the aforementioned problems. To achieve this objective, the technical solution adopted by this invention is as follows:

[0005] Firstly, this application provides an inventory management method based on a robust optimization model, including:

[0006] Construct a heterogeneous supply chain network diagram using inventory demand data, supply data, and external environment data;

[0007] Based on the two-layer message passing mechanism in the spatiotemporal graph convolutional network, the multidimensional relationship characteristics of the heterogeneous supply chain network graph are identified to identify potential risk transmission paths and output risk perception requirements.

[0008] A neural symbolic model based on a hierarchical task sampling strategy performs virtual reasoning on preset historical data and risk transmission paths to generate virtual extreme scenario samples.

[0009] By using the influencing factors of the demand data and the supply data as graph nodes, and combining the virtual extreme scenario samples with the shelf life, storage conditions and historical loss data of perishable goods, a deep joint prediction is performed to obtain the supply and demand prediction results.

[0010] Based on the supply and demand forecast results and the risk perception requirements, the supply and demand matching degree is calculated, and the safety stock level and order lead time are dynamically optimized through a robust optimization algorithm to generate an inventory operation strategy.

[0011] The inventory operation strategy is input into a preset lightweight inventory optimization model for real-time reasoning to generate an executable inventory management solution.

[0012] Secondly, this application also provides an inventory management system based on a robust optimization model, including:

[0013] The building module is used to construct heterogeneous supply chain network diagrams based on inventory demand data, supply data, and external environment data;

[0014] The identification module is used to identify potential risk transmission paths based on the two-layer message passing mechanism in the spatiotemporal graph convolutional network, and output risk perception requirements.

[0015] The virtual reasoning module is used to perform virtual reasoning on preset historical data and risk transmission paths based on a hierarchical task sampling strategy neural symbolic model, and generate virtual extreme scenario samples.

[0016] The prediction module is used to take the influencing factors of the demand data and the supply data as graph nodes, and combine them with the virtual extreme scenario samples and the shelf life, storage conditions and historical loss data of perishable goods to perform deep joint prediction and obtain the supply and demand prediction results.

[0017] The calculation module is used to calculate the supply and demand matching degree based on the supply and demand forecast results and the risk perception requirements, and dynamically optimize the safety stock level and order lead time through a robust optimization algorithm to generate an inventory operation strategy.

[0018] The real-time inference module is used to input the inventory operation strategy into a preset lightweight inventory optimization model for real-time inference and generate an executable inventory management solution.

[0019] The beneficial effects of this invention are as follows:

[0020] This invention utilizes a two-layer message passing mechanism in a spatiotemporal graph convolutional network to automatically mine potential risk transmission paths from heterogeneous supply chain networks, achieving the output of risk-aware demand from data modeling. Based on a neural symbolic model, it performs virtual reasoning on historical data and risk paths, autonomously generating samples covering extreme scenarios, effectively solving the problems of small sample size and difficult labeling. Furthermore, it graphs the influencing factors of supply and demand, integrating virtual extreme scenarios, shelf life of perishable goods, storage conditions, and historical losses for deep joint prediction, reducing demand forecasting bias and inventory expiration losses, and improving the model's adaptability to dynamic environments and perishable characteristics. Driven by supply-demand matching, it dynamically optimizes safety stock levels and order lead times through robust optimization algorithms, generating inventory operation strategies that balance risk resistance and cost control. Finally, the strategy is input into a lightweight inventory optimization model for real-time reasoning, generating directly executable management solutions. In summary, this invention solves the problem of failing to meet the needs of refined, dynamic, and highly robust inventory management in heterogeneous supply chains.

[0021] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a schematic diagram of the inventory management method based on a robust optimization model as described in an embodiment of the present invention;

[0024] Figure 2 This is a schematic diagram of the inventory management equipment structure based on the robust optimization model described in an embodiment of the present invention.

[0025] The diagram is labeled as follows: 800, Inventory management device based on robust optimization model; 801, Processor; 802, Memory; 803, Multimedia component; 804, I / O interface; 805, Communication component. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0027] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0028] Specific application scenarios:

[0029] In the existing heterogeneous supply chain inventory management technology environment, where the supply chain network involves multi-dimensional dynamic data interaction and needs to cope with sudden extreme risks such as the shelf life of perishable goods, loss control, supplier interruptions, and logistics delays, existing technical solutions face multiple limitations: First, the assumption of independent and identically distributed historical data in deep learning models fails, making it unable to support emergency scheduling; second, multi-focused demand forecasting only optimizes a single dimension and fails to systematically integrate supply-side risks such as supplier fulfillment fluctuations and logistics delays, leading to a structural contradiction of accurate demand forecasting but delayed replenishment; finally, existing robust optimization methods cannot adapt to dynamically changing out-of-distribution scenarios and emergency rules, cannot correct deviations in real time and generate reliable strategies, ultimately leading to problems such as inventory backlog, stockout losses, wasted ordering costs, excessive loss of perishable goods, and failure of shelf-life control, thus failing to meet the needs of refined, dynamic, and highly robust inventory management.

[0030] Example 1:

[0031] This embodiment provides an inventory management method based on a robust optimization model.

