Trade commodity dynamic inventory management and control system based on machine learning

Abstract

The invention discloses a trade commodity dynamic inventory management and control system based on machine learning, and the system comprises the following modules: a structure state module which is used for obtaining the inventory structure state data of each commodity and generating a corresponding inventory structure state vector; the structure diagram module is used for constructing a structure connecting edge and generating a commodity dynamic diagram; the structure aggregation module is used for performing structure feature aggregation on the commodity adjacency relation to generate structure aggregation features; the demand prediction module is used for outputting a demand prediction result and demand plasticity parameters based on the improved iTransform; the inventory decision module is used for executing inventory decision calculation and generating an inventory adjustment track; and the execution feedback module is used for generating an inventory control instruction according to the inventory adjustment track, executing an inventory control operation and updating the inventory structure state data. According to the invention, through machine learning prediction of inventory structure perception, dynamic executable management of trade commodity demands and inventory regulation and control is realized.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a dynamic inventory management system for traded goods based on machine learning. Background Technology

[0002] With the continuous expansion of trade volume and the increasing complexity of commodity types and distribution channels, inventory management has gradually evolved from a traditional experience-driven model to a data-driven and intelligent approach. Existing trade commodity inventory control systems typically estimate commodity demand based on historical sales data, fixed safety stock rules, or simple forecasting models, and then execute inventory operations such as replenishment, allocation, or release accordingly. Some systems have begun to introduce machine learning methods, using time series forecasting models to predict commodity demand, thereby improving the foresight and automation of inventory decisions and alleviating inventory backlog or stockout problems to some extent.

[0003] However, existing technologies in inventory modeling and forecasting generally treat commodities as independent decision-making units, relying primarily on historical demand data for single commodities, lacking a systematic characterization of the inventory structure itself. In actual trade scenarios, commodities are often constrained by multiple structural factors such as replenishment cycles, warehousing coverage, supply routes, inventory transfer feasibility, and sales restrictions or quotas. Different commodities may also have substitution or transfer relationships. Existing inventory forecasting models and decision-making methods struggle to effectively express and utilize this inventory structure information, leading to discrepancies between demand forecasts and actual executable inventory conditions. Inventory decisions are frequently constrained by structural limitations during the execution phase, impacting the overall efficiency and stability of inventory management.

[0004] Existing machine learning-based inventory management solutions mostly focus on demand forecasting itself, lacking the ability to characterize demand changes under different inventory structures. They struggle to reflect the varying responses of demand to replenishment, allocation, and other inventory operations, and also lack the ability to dynamically adjust the relationship between the model and inventory structure based on execution feedback. This makes it difficult to form an effective closed loop in the inventory decision-making process, hindering the continuous optimization of inventory allocation in complex and ever-changing inventory structure environments.

[0005] Therefore, how to provide a dynamic inventory management system for traded goods based on machine learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a dynamic inventory management system for traded goods based on machine learning. This invention comprehensively utilizes machine learning prediction technology, inventory structure modeling technology, and graph structure feature aggregation method to uniformly model and jointly analyze the inventory structure status, inter-commodity relationships, and historical demand information of traded goods. By constructing a dynamic commodity graph with inventory structure awareness and introducing an improved iTransformer model, it achieves accurate prediction of commodity demand and demand plasticity. Based on this, it performs inventory decision calculation and execution feedback updates under multiple constraints, thereby completing the dynamic closed-loop optimization of inventory management. It has the advantages of being able to characterize complex inventory structures, improving the consistency between demand forecasting and inventory decision-making, enhancing the executability of inventory control, and reducing inventory fluctuations and resource consumption.

[0007] A machine learning-based dynamic inventory management system for traded goods, according to an embodiment of the present invention, includes the following modules:

[0008] The structure state module is used to obtain the inventory structure state data of each product and generate the corresponding inventory structure state vector.

[0009] The structure graph module is used to construct structural connection edges with products as nodes and inventory structure state vectors as node attributes, and generate a dynamic product graph.

[0010] The structure aggregation module is used to acquire historical demand time series data of products, aggregate structural features of product adjacency relationships based on the product dynamic graph, and generate structure aggregation features.

[0011] The demand forecasting module is used to input historical demand time series, structural aggregation features and inventory structure state vectors into the improved iTransformer, and output demand forecasting results and demand plasticity parameters.

[0012] The inventory decision module is used to perform inventory decision calculations based on demand forecast results, demand plasticity parameters, and inventory control constraints, and generate inventory adjustment trajectories.

[0013] The execution feedback module is used to generate inventory control instructions based on the inventory adjustment trajectory, execute inventory management operations, collect execution results, and update inventory structure status data.

[0014] Optionally, modules can be integrated using the following methods:

[0015] Obtain the inventory structure status data corresponding to each commodity in the trading system, and construct a corresponding inventory structure status vector for each commodity based on the inventory structure status data.

[0016] Using each product as a graph node and the inventory structure state vector of each product as a node attribute, structural connection edges between products are constructed based on the structural reachability relationship of the inventory structure state vectors between different products, thereby generating a dynamic product graph with inventory structure awareness.

[0017] Obtain historical demand time series data for each commodity, and aggregate structural features of the adjacency relationships of each commodity node based on the commodity dynamic graph to obtain structural aggregation features that characterize the supply availability of the commodity under the current inventory structure.

[0018] The improved iTransformer prediction model takes historical demand time series data, structural aggregation features, and inventory structure state vectors of each commodity as input data, performs attention modeling based on commodity variable dimensions and feature fusion processing that introduces commodity dynamic graph structure constraints, and outputs demand prediction results and demand plasticity parameters for each commodity.

[0019] Based on demand forecast results, demand plasticity parameters, and preset inventory control constraints, inventory decision calculations are performed to jointly determine inventory replenishment, transfer, and release operations for each commodity at multiple time points, and output the corresponding inventory adjustment trajectory.

[0020] Based on the inventory adjustment trajectory, corresponding inventory control instructions are generated and sent to the inventory execution system to complete the inventory management operation. The inventory execution results are collected and the inventory structure status data is updated.

[0021] Optionally, the inventory structure status data includes replenishment cycle parameters, minimum replenishment interval parameters, and replenishment frequency parameters corresponding to the product; warehouse node information and sales area identification information covered by the product; supply path quantity, supply path level, and supply path switching status information corresponding to the product; inventory transfer feasibility identification and transfer constraint information between different warehouse nodes; and inventory lock status, sales restriction status, or quota control status information corresponding to the product.

[0022] Optionally, the construction of the inventory structure state vector includes:

[0023] Standardize the inventory structure status data to convert inventory structure status data of different dimensions into structural feature values ​​of a unified dimension;

[0024] According to the preset inventory structure feature dimension order, the standardized replenishment cycle parameters, warehouse coverage information, supply path information, inventory transfer information and inventory restriction control information are mapped to the corresponding vector dimensions respectively;

[0025] The structural feature values ​​of each vector dimension are concatenated and combined to generate an inventory structure state vector corresponding to a single product.

[0026] Optionally, generating the inventory structure-aware dynamic product graph includes:

[0027] The inventory structure state vector of each commodity is parsed, and the replenishment cycle information, warehouse coverage information, supply path information, inventory transfer information and inventory restriction information are mapped to the corresponding structural feature fields respectively;

[0028] Based on structural feature fields, rule-based matching is performed on the degree of overlap in warehouse coverage, the degree of matching in supply routes, and the degree of similarity in replenishment cycles for any two commodities to obtain candidate association weights;

[0029] The candidate association weights are filtered using inventory transfer feasibility indicators, and only product pairs that can achieve demand transfer under the current inventory rules are retained. A set of structural connection attributes, including association weight, substitution priority and transfer cost level, is established for each retained product pair.

[0030] Read the feedback factors generated based on historical inventory adjustment results, perform adaptive correction processing on the set of structural connection attributes, and obtain the target connection attribute set that can be automatically updated with changes in inventory structure.

[0031] Using a set of goods as the set of nodes and pairs of goods with target connection attributes as the set of edges, multi-attribute product relationship edges are constructed according to the association weight, substitution priority, and allocation cost level in the edge set, generating a dynamic product graph with inventory structure awareness.

[0032] Optionally, the obtained structural aggregation features characterizing the supply availability of goods under the current inventory structure include:

[0033] Read the inventory structure state vector of each product and the historical demand time series of each product, slice the historical demand time series according to the preset time granularity, and generate the time sequence code input of each product;

[0034] Read the dynamic graph of goods perceived by the inventory structure, determine the set of adjacency relationships for each goods, sort and stratify the set of adjacency relationships according to the association weight, substitution priority and transfer cost level, and remove adjacency relationships that do not meet the reachability conditions according to the transfer feasibility indicator.

[0035] For each product, structural feature aggregation processing is performed on the set of adjacency relationships after sorting and layering. Using the association weight and substitution priority of each layer as control parameters, the temporal coding results of adjacent products within the layer are fused within the layer, and the corresponding layer aggregation features are output.

[0036] Under the conditions of satisfying the preset reachability rules and cost limits, structural feature aggregation processing is performed on each product based on the product dynamic graph. Path-level fusion is performed on the path composed of two continuous structural connecting edges to generate path aggregation features. The feedback factor is used as the masking condition to perform masking processing on the path that does not meet the stability requirements. The contribution of the masked path is set to invalid to form the effect of structural attention masking.

[0037] The product's structural aggregate features are generated by merging its own time-series encoding results, layer aggregation features, path aggregation features, and inventory structure state vector.