[0032] See Figure 1 The figure shows that the method includes steps S1 to S6, including:

[0033] S1: Construct a heterogeneous supply chain network diagram based on inventory demand data, supply data, and external environment data;

[0034] In this step, data dimension entity features and relationship features are extracted from the demand data, supply data and external environment data. A node feature matrix with static attributes and multi-scale time-series features is constructed by extracting the features. Then, time-varying weights are set by combining the correlation strength and influence weight of nodes and relationships in actual business, and finally a heterogeneous supply chain network diagram is formed.

[0035] The demand data includes historical sales, promotional activities, seasonal factors, and macroeconomic indicators; the supply data includes supplier historical performance records, capacity utilization rates, logistics tracking data, and IoT sensor data from production facilities; and the external environment data includes geopolitical events, public health events, natural disaster warnings, and changes in trade policies.

[0036] S2: Based on the two-layer message passing mechanism in the spatiotemporal graph convolutional network, identify potential risk transmission paths for the multidimensional relationship features of the heterogeneous supply chain network graph and output risk perception requirements;

[0037] To clarify the specific methods for obtaining risk perception needs, step S2 includes S21 to S24, specifically:

[0038] S21: Based on the node and edge information interaction rules of the two-layer message passing mechanism, feature aggregation is performed on the relationship between supplier, production facility and warehouse nodes in the heterogeneous supply chain network diagram to extract supply and demand spatial coupling features.

[0039] In this step, based on the node and edge information interaction rules of the two-layer message passing mechanism, the spatial association links of supplier, production facility, and warehouse nodes are established in the heterogeneous supply chain network graph, including the supply relationship between suppliers and production facilities, the production relationship between production facilities and logistics centers, and the distribution relationship between logistics centers and warehouses. Then, the heterogeneous relationships between each node are captured by the relational graph attention network (R-GAT). The multi-head attention mechanism calculates the attention coefficients under different relationships and aggregates them with weights. At the same time, the node features are updated by combining the static attributes and temporal characteristics of the nodes themselves, and the spatial coupling characteristics of supply and demand that reflect the mutual influence between supply and demand are extracted for the spatial association characteristics of each node.

[0040] The relational graph attention network captures heterogeneous relationships between nodes in the following specific way:

[0041] ;

[0042] In the above formula, For nodes In the graph neural network The feature vector of the layer, It can be any kind of relationship (such as supply, production, distribution). It is a non-linear activation function. It is the set of all relationship types in the heterogeneous supply chain network diagram. For nodes In relationship The set of adjacent nodes below, For relationship Attention coefficient under, For relationship In the The learnable weight matrix of the layer, For nodes In the graph neural network The original feature vector of the layer, For the node's own characteristics in the th The learnable weight matrix of the layer, For nodes In the graph neural network The original feature vector of the layer;

[0043] Among them, relationship The attention coefficients are calculated using a multi-head attention mechanism.

[0044] S22: Perform time-domain evolution modeling on the aforementioned supply and demand spatial coupling characteristics to obtain the time evolution results of node states;

[0045] In this step, the supply and demand spatial coupling characteristics are used as the neighbor information aggregated by each node at the corresponding time point. The memory state of the node at the current time is updated by using a gated recurrent unit (GRU) and the memory state of the node at the previous time point, and the change law of node characteristics over time is tracked. In this way, the temporal evolution of the node state is modeled, and the temporal evolution result of the node state is output.

[0046] The expression for the time evolution result of the node state is:

[0047] ;

[0048] In the above formula, For nodes At the present moment memory state, For gated loop unit, For nodes In the previous moment memory state, For time node The neighbor information obtained by aggregation.

[0049] S23: The supplier's interruption probability, logistics delay distribution, and production failure rate are encoded as node risks and embedded into a graph neural network to simulate the cascading propagation of risks, thus constructing a supply-demand coupled risk propagation model.

[0050] In this step, supply-side risk indicators such as supplier disruption probability, logistics delay distribution, and production failure rate are numerically encoded to generate node risks with unified dimensions, which are then embedded into the node features of a graph neural network. A graph diffusion model incorporating an attention mechanism is built, and the risk propagation relationship between suppliers, production facilities, warehouses, and supply chain nodes is characterized by a normalized adjacency matrix. Through multi-step iterative calculations, the cascading diffusion and transmission process of risks among nodes is simulated, and the node and global risk propagation status after each iteration is quantified synchronously. Finally, a supply-demand coupled risk propagation model is constructed, which reflects the transmission law of supply-side risks to the entire supply chain.

[0051] S24: Based on the supply and demand coupling risk propagation model, identify potential risk transmission paths from the node state time evolution results, and couple them with a preset demand prediction model to obtain risk perception demand.

[0052] In this step, the node state time evolution results are input into the supply and demand coupling risk propagation model for iterative risk propagation calculation. The calculation results are used to uncover the core paths and corresponding impact magnitudes of risk transmission from the source node to other upstream and downstream nodes within the supply chain network. All potential risk transmission paths are identified, and a node comprehensive risk score is generated. The node comprehensive risk score and historical time-series data on the demand side are input into a preset demand forecasting model for explicit modeling of demand impact. Through the bidirectional coupling and linkage operation of the supply and demand coupling risk propagation model and the demand forecasting model, the risk perception demand that integrates risk factors of the entire supply chain is finally output.