[0038] Optionally, the output of demand forecast results and demand plasticity parameters for each commodity includes:

[0039] The structural aggregation features, inventory structure state vector, and historical demand time series data of each product are sequentially written into the input cache. The historical demand time series is divided into fixed-length windows and identified by the product dimension. The data stream output by the cache is defined as a variable-level input token sequence.

[0040] The historical demand fragments, structural aggregation features and inventory structure state vectors in the variable-level input token are concatenated in a preset order. Dense mapping is performed on the product identifier and a three-dimensional structural bias consisting of warehouse coverage identifier, supply path level and allocation constraint level is superimposed. Then, the corresponding time position code is added to obtain a unified dimension of embedded vector sequence.

[0041] A sequence of embedding vectors of a uniform dimension is fed into an improved iTransformer prediction model. The improved iTransformer prediction model consists of a backbone composed of six variable-dimensional self-attention blocks and six feedforward blocks stacked alternately. An inventory structure adaptive gating module and a demand fluctuation suppression attention module are inserted after the third and fifth self-attention blocks, respectively. The inventory structure adaptive gating module dynamically adjusts the attention distribution according to the correlation weights between commodities, and the demand fluctuation suppression attention module dynamically suppresses high-frequency fluctuation features according to the historical demand variance. All attention blocks are loaded with a structural bias matrix to limit the effective range of cross-commodity information propagation.

[0042] The output of the backbone network is fed into the dual-branch output head. The first branch uses a one-dimensional convolutional layer with linear mapping to generate a demand forecast sequence for multiple time steps in the future using a sliding window method. The second branch uses a two-layer fully connected network to generate a demand plasticity parameter vector that represents the demand response to changes in inventory structure.

[0043] Using historical demand sequences and corresponding actual demand values ​​and inventory adjustment records as supervision signals, gradient training is performed on the improved iTransformer. The training objective is to minimize the comprehensive loss consisting of demand forecasting error, demand plasticity error, and structural consistency constraint. After training, the demand forecasting sequence and corresponding demand plasticity parameter vector for each commodity are output.

[0044] Optionally, the inventory adjustment trajectory corresponding to the output includes:

[0045] Read the demand forecast sequence and demand plasticity parameters of each commodity, read the commodity dynamic diagram and feedback factors of inventory structure perception, read the inventory structure state vector and inventory control constraints, set the planning cycle and time granularity and establish the decision time axis.

[0046] Based on demand plasticity parameters and commodity dynamics, a plasticity constraint set and a structural compatibility constraint set are generated for each commodity at each point in time, forming a constraint list that includes allowable replenishment range, allowable transfer range, allowable release range and prohibited sections. The consistency of the constraint list is verified based on the upper limit of storage capacity, upper limit of capital occupation, supply lead time, minimum purchase batch, lower limit of demand fulfillment rate and transfer route permission.

[0047] Under the premise of satisfying the constraint list, a set of candidate actions is generated for each product on the decision timeline. The set of candidate actions consists of discretized replenishment operations, transfer operations and release operations. For each candidate action, four evaluation values ​​are calculated: service level impact, inventory holding impact, transfer execution impact and stability impact, and a candidate action pool is constructed.

[0048] Inventory decision calculation is performed on the candidate action pool. A two-stage path solving method is used to determine the inventory adjustment trajectory. In the first stage, an reachable path search is performed under structural compatibility constraints to obtain an initial trajectory that meets the service level lower limit and does not enter the prohibited section. In the second stage, local improvement and smoothing are performed under the constraints of storage capacity and capital occupation to reduce the fluctuation of the trajectory. Backtracking consistency checks are performed on each action to avoid triggering demand amplification.

[0049] The action sequence after two-stage path solving and backtracking consistency check is combined in chronological order to form an inventory adjustment trajectory, forming an execution list that includes time points, warehouse nodes, product identifiers and quantity instructions.

[0050] Optionally, the step of generating corresponding inventory control instructions based on inventory adjustment trajectories and sending the inventory control instructions to the inventory execution system to complete inventory management operations includes:

[0051] Read the inventory adjustment trajectory, parse the time points, warehouse nodes, product identifiers and quantity instructions contained in the inventory adjustment trajectory into an execution list, and generate an inventory control instruction set according to the preset interface protocol and field format;

[0052] The inventory control instruction set is sent to the inventory execution system, which then performs replenishment, transfer, and release operations on the corresponding warehouse nodes at the corresponding time points. During the execution process, the actual execution status, actual execution quantity, and abnormal information of each inventory control instruction are recorded.

[0053] The inventory execution results, which correspond one-to-one with the inventory control instruction set, are collected from the inventory execution system. The actual execution status, actual execution quantity, and abnormal information are compared with the execution list to generate summary data of execution results, including execution deviations, unexecuted instructions, and partial execution instruction identifiers.

[0054] The inventory structure status data is updated based on the summary data of the execution results. Changes in warehouse coverage, supply path, allocation feasibility and inventory restriction status caused by inventory execution are written into the new inventory structure status vector, and the feedback factors used by the commodity dynamic graph of inventory structure perception are updated.

[0055] The beneficial effects of this invention are:

[0056] This invention systematically models the inventory structure of traded goods, mapping multi-dimensional structural factors such as replenishment cycle, warehousing coverage, supply path, inventory allocation, and inventory constraints into a unified inventory structure state vector. Based on this vector, a dynamic commodity graph with inventory structure perception is constructed. This allows inventory management to move beyond a single commodity or time series dimension, instead depicting the substitutability and supply availability relationships between commodities at a global structural level. This fundamentally improves the completeness and accuracy of inventory state representation, providing a structured input foundation that is highly consistent with actual executable conditions for subsequent demand forecasting and inventory decisions.

[0057] At the demand forecasting and decision-making level, this invention introduces an improved iTransformer model that incorporates constraints from the dynamic graph structure of commodities. This model jointly models historical demand time series, structural aggregation features, and inventory structure state vectors. It not only enables multi-time-step forecasting of future commodity demand but also outputs demand plasticity parameters that characterize the ability of demand to respond to changes in inventory structure. This allows the forecast results to reflect the differences in demand fluctuations under different inventory structure conditions, thereby significantly reducing the deviation between the forecast results and inventory execution conditions and improving the rationality and stability of inventory replenishment, allocation, and release decisions under complex constraints.

[0058] This invention transforms inventory decision-making results into inventory adjustment trajectories and introduces an execution feedback mechanism to dynamically update deviations and structural changes during inventory execution. This allows the inventory structure status, product dynamic diagrams, and model input parameters to be continuously corrected based on actual execution results, forming a closed-loop control mechanism that links prediction, decision-making, and execution. This effectively suppresses the demand amplification effect during inventory adjustment, enhances the adaptability and long-term operational stability of inventory control, and has higher application value and promotion significance in trade scenarios with multiple commodities, multiple warehouses, and multiple constraints. Attached Figure Description

[0059] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0060] Figure 1 This is a schematic diagram of the structure of a machine learning-based dynamic inventory management system for traded goods proposed in this invention.

[0061] Figure 2 This is a flowchart illustrating a machine learning-based dynamic inventory management method for traded goods proposed in this invention. Detailed Implementation

[0062] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0063] refer to Figure 1 A machine learning-based dynamic inventory management system for traded goods includes the following modules:

[0064] The structure state module is used to obtain the inventory structure state data of each product and generate the corresponding inventory structure state vector.

[0065] The structure graph module is used to construct structural connection edges with products as nodes and inventory structure state vectors as node attributes, and generate a dynamic product graph.

[0066] The structure aggregation module is used to acquire historical demand time series data of products, aggregate structural features of product adjacency relationships based on the product dynamic graph, and generate structure aggregation features.

[0067] The demand forecasting module is used to input historical demand time series, structural aggregation features and inventory structure state vectors into the improved iTransformer, and output demand forecasting results and demand plasticity parameters.

[0068] The inventory decision module is used to perform inventory decision calculations based on demand forecast results, demand plasticity parameters, and inventory control constraints, and generate inventory adjustment trajectories.

[0069] The execution feedback module is used to generate inventory control instructions based on the inventory adjustment trajectory, execute inventory management operations, collect execution results, and update inventory structure status data.

[0070] refer to Figure 2 A machine learning-based method for dynamic inventory management of traded goods includes:

[0071] Obtain the inventory structure status data corresponding to each commodity in the trading system, and construct a corresponding inventory structure status vector for each commodity based on the inventory structure status data.

[0072] Using each product as a graph node and the inventory structure state vector of each product as a node attribute, structural connection edges between products are constructed based on the structural reachability relationship of the inventory structure state vectors between different products, thereby generating a dynamic product graph with inventory structure awareness.

[0073] Obtain historical demand time series data for each commodity, and aggregate structural features of the adjacency relationships of each commodity node based on the commodity dynamic graph to obtain structural aggregation features that characterize the supply availability of the commodity under the current inventory structure.

[0074] The improved iTransformer prediction model takes historical demand time series data, structural aggregation features, and inventory structure state vectors of each commodity as input data, performs attention modeling based on commodity variable dimensions and feature fusion processing that introduces commodity dynamic graph structure constraints, and outputs demand prediction results and demand plasticity parameters for each commodity.

[0075] Based on demand forecast results, demand plasticity parameters, and preset inventory control constraints, inventory decision calculations are performed to jointly determine inventory replenishment, transfer, and release operations for each commodity at multiple time points, and output the corresponding inventory adjustment trajectory.

[0076] Based on the inventory adjustment trajectory, corresponding inventory control instructions are generated and sent to the inventory execution system to complete the inventory management operation. The inventory execution results are collected and the inventory structure status data is updated.