[0053] S3: A neural symbolic model based on a hierarchical task sampling strategy performs virtual reasoning on preset historical data and risk transmission paths to generate virtual extreme scenario samples;

[0054] To clarify the specific method for obtaining virtual extreme scenario samples, step S3 includes S31 to S35, specifically:

[0055] S31: Based on the neural symbolic model, hierarchical sampling is performed on the demand distribution and supply network structure in historical data to obtain a hierarchical task sampling set;

[0056] In this step, the demand distribution and supply network structure in historical data are analyzed and sampled hierarchically based on the neural symbolic model. The inventory optimization problem is modeled as a family of Markov decision processes. A meta-training task set is constructed using normal inventory scenarios with different inventory units, regions, and seasons in the historical data. A meta-validation task set is constructed with hyperparameter tuning and early stop verification as objectives. At the same time, a meta-testing task set is constructed by simulating out-of-distribution scenarios of supply chain disruption and demand surge. Diverse samples are extracted from the task pools of the meta-training task set, the meta-validation task set, and the meta-testing task set through hierarchical sampling and integrated to cover normal and potential abnormal scenarios, forming a hierarchical task sampling set.

[0057] S32: The hierarchical task sampling set and the preset given task set are iteratively updated by gradient descent according to the model-independent meta-learning algorithm to obtain task-specific parameters;

[0058] In this step, the hierarchical task sampling set and a preset given task set are used as training inputs for inner loop training based on the Model Independent Meta-Learning (MAML) algorithm. A loss function is calculated for each specific task and used as the optimization direction. The initial parameters of the target model are updated once or in small steps using gradient descent to obtain task-specific parameters, enabling the target model's parameters to quickly adapt to the characteristics and patterns of the task. Subsequently, an outer loop optimization is performed based on the performance of all task-specific parameters on the task. The initial parameters of the target model are then globally updated again using gradient descent, giving them better cross-task generalization performance. Finally, the task-specific parameters are obtained, achieving the initial adaptation of the target model to different tasks.

[0059] The inner loop parameter update follows the formula:

[0060] ;

[0061] In the above formula, For the first The updated task-specific parameters for each task. These are the initial common parameters of the model. For the inner loop learning rate, For the gradient operator with respect to the initial parameters, For the first The loss function for each task As task weight, For models based on initial parameters, For the first Support set data for each task;

[0062] The external circulation parameters are updated according to the following formula:

[0063] ;

[0064] In the above formula, These are the initial common parameters of the model. For the outer loop learning rate, For the gradient operator with respect to the initial parameters, As task weight, For the first The loss function for each task For the first The task-specific parameters obtained after the task undergoes a partial update via the inner loop. The constructed target model, For the first The query set data for each task, where ∑ represents the sum of losses for all tasks.

[0065] S33: Based on the policy gradient and value function loss, the task-specific parameters are iteratively updated to obtain the evaluation and adaptation parameters;

[0066] In this step, a task-specific loss function is constructed based on policy gradient loss and value function loss. During execution, the task-specific parameters are input into the task-specific loss function to carry out multiple rounds of outer loop iterative optimization. After each round of iterative optimization, the updated parameters are subjected to targeted performance evaluation on the meta-test task set. The parameter values ​​are adjusted in real time based on the evaluation results, forming a closed-loop linkage between iterative optimization and performance verification. The parameter iteration and adjustment work is continuously promoted until the model's performance on the task meets the preset performance standard, and the evaluation adaptation parameters are obtained.

[0067] The policy gradient loss is responsible for anchoring the profit orientation of the model decision and optimizing the effectiveness of the model decision, while the value function loss is used to reduce the state value prediction error and improve the prediction accuracy.

[0068] S34: Based on the neural symbolic fusion decision-making mechanism and the aforementioned evaluation adaptation parameters, extract state features to generate supply chain inventory decisions;

[0069] To clarify the specific methods for obtaining supply chain inventory decisions, step S34 includes S341 to S345, specifically:

[0070] S341: Obtain supply chain status data;

[0071] S342: Based on the neural symbol fusion decision-making mechanism and the evaluation adaptation parameters, extract state features from the supply chain state data to obtain a high-dimensional state feature vector;

[0072] In this step, based on the neural perception layer of the neural symbol fusion decision-making mechanism, the evaluation adaptation parameters are input into the dynamic heterogeneous graph neural network model to extract and fuse features of inventory level, in-transit inventory, supply risk score and cost parameters in the supply chain status data. Implicit correlation features between data nodes are extracted through graph convolution operation, and the temporal evolution law of data is captured by combining time series modeling, and finally a high-dimensional state feature vector is output.

[0073] S343: Iteratively optimize the high-dimensional state feature vector according to inventory theory rules and business logic constraints to generate the original inventory action;

[0074] In this step, the standardized inventory theory rules and business logic constraints adapted to supply chain practices are embedded in the call layer of the symbolic reasoning layer of the neural symbolic fusion decision-making mechanism. The high-dimensional state feature vector is input into the rule system for logical reasoning and constraint verification. The symbolic rules are transformed into a soft constraint embedding loss function through a differentiable logic layer. The soft constraint embedding loss function performs multiple rounds of iterative optimization on the feature decision results corresponding to the high-dimensional state feature vector to generate the original inventory action.

[0075] S344: Project the original inventory actions onto the feasible domain under the inventory capacity constraint and the zero order quantity constraint for feasibility verification and correction, and obtain the effective inventory actions;

[0076] In this step, the original inventory actions are projected onto the feasible regions of inventory capacity constraints, minimum order quantity, and zero order quantity constraints for verification one by one. Actions that do not meet the hard constraints, such as order quantities exceeding the warehouse capacity limit or order values ​​below the minimum order quantity and not zero, are corrected in real time. End-to-end constraint adaptation is performed through a differentiable optimization layer to obtain effective inventory actions.