[0077] In this embodiment, the inventory structure status data includes the replenishment cycle parameters, minimum replenishment interval parameters, and replenishment frequency parameters corresponding to the product; warehouse node information and sales area identification information covered by the product; the number of supply paths, supply path levels, and supply path switching status information corresponding to the product; inventory transfer feasibility identification and transfer constraint information between different warehouse nodes; and inventory lock status, sales restriction status, or quota control status information corresponding to the product.

[0078] In this embodiment, the construction of the inventory structure state vector includes:

[0079] Standardize the inventory structure status data to convert inventory structure status data of different dimensions into structural feature values ​​of a unified dimension;

[0080] According to the preset inventory structure feature dimension order, the standardized replenishment cycle parameters, warehouse coverage information, supply path information, inventory transfer information and inventory restriction control information are mapped to the corresponding vector dimensions respectively;

[0081] The structural feature values ​​of each vector dimension are concatenated and combined to generate an inventory structure state vector corresponding to a single product.

[0082] In this embodiment, generating the inventory structure-aware dynamic product graph includes:

[0083] The inventory structure state vector of each commodity is parsed, and the replenishment cycle information, warehouse coverage information, supply path information, inventory transfer information and inventory restriction information are mapped to the corresponding structural feature fields respectively;

[0084] Based on structural feature fields, rule-based matching is performed on the degree of overlap in warehouse coverage, the degree of matching in supply routes, and the degree of similarity in replenishment cycles for any two commodities to obtain candidate association weights;

[0085] The candidate association weights are filtered using inventory transfer feasibility indicators, and only product pairs that can achieve demand transfer under the current inventory rules are retained. A set of structural connection attributes, including association weight, substitution priority and transfer cost level, is established for each retained product pair.

[0086] Read the feedback factors generated based on historical inventory adjustment results, perform adaptive correction processing on the set of structural connection attributes, and obtain the target connection attribute set that can be automatically updated with changes in inventory structure.

[0087] Using a set of goods as the set of nodes and pairs of goods with target connection attributes as the set of edges, multi-attribute product relationship edges are constructed according to the association weight, substitution priority, and allocation cost level in the edge set, generating a dynamic product graph with inventory structure awareness.

[0088] In this embodiment, obtaining the structural aggregation features characterizing the supply availability of goods under the current inventory structure includes:

[0089] Read the inventory structure state vector of each product and the historical demand time series of each product. Slice the historical demand time series according to the preset time granularity to generate the time sequence code input of each product. The preset time granularity is the daily time granularity, that is, the historical demand data is divided with a single day as the smallest time unit.

[0090] Read the dynamic graph of goods perceived by the inventory structure, determine the set of adjacency relationships for each goods, sort and stratify the set of adjacency relationships according to the association weight, substitution priority and transfer cost level, and remove adjacency relationships that do not meet the reachability conditions according to the transfer feasibility indicator.

[0091] For each product, structural feature aggregation processing is performed on the set of adjacency relationships after sorting and layering. Using the association weight and substitution priority of each layer as control parameters, the temporal coding results of adjacent products within the layer are fused within the layer, and the corresponding layer aggregation features are output.

[0092] Under the conditions of satisfying the preset reachability rules and cost limits, structural feature aggregation processing is performed on each product based on the product dynamic graph. Path-level fusion is performed on the path composed of two continuous structural connecting edges to generate path aggregation features. The feedback factor is used as the masking condition to perform masking processing on the path that does not meet the stability requirements. The contribution of the masked path is set to invalid to form the effect of structural attention masking.

[0093] The product's structural aggregate features are generated by merging its own time-series encoding results, layer aggregation features, path aggregation features, and inventory structure state vector.

[0094] In this embodiment, the output of demand forecast results and demand plasticity parameters for each commodity includes:

[0095] The structural aggregation features, inventory structure state vector, and historical demand time series data of each product are sequentially written into the input cache. The historical demand time series is divided into fixed-length windows and identified by the product dimension. The data stream output by the cache is defined as a variable-level input token sequence.

[0096] The historical demand fragments, structural aggregation features and inventory structure state vectors in the variable-level input token are concatenated in a preset order. Dense mapping is performed on the product identifier and a three-dimensional structural bias consisting of warehouse coverage identifier, supply path level and allocation constraint level is superimposed. Then, the corresponding time position code is added to obtain a unified dimension of embedded vector sequence.

[0097] A sequence of embedding vectors of uniform dimension is fed into an improved iTransformer prediction model. This improved model comprises a backbone consisting of six variable-dimensional self-attention blocks and six feedforward blocks stacked alternately. An inventory structure adaptive gating module and a demand fluctuation suppression attention module are inserted after the third and fifth self-attention blocks, respectively. The inventory structure adaptive gating module dynamically adjusts the attention distribution based on the inter-product correlation weights, while the demand fluctuation suppression attention module dynamically suppresses high-frequency fluctuations based on historical demand variance. All attention blocks are loaded with a structural bias matrix to limit the effective range of cross-product information propagation.

[0098] The inventory structure adaptive gating module dynamically adjusts the attention distribution based on the correlation weights between products. Specifically, during the self-attention calculation process of variable dimensions, a correlation weight matrix based on the product dynamic graph is introduced as a gating control factor to adjust the attention weights between different product variables. This allows product variables with higher structural correlation or accessibility to receive higher attention allocation, while the attention weights corresponding to product variables with lower structural correlation or no accessibility are attenuated or suppressed. This enables the model to adaptively perceive inventory structure relationships during the attention modeling stage.

[0099] The demand fluctuation suppression attention module dynamically suppresses high-frequency fluctuation characteristics based on historical demand variance. Specifically, it calculates the demand variance and fluctuation amplitude index of each commodity's historical demand time series within a preset time window, generates a demand fluctuation intensity coefficient, and introduces the demand fluctuation intensity coefficient into the attention weight calculation process to dynamically attenuate the attention response corresponding to high fluctuation and high noise characteristics, thereby reducing the impact of short-term abnormal fluctuations, promotional noise, or occasional demand shocks on the prediction results.

[0100] All attention blocks are loaded with a structural bias matrix to limit the effective scope of cross-product information propagation. Specifically, a structural bias matrix consisting of warehouse coverage identifier, supply path level and transfer constraint level is introduced into the attention block for each variable dimension. The structural bias matrix is ​​superimposed on the attention score calculation item. Negative bias or masking is applied to product variables that do not meet the inventory structure reachability conditions or violate transfer constraints.

[0101] The output of the backbone network is fed into the dual-branch output head. The first branch uses a one-dimensional convolutional layer with linear mapping to generate a demand forecast sequence for multiple time steps in the future using a sliding window method. The second branch uses a two-layer fully connected network to generate a demand plasticity parameter vector that represents the demand response to changes in inventory structure.

[0102] Using historical demand sequences and corresponding actual demand values ​​and inventory adjustment records as supervision signals, gradient training is performed on the improved iTransformer. The training objective is to minimize the comprehensive loss consisting of demand forecasting error, demand plasticity error, and structural consistency constraint. After training, the demand forecasting sequence and corresponding demand plasticity parameter vector for each commodity are output.

[0103] This invention improves the iTransformer model by introducing inventory structure awareness and commodity association constraint mechanisms. Specifically, it unifies historical demand time series, inventory structure state vectors, and structural aggregation features obtained from commodity dynamic graphs into variable-level input tokens. During the embedding stage, structural biases formed by warehouse coverage, supply path levels, and allocation constraint levels are superimposed, enabling the model to explicitly perceive inventory structure relationships during variable-dimensional modeling. Simultaneously, an inventory structure adaptive gating module is inserted into the backbone network to dynamically adjust the attention distribution based on the association weights between commodities. A demand fluctuation suppression attention module is introduced to reduce the impact of high-frequency demand fluctuations on prediction stability. Finally, through a dual-branch output structure, demand plasticity parameters are output while predicting future multi-time-step demand. Joint optimization is performed during the training stage in conjunction with structural consistency constraints, enabling the model to achieve more stable and executable demand forecasting under complex inventory structure conditions.

[0104] In this embodiment, the output corresponding inventory adjustment trajectory includes:

[0105] Read the demand forecast sequence and demand plasticity parameters of each commodity, read the commodity dynamic diagram and feedback factors of inventory structure perception, read the inventory structure state vector and inventory control constraints, set the planning cycle and time granularity and establish the decision time axis.

[0106] Based on demand plasticity parameters and commodity dynamics, a plasticity constraint set and a structural compatibility constraint set are generated for each commodity at each point in time, forming a constraint list that includes allowable replenishment range, allowable transfer range, allowable release range and prohibited sections. The consistency of the constraint list is verified based on the upper limit of storage capacity, upper limit of capital occupation, supply lead time, minimum purchase batch, lower limit of demand fulfillment rate and transfer route permission.

[0107] Under the premise of satisfying the constraint list, a set of candidate actions is generated for each product on the decision timeline. The set of candidate actions consists of discretized replenishment operations, transfer operations and release operations. For each candidate action, four evaluation values ​​are calculated: service level impact, inventory holding impact, transfer execution impact and stability impact, and a candidate action pool is constructed.

[0108] Inventory decision-making is performed on the candidate action pool, employing a two-stage path solving approach to determine the inventory adjustment trajectory. The first stage involves a reachable path search under structural compatibility constraints to obtain an initial trajectory that satisfies the service level lower bound and avoids entering prohibited areas. The second stage involves local improvements and smoothing under warehouse capacity and capital constraints to reduce trajectory volatility. A backtracking consistency check is performed on each action to prevent demand amplification. Specifically:

[0109] The first stage involves searching for reachable paths under structural compatibility constraints to obtain an initial trajectory that meets the service level lower limit and does not enter prohibited areas. Specifically, based on the product dynamic graph and the set of structural compatibility constraints, the reachability of replenishment, transfer, and release actions of each product in the candidate action pool at adjacent time nodes is determined. Action combinations that violate the inventory structure status, transfer feasibility indicators, or demand plasticity constraints are eliminated. In the remaining action space, a path search method that expands step by step is adopted to prioritize the action sequence that can meet the preset demand satisfaction rate lower limit and does not trigger the inventory prohibited area, thus forming an initial inventory adjustment trajectory that is executable under structural constraints.