[0077] S345: Interact the effective inventory actions with the supply chain environment and extract interpretable decision rules to form supply chain inventory decisions.

[0078] In this step, the effective inventory actions are observed in relation to the actual supply chain environment, and data on the benefits and status changes in the environment are collected. At the same time, an inductive logic programming method is used to extract a rule set from the trained strategy network. The rule set is then checked for consistency with the inventory management expert knowledge base, and invalid rules are eliminated and reasonable rules are improved to form a decision rule system. Combined with the optimization and adjustment based on the feedback from the actual environment, the final supply chain inventory decision is formed.

[0079] S35: Based on the joint reasoning of the supply chain inventory decision and the risk transmission path, simulate the supply chain operation status under different extreme risks to generate virtual extreme scenario samples.

[0080] In this step, the supply chain inventory decision is input into a supply chain risk propagation simulator for multi-round joint reasoning. This simulates the occurrence of different extreme risks by triggering various extreme risk events, including supplier disruptions, logistics disruptions, and demand surges. The simulation dynamically extrapolates the propagation patterns and impact range of cascading failures along the potential risk transmission paths, as well as the actual buffering and adjustment effects of inventory strategies under risk disturbances. During the simulation, key operational data on the status changes, response times, and recovery capabilities of supply chain nodes are recorded in real time and subjected to cluster analysis and feature extraction to generate virtual extreme scenario samples covering different types of extreme risks and varying degrees of risk impact.

[0081] S4: Using the influencing factors of the demand data and the supply data as graph nodes, and combining the virtual extreme scenario samples with the shelf life, storage conditions and historical loss data of perishable goods, a deep joint prediction is performed to obtain the supply and demand prediction results.

[0082] To clarify the specific methods for obtaining supply and demand forecast results, step S4 includes S41 to S44, specifically:

[0083] S41: Using the influencing factors of the demand data and the supply data as graph nodes, the mutual influence relationship between the demand side and the supply side is quantified through the business logic of inventory management to obtain a supply and demand characteristic graph.

[0084] In this step, the influencing factors of the demand data (market demand, consumer preferences, seasonal factors, etc.) and the influencing factors of the supply data (capacity, supply chain efficiency, raw material reserves, etc.) are used as graph nodes. Based on the supply and demand balance relationship in the inventory management business logic, the mutual influence relationship between the demand side and the supply side is quantified according to the inventory turnover law, and a dual-end characteristic graph of supply and demand is constructed.

[0085] S42: Based on the virtual extreme scenario samples and the shelf life, storage conditions and historical loss data of perishable goods, feature alignment is performed, and the fused feature map is obtained by fusing the feature-aligned data with the supply and demand feature map.

[0086] In this step, feature alignment is performed based on extreme weather and sudden market fluctuations in the virtual extreme scenario sample, as well as the shelf life, storage conditions, and historical loss data of perishable goods, thereby eliminating the differences between different data dimensions. Then, the aligned data is fused with the supply and demand feature map to obtain the fused feature map.

[0087] S43: Input the fused feature map into the hybrid prediction model constructed by graph neural network and gated recurrent unit to perform demand prediction and supply capacity prediction, and obtain the demand prediction value and the supply-side indicator prediction value.

[0088] In this step, the fused feature map is input into a hybrid prediction model constructed by a graph neural network (GNN) and a gated recurrent unit (GRU) for demand and supply capacity prediction. The graph neural network is used to mine the complex correlation features between nodes in the map, and the gated recurrent unit is responsible for capturing time series features. After feature extraction and analysis, the predicted demand value and the predicted supply index value are output.

[0089] S44: Based on the demand forecast and the supply-side indicator forecast, perform in-depth joint deduction and correction to obtain the supply and demand forecast results.

[0090] In this step, a deep joint extrapolation is performed based on the demand forecast and the supply-side indicator forecast. Targeted error corrections are made to the initial forecast deviations generated during the extrapolation process, taking into account the shelf-life constraints of perishable goods, storage limitations, and historical loss patterns. By synchronizing the special commodity attributes of perishable goods with actual warehousing operation scenarios, forecast deviations under extreme supply and demand scenarios are effectively avoided. The final supply and demand forecast results address the technical shortcomings of traditional forecasting methods, such as ignoring the inherent correlation between supply and demand, poor adaptability to extreme scenarios, and failure to consider the impact of the specific characteristics of perishable goods. This achieves a high degree of matching between the forecast results and actual business scenarios.

[0091] S5: Calculate the supply and demand matching degree based on the supply and demand forecast results and the risk perception demand, and dynamically optimize the safety stock level and order lead time through a robust optimization algorithm to generate an inventory operation strategy;

[0092] To clarify the specific methods for obtaining inventory operation strategies, step S5 includes S51 to S54, specifically:

[0093] S51: Based on the supply and demand forecast results and the risk perception demand, the current supply and demand matching degree is calculated in real time.

[0094] In this step, the supply and demand forecast results are compared with the risk perception demand in real time, and the current supply and demand matching is quantified to obtain the current matching degree between supply and demand. At the same time, the impact of risk perception demand on the overall supply and demand balance is quantified to obtain the current supply and demand matching degree.