[0110] The second stage involves local improvements and smoothing under the constraints of storage capacity and capital occupation to reduce the volatility of the trajectory. Specifically, while maintaining the accessibility of the initial inventory adjustment trajectory structure, the inventory operation quantity at each time node in the trajectory is locally adjusted. By applying smoothing constraints to the replenishment magnitude, transfer quantity, and release scale of adjacent time nodes, the inventory change magnitude is ensured not to exceed the preset threshold. At the same time, the upper limit of storage capacity and the upper limit of capital occupation are checked. For nodes that exceed the constraints, backtracking correction and neighbor replacement processing are performed. This reduces the overall volatility of the inventory adjustment trajectory while meeting the inventory execution constraints. Furthermore, backtracking consistency checks prevent the demand amplification effect caused by continuous adjustments.

[0111] The action sequence after two-stage path solving and backtracking consistency check is combined in chronological order to form an inventory adjustment trajectory, forming an execution list that includes time points, warehouse nodes, product identifiers and quantity instructions.

[0112] In this embodiment, the step of generating corresponding inventory control instructions based on inventory adjustment trajectories and sending the inventory control instructions to the inventory execution system to complete inventory management operations includes:

[0113] Read the inventory adjustment trajectory, parse the time points, warehouse nodes, product identifiers, and quantity instructions contained in the trajectory into an execution list, and generate an inventory control instruction set according to the preset interface protocol and field format, including:

[0114] The default interface protocol is a standardized interface call protocol based on HTTP, which adopts a request-response communication method and uses JSON as the data exchange carrier;

[0115] The field format is a structured instruction field format, including instruction number field, instruction type field, execution time field, warehouse node identifier field, product identifier field, operation type field, and operation quantity field;

[0116] The inventory control instruction set is sent to the inventory execution system, which then performs replenishment, transfer, and release operations on the corresponding warehouse nodes at the corresponding time points. During the execution process, the actual execution status, actual execution quantity, and abnormal information of each inventory control instruction are recorded.

[0117] The inventory execution results, which correspond one-to-one with the inventory control instruction set, are collected from the inventory execution system. The actual execution status, actual execution quantity, and abnormal information are compared with the execution list to generate summary data of execution results, including execution deviations, unexecuted instructions, and partial execution instruction identifiers.

[0118] The inventory structure status data is updated based on the summary data of the execution results. Changes in warehouse coverage, supply path, allocation feasibility and inventory restriction status caused by inventory execution are written into the new inventory structure status vector, and the feedback factors used by the commodity dynamic graph of inventory structure perception are updated.

[0119] Example 1:

[0120] To verify the feasibility of this invention in practice, it was applied to a trading company that has long faced problems such as complex inventory structure, implicit product relationships, and a disconnect between inventory decision-making and execution. The company's products cover multiple categories, including fast-moving consumer goods and daily necessities. Different products exhibit significant differences in replenishment cycles, warehousing coverage, supply path quantity, and inventory transfer feasibility. Furthermore, limited by warehousing capacity, capital occupation, and transfer rules, the inventory structure exhibits highly dynamic characteristics. In actual operation, the company's original inventory management method mainly relied on historical demand time series forecasts for individual products, combined with fixed safety stock rules for replenishment and transfer operations. This failed to fully consider the inventory structure and the substitutability between products, leading to a mismatch between forecast results and actual executable inventory conditions. During inventory adjustments, frequent transfer obstacles, increased manual intervention, and amplified inventory fluctuations occurred.

[0121] In this business scenario, the machine learning-based dynamic inventory management system for trade goods described in this invention is introduced to manage the inventory of the enterprise's core products. The system first continuously acquires replenishment cycle parameters, inventory constraint status, warehouse coverage information, supply path structure, and inventory transfer feasibility data for each product from the order system and inventory system, and constructs the above multi-dimensional inventory structure information into an inventory structure state vector. Based on the inventory structure state vector, the system uses products as nodes and combines the degree of warehouse coverage overlap, supply path matching relationship, and transfer feasibility constraints to construct a dynamic product graph with inventory structure awareness, explicitly expressing the potential demand transfer and supply availability relationships between products within the system.