[0095] S52: Taking the holding cost, stockout cost, and ordering cost of inventory as optimization objectives, the robust optimization algorithm is improved by using uncertainty set theory and risk preference coefficient to obtain an adaptive robust optimization model;

[0096] In this step, the total inventory cost function is constructed with the inventory holding cost, stockout cost, and ordering cost as optimization objectives. The total inventory cost function is solved using a robust optimization algorithm. Then, the range of supply and demand fluctuations is quantitatively defined using uncertainty set theory to obtain the uncertainty boundary between supply and demand. A risk preference coefficient is then used to characterize the decision-maker's risk attitude; a larger risk preference coefficient indicates a higher degree of risk aversion. By integrating uncertainty set theory with the risk preference coefficient, the traditional robust optimization algorithm is improved to obtain an adaptive robust optimization model. This adaptive robust optimization model is used to adaptively match different risk preferences with different supply and demand fluctuation scenarios.

[0097] S53: Input the current supply and demand matching degree, the preset safety stock level and the order lead time into the adaptive robust optimization model for iterative solution to obtain the robust optimization solution;

[0098] In this step, the current supply and demand matching degree, the preset safety stock level, and the order lead time are input into the adaptive robust optimization model. Under the premise of supply and demand fluctuation range and risk preference constraints, the robust optimization solution is solved through iterative calculation. The robust optimization solution is used to ensure that the optimization result still has robustness and optimality within the range of fluctuations in demand and supply, thereby ensuring the stability and rationality of inventory decisions under uncertain scenarios.

[0099] S54: Based on the robust optimization solution, generate strategies for the preset target safety stock level, order replenishment batch and replenishment frequency to obtain the inventory operation strategy.

[0100] In this step, based on the robust optimization solution and the current supply and demand matching status and risk preference information, the preset target safety stock level, order replenishment batch and replenishment frequency are adaptively adjusted and strategies are generated to obtain an inventory operation strategy for the current supply and demand status and risk preference. The inventory operation strategy realizes precise control of inventory operation costs, improves the robustness and adaptability of inventory management, and enables the inventory operation strategy to dynamically adapt to supply and demand fluctuations and risk changes. This solves the technical problems of traditional inventory strategies, such as incomplete cost optimization, poor adaptability to supply and demand fluctuations and risks, and inventory backlog, excessive stockout losses or waste of ordering costs caused by lag in strategy adjustment.

[0101] S6: Input the inventory operation strategy into the preset lightweight inventory optimization model for real-time reasoning to generate an executable inventory management solution.

[0102] To clarify the specific method for obtaining an executable inventory management plan, step S6 includes S61 to S64, specifically:

[0103] S61: Input the inventory operation strategy into the preset lightweight inventory optimization model, and perform parameter iterative optimization by combining the anti-interference characteristics of the robust optimization algorithm to obtain preliminary optimization results;

[0104] In this step, the inventory operation strategy (optimal safety stock level, optimal order lead time, and replenishment batch size) is input into a preset lightweight inventory optimization model. The model has the characteristics of being resistant to interference and adapting to supply and demand fluctuations. With the dual objectives of minimizing model inference error and optimizing the feasibility of implementing the inventory strategy, iterative optimization calculations are performed on the internal core parameters. Through multiple rounds of iteration and convergence, preliminary optimization results that fit the actual inventory operation scenario are obtained.

[0105] S62: Compare and judge the preliminary optimization result with the preset parameter constraints. If the preliminary optimization result does not meet the parameter constraints, the parameters are reallocated by weight adjustment until the constraints are met and the inference result is output.

[0106] In this step, the preliminary optimization results are compared and verified one by one with the preset parameter constraints. When the verification determines that the preliminary optimization results do not meet any or all of the constraints, the allocation weights of each inventory parameter are adjusted based on the current supply and demand matching status and risk preference. The parameter optimization allocation and model calculation are re-executed, and the constraint verification is carried out again. The above weight adjustment, parameter redistribution and constraint verification process is repeated until the preliminary optimization results fully meet all preset parameter constraints. Finally, the model inference results adapted to the actual inventory operation scenario are output.

[0107] The parameter constraints include inventory capacity constraints, replenishment batch constraints, and operating cost constraints. The inventory capacity constraint corresponds to the maximum upper limit of actual inventory capacity. The replenishment batch constraint is used to limit the minimum and maximum reasonable range of a single replenishment. The cost constraint is the maximum total cost threshold allowed throughout the entire inventory operation process.

[0108] S63: Based on the reasoning results, the initial inventory plan is generated by matching with the preset replenishment plan, inventory allocation plan and inventory clearance plan;

[0109] In this step, the reasoning results are matched and selected from the preset replenishment plan, inventory allocation plan library and inventory clearance plan library. At the same time, the shelf life attribute of perishable goods, actual warehousing conditions and storage requirements are comprehensively adapted to generate an initial inventory plan. The initial inventory plan includes order replenishment time, allocation path and clearance priority.

[0110] S64: Perform a feasibility check on the execution cost and operational difficulty of the initial inventory plan, and generate a final executable inventory management plan when the preset target guidance parameters are met.

[0111] In this step, the feasibility of the initial inventory plan is verified in terms of execution cost and operational difficulty. This is achieved by quantifying the execution cost and operational difficulty coefficient. The preset target-oriented parameters are further subdivided into target execution cost thresholds and target operational difficulty coefficient thresholds, corresponding to the maximum acceptable execution cost and maximum feasible operational difficulty for perishable goods inventory operations, respectively. During the verification process, adaptability checks are simultaneously performed based on perishable goods shelf-life control and actual warehousing conditions. If the execution cost of the initial inventory plan does not exceed the target execution cost threshold and the operational difficulty coefficient does not exceed the target operational difficulty coefficient threshold, and both dimensions meet the preset target requirements, a final inventory management plan adapted to the actual warehousing operation scenario is generated. If either dimension or both dimensions fail to meet the requirements, the core parameters of the initial inventory plan are optimized based on the verification results, and the verification is repeated until the requirements are met. This addresses the problems of traditional inventory strategies being disconnected from actual warehousing execution, poor plan feasibility, low execution efficiency, and plans failing to implement due to a lack of adaptation to the special attributes of perishable goods, or exceeding cost limits or becoming cumbersome after implementation.

[0112] Example 2:

[0113] This embodiment provides an inventory management system based on a robust optimization model, the system comprising:

[0114] The building module is used to construct heterogeneous supply chain network diagrams based on inventory demand data, supply data, and external environment data;

[0115] The identification module is used to identify potential risk transmission paths based on the two-layer message passing mechanism in the spatiotemporal graph convolutional network, and output risk perception requirements.

[0116] To clarify the specific methods for obtaining the identification module, the following are included:

[0117] The aggregation unit is used to perform feature aggregation on the relationship between suppliers, production facilities and warehouse nodes in the heterogeneous supply chain network diagram according to the node and edge information interaction rules of the two-layer message passing mechanism, and extract the spatial coupling features of supply and demand.

[0118] The modeling unit is used to perform time-domain evolution modeling on the supply and demand spatial coupling characteristics to obtain the time evolution results of node states;

[0119] The simulation unit is used to encode the supplier's interruption probability, logistics delay distribution, and production failure rate as node risks and embed them into the graph neural network to simulate the cascading propagation of risks and build a supply-demand coupled risk propagation model.

[0120] The identification unit is used to identify potential risk transmission paths based on the time evolution results of the node state according to the supply and demand coupling risk propagation model, and to couple them with a preset demand prediction model to obtain risk perception demand.

[0121] The virtual reasoning module is used to perform virtual reasoning on preset historical data and risk transmission paths based on a hierarchical task sampling strategy neural symbolic model, and generate virtual extreme scenario samples.

[0122] To clarify the specific methods for obtaining the virtual inference module, the following are included:

[0123] The sampling unit is used to perform hierarchical sampling of the demand distribution and supply network structure in historical data based on the neural symbol model, to obtain a hierarchical task sampling set;

[0124] The update unit is used to perform gradient descent iterative updates on the hierarchical task sampling set and the preset given task set according to the model-independent meta-learning algorithm to obtain task-specific parameters;

[0125] The iterative update unit is used to iteratively update the task-specific parameters based on the policy gradient and value function loss to obtain the evaluation and adaptation parameters.

[0126] The extraction unit is used to extract state features based on the neural symbolic fusion decision-making mechanism and the evaluation adaptation parameters to generate supply chain inventory decisions;

[0127] The reasoning unit is used to generate virtual extreme scenario samples by jointly reasoning based on the supply chain inventory decision and the risk transmission path to simulate the supply chain operation status under different extreme risks.

[0128] The prediction module is used to take the influencing factors of the demand data and the supply data as graph nodes, and combine them with the virtual extreme scenario samples and the shelf life, storage conditions and historical loss data of perishable goods to perform deep joint prediction and obtain the supply and demand prediction results.

[0129] The calculation module is used to calculate the supply and demand matching degree based on the supply and demand forecast results and the risk perception requirements, and dynamically optimize the safety stock level and order lead time through a robust optimization algorithm to generate an inventory operation strategy.

[0130] To clarify the specific methods for obtaining the calculation module, the following are included:

[0131] The comparison unit is used to compare the supply and demand forecast results with the risk perception requirements in real time and calculate the current supply and demand matching degree.

[0132] An improvement unit is used to improve the robust optimization algorithm by using uncertainty set theory and risk preference coefficients, with inventory holding costs, stockout costs and ordering costs as optimization objectives, to obtain an adaptive robust optimization model.

[0133] The iterative unit is used to input the current supply and demand matching degree, the preset safety stock level and the order lead time into the adaptive robust optimization model for iterative solution to obtain the robust optimization solution;

[0134] The strategy unit is used to generate a strategy based on the robust optimization solution for the preset target safety stock level, order replenishment batch and replenishment frequency, so as to obtain the inventory operation strategy.

[0135] The real-time inference module is used to input the inventory operation strategy into a preset lightweight inventory optimization model for real-time inference and generate an executable inventory management solution.

[0136] It should be noted that the specific methods by which each module performs operations in the system described in the above embodiments have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0137] Example 3:

[0138] Corresponding to the above method embodiments, this embodiment also provides an inventory management device based on a robust optimization model. The inventory management device based on a robust optimization model described below and the inventory management method based on a robust optimization model described above can be referred to in correspondence.

[0139] Figure 2 This is a block diagram illustrating an inventory management device 800 based on a robust optimization model, according to an exemplary embodiment. Figure 2 As shown, the inventory management device 800 based on a robust optimization model may include: a processor 801 and a memory 802. The inventory management device 800 may also include one or more of a multimedia component 803, an I / O interface 804, and a communication component 805.

[0140] The processor 801 controls the overall operation of the robust optimization model-based inventory management device 800 to complete all or part of the steps in the robust optimization model-based inventory management method described above. The memory 802 stores various types of data to support the operation of the robust optimization model-based inventory management device 800. This data may include, for example, instructions for any application or method operating on the robust optimization model-based inventory management device 800, as well as application-related data such as contact data, sent and received messages, images, audio, video, etc. The memory 802 can be implemented using any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The multimedia component 803 may include a screen and an audio component. The screen may be, for example, a touchscreen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals. The received audio signals may be further stored in the memory 802 or transmitted via the communication component 805. The audio component also includes at least one speaker for outputting audio signals. I / O interface 804 provides an interface between processor 801 and other interface modules, such as keyboards, mice, and buttons. These buttons can be virtual or physical. Communication component 805 is used for wired or wireless communication between the robust optimization model-based inventory management device 800 and other devices. Wireless communication includes Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, or 4G, or a combination thereof. Therefore, the corresponding communication component 805 may include a Wi-Fi module, a Bluetooth module, and an NFC module.

[0141] In an exemplary embodiment, the inventory management device 800 based on the robust optimization model may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to execute the aforementioned inventory management method based on the robust optimization model.

[0142] Example 4:

[0143] Corresponding to the above method embodiments, this embodiment also provides a medium. The medium described below can be referred to in conjunction with the inventory management method based on a robust optimization model described above.

[0144] A medium storing a computer program, which, when executed by a processor, implements the steps of the inventory management method based on a robust optimization model as described in the above method embodiments.

[0145] The medium can specifically be any medium capable of storing program code, such as a USB flash drive, external hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0146] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0147] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An inventory management method based on a robust optimization model, characterized in that, include: Construct a heterogeneous supply chain network diagram using inventory demand data, supply data, and external environment data; Based on the two-layer message passing mechanism in the spatiotemporal graph convolutional network, the multidimensional relationship characteristics of the heterogeneous supply chain network graph are identified to identify potential risk transmission paths and output risk perception requirements. A neural symbolic model based on a hierarchical task sampling strategy performs virtual reasoning on preset historical data and risk transmission paths to generate virtual extreme scenario samples. The factors influencing the demand data and supply data are used as graph nodes. The virtual extreme scenario samples are combined with the shelf life, storage conditions and historical loss data of perishable goods to perform deep joint prediction and obtain the supply and demand prediction results. Based on the supply and demand forecast results and the risk perception requirements, the supply and demand matching degree is calculated, and the safety stock level and order lead time are dynamically optimized through a robust optimization algorithm to generate an inventory operation strategy. The inventory operation strategy is input into a preset lightweight inventory optimization model for real-time reasoning to generate an executable inventory management solution.

2. The inventory management method based on a robust optimization model according to claim 1, characterized in that, Based on the supply and demand forecast results and the perceived risk demand, the supply and demand matching degree is calculated. A robust optimization algorithm is used to dynamically optimize the safety stock level and order lead time, generating an inventory operation strategy, including: The current supply and demand matching degree is calculated by comparing the supply and demand forecast results with the risk perception demand in real time. With inventory holding costs, stockout costs, and ordering costs as optimization objectives, the robust optimization algorithm is improved using uncertainty set theory and risk preference coefficients to obtain an adaptive robust optimization model. The current supply and demand matching degree, the preset safety stock level, and the order lead time are input into the adaptive robust optimization model for iterative solution to obtain the robust optimization solution; Based on the robust optimization solution, strategies are generated for the preset target safety stock level, order replenishment batch size, and replenishment frequency to obtain the inventory operation strategy.

3. The inventory management method based on a robust optimization model according to claim 1, characterized in that, Based on the two-layer message passing mechanism in the spatiotemporal graph convolutional network, the multidimensional relationship characteristics of the heterogeneous supply chain network graph are identified to determine potential risk transmission paths, and risk perception requirements are output, including: Based on the node and edge information interaction rules of the two-layer message passing mechanism, feature aggregation is performed on the relationship between supplier, production facility and warehouse nodes in the heterogeneous supply chain network diagram to extract supply and demand spatial coupling features. The spatial coupling characteristics of supply and demand are modeled in the time domain to obtain the temporal evolution results of node states; Supplier disruption probability, logistics delay distribution, and production failure rate are encoded as node risks and embedded into a graph neural network to simulate risk cascading propagation, thus constructing a supply-demand coupled risk propagation model. Based on the supply and demand coupling risk propagation model, potential risk transmission paths are identified by analyzing the temporal evolution results of the node state. These paths are then coupled with a preset demand prediction model to obtain risk-perceived demand.

4. The inventory management method based on a robust optimization model according to claim 1, characterized in that, A neural symbolic model based on a hierarchical task sampling strategy performs virtual reasoning on preset historical data and risk transmission paths to generate virtual extreme scenario samples, including: Based on the aforementioned neural symbolic model, hierarchical sampling is performed on the demand distribution and supply network structure in historical data to obtain a hierarchical task sampling set; The hierarchical task sampling set and the preset given task set are iteratively updated using gradient descent based on the model-independent meta-learning algorithm to obtain task-specific parameters. The task-specific parameters are iteratively updated based on the policy gradient and value function loss to obtain the evaluation and adaptation parameters. Based on the neural symbolic fusion decision-making mechanism and the aforementioned evaluation adaptation parameters, state features are extracted to generate supply chain inventory decisions. Based on the joint reasoning of the supply chain inventory decision and the risk transmission path, virtual extreme scenario samples are generated by simulating the supply chain operation status under different extreme risks.

5. The inventory management method based on a robust optimization model according to claim 4, characterized in that, Based on the neural symbolic fusion decision-making mechanism and the state features extracted from the evaluation adaptation parameters, a supply chain inventory decision is generated, including: Obtain supply chain status data; Based on the neural symbol fusion decision-making mechanism and the evaluation adaptation parameters, state features are extracted from the supply chain state data to obtain a high-dimensional state feature vector. The high-dimensional state feature vector is iteratively optimized based on inventory theory rules and business logic constraints to generate the original inventory actions. The original inventory actions are projected onto the feasible domain under inventory capacity constraints and zero order quantity constraints for feasibility verification and correction, resulting in effective inventory actions. The effective inventory actions are interacted with the supply chain environment, and interpretable decision rules are extracted to form supply chain inventory decisions.

6. The inventory management method based on a robust optimization model according to claim 1, characterized in that, The inventory operation strategy is input into a preset lightweight inventory optimization model for real-time inference to generate an executable inventory management solution, including: The inventory operation strategy is input into a preset lightweight inventory optimization model, and the parameters are iteratively optimized by combining the anti-interference characteristics of the robust optimization algorithm to obtain preliminary optimization results. The preliminary optimization results are compared with the preset parameter constraints. If the preliminary optimization results do not meet the parameter constraints, the parameters are reallocated through weight adjustment until the constraints are met and the inference results are output. Based on the reasoning results, the initial inventory plan is generated by matching it with the preset replenishment plan, inventory allocation plan and inventory clearance plan. The feasibility of the initial inventory plan is verified in terms of execution cost and operational difficulty. When the preset target guidance parameters are met, the final executable inventory management plan is generated.

7. An inventory management system based on a robust optimization model, characterized in that, include: The building module is used to construct heterogeneous supply chain network diagrams based on inventory demand data, supply data, and external environment data; The identification module is used to identify potential risk transmission paths based on the two-layer message passing mechanism in the spatiotemporal graph convolutional network, and output risk perception requirements. The virtual reasoning module is used to perform virtual reasoning on preset historical data and risk transmission paths based on a hierarchical task sampling strategy neural symbolic model, and generate virtual extreme scenario samples. The prediction module is used to take the influencing factors of the demand data and the supply data as graph nodes, and combine them with the virtual extreme scenario samples and the shelf life, storage conditions and historical loss data of perishable goods to perform deep joint prediction and obtain the supply and demand prediction results. The calculation module is used to calculate the supply and demand matching degree based on the supply and demand forecast results and the risk perception requirements, and dynamically optimize the safety stock level and order lead time through a robust optimization algorithm to generate an inventory operation strategy. The real-time inference module is used to input the inventory operation strategy into a preset lightweight inventory optimization model for real-time inference and generate an executable inventory management solution.

8. The inventory management system based on a robust optimization model according to claim 7, characterized in that, The computing module includes: The comparison unit is used to compare the supply and demand forecast results with the risk perception requirements in real time and calculate the current supply and demand matching degree. An improvement unit is used to improve the robust optimization algorithm by using uncertainty set theory and risk preference coefficients, with inventory holding costs, stockout costs and ordering costs as optimization objectives, to obtain an adaptive robust optimization model. The iterative unit is used to input the current supply and demand matching degree, the preset safety stock level and the order lead time into the adaptive robust optimization model for iterative solution to obtain the robust optimization solution; The strategy unit is used to generate a strategy based on the robust optimization solution for the preset target safety stock level, order replenishment batch and replenishment frequency, so as to obtain the inventory operation strategy.

9. The inventory management system based on a robust optimization model according to claim 7, characterized in that, The identification module includes: The aggregation unit is used to perform feature aggregation on the relationship between suppliers, production facilities and warehouse nodes in the heterogeneous supply chain network diagram according to the node and edge information interaction rules of the two-layer message passing mechanism, and extract the spatial coupling features of supply and demand. The modeling unit is used to perform time-domain evolution modeling on the supply and demand spatial coupling characteristics to obtain the time evolution results of node states; The simulation unit is used to encode the supplier's interruption probability, logistics delay distribution, and production failure rate as node risks and embed them into the graph neural network to simulate the cascading propagation of risks and build a supply-demand coupled risk propagation model. The identification unit is used to identify potential risk transmission paths based on the time evolution results of the node state according to the supply and demand coupling risk propagation model, and to couple them with a preset demand prediction model to obtain risk perception demand.

10. The inventory management system based on a robust optimization model according to claim 7, characterized in that, The virtual inference module includes: The sampling unit is used to perform hierarchical sampling of the demand distribution and supply network structure in historical data based on the neural symbol model, to obtain a hierarchical task sampling set; The update unit is used to perform gradient descent iterative updates on the hierarchical task sampling set and the preset given task set according to the model-independent meta-learning algorithm to obtain task-specific parameters; The iterative update unit is used to iteratively update the task-specific parameters based on the policy gradient and value function loss to obtain the evaluation and adaptation parameters. The extraction unit is used to extract state features based on the neural symbolic fusion decision-making mechanism and the evaluation adaptation parameters to generate supply chain inventory decisions; The reasoning unit is used to generate virtual extreme scenario samples by jointly reasoning based on the supply chain inventory decision and the risk transmission path to simulate the supply chain operation status under different extreme risks.

Images