[0122] In the demand forecasting and feature modeling process, the system uses the historical demand time series and inventory structure state vector of each commodity as input. It aggregates structural features of adjacent commodities through a commodity dynamic graph, generating structural aggregation features that characterize the supply availability of commodities under the current inventory structure. These features are then input into the improved iTransformer forecasting model, which performs attention modeling on the variable dimensions. During the forecasting process, an adaptive gating mechanism for inventory structure and a demand fluctuation suppression mechanism are introduced to limit information interference between commodities that lack structural reachability while ensuring forecasting accuracy. The model outputs demand forecasts for multiple future time steps and simultaneously generates demand plasticity parameters to characterize the responsiveness of demand under different inventory structure changes.

[0123] During the inventory decision-making phase, the system, based on demand forecast results, demand plasticity parameters, and inventory structure state vectors, and combined with inventory control constraints such as storage capacity, capital occupation, minimum purchase quantity, and demand fulfillment rate, jointly calculates the replenishment, allocation, and inventory release operations for each commodity within the planning cycle, generating an inventory adjustment trajectory with time sequence and quantity instructions. This inventory adjustment trajectory is converted into inventory control instructions and sent to the inventory execution system. Execution results are collected in real time and used to update the inventory structure state data and feedback factors in the commodity dynamic graph, enabling the system to continuously adjust its forecasting and decision-making logic based on actual execution, forming a dynamic closed loop of inventory forecasting, decision-making, and execution.

[0124] Table 1 Comparison of the operational effects of different inventory control schemes

[0125] Indicator Name Traditional solution Invention Solution Improvement range Statistical period Data source Average error rate of demand forecast 18.7% 14.9% ↓3.8% 3 months Order records Average inventory turnover days 31.4 days 27.6 days ↓3.8 days 3 months Inventory system Inventory transfer execution success rate 86.2% 92.5% ↑6.3% 3 months Transfer Records Number of abnormal inventory fluctuations 14 times 6 times ↓8 times 3 months Run log Frequency of human intervention 11 times 4 times ↓7 times 3 months Operation and maintenance records Service level achievement rate 93.1% 95.8% ↑2.7% 3 months Delivery statistics

[0126] As shown in Table 1, from the perspective of demand forecasting and inventory efficiency, the traditional approach, relying solely on historical demand time series for single commodities, has an average demand forecasting error rate of 18.7%. The forecasting results are easily affected by the lack of constraints on inventory structure, leading to replenishment and allocation decisions deviating from actual feasibility. The proposed solution, by jointly modeling inventory structure status, commodity relationships, and historical demand information, reduces the average demand forecasting error rate to 14.9%, achieving stable improvement within the same statistical period. Simultaneously, the average inventory turnover days decreased from 31.4 days to 27.6 days, indicating a more reasonable inventory adjustment rhythm, effective control of inventory resource utilization, and an overall improvement in inventory operating efficiency.

[0127] Regarding inventory execution stability, the traditional approach achieves an 86.2% success rate for inventory transfers. However, a certain percentage of transfers still fail or experience abnormal fluctuations, leading to a high frequency of abnormal inventory fluctuations and manual interventions. By adopting the solution of this invention, the success rate of inventory transfers increases to 92.5%, the number of abnormal inventory fluctuations decreases from 14 to 6, and the frequency of manual interventions decreases from 11 to 4. This demonstrates that introducing structural accessibility constraints and demand plasticity parameters during the inventory decision-making stage helps reduce unexecutable or high-risk inventory operations, making the inventory adjustment process more stable and controllable.

[0128] In terms of service level and overall operational effectiveness, the traditional solution achieves a service level attainment rate of 93.1%, but still suffers from insufficient supply response in scenarios with fluctuating demand. The solution of this invention improves the service level attainment rate to 95.8% without significantly increasing inventory size, and maintains good stability over continuous statistical periods. This demonstrates the continuous optimization capability of the prediction, decision-making, and execution closed-loop mechanism for inventory management quality, verifying the practicality and feasibility of this invention in complex inventory structure environments.

[0129] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A dynamic inventory management system for traded goods based on machine learning, characterized in that, Includes the following modules: The structure state module is used to obtain the inventory structure state data of each product and generate the corresponding inventory structure state vector. The structure graph module is used to construct structural connection edges with products as nodes and inventory structure state vectors as node attributes, and generate a dynamic product graph. The structure aggregation module is used to acquire historical demand time series data of products, aggregate structural features of product adjacency relationships based on the product dynamic graph, and generate structure aggregation features. The demand forecasting module is used to input historical demand time series, structural aggregation features and inventory structure state vectors into the improved iTransformer, and output demand forecasting results and demand plasticity parameters. The inventory decision module is used to perform inventory decision calculations based on demand forecast results, demand plasticity parameters, and inventory control constraints, and generate inventory adjustment trajectories. The execution feedback module is used to generate inventory control instructions based on the inventory adjustment trajectory, execute inventory management operations, collect execution results, and update inventory structure status data.

2. A machine learning-based dynamic inventory management method for traded goods, applied to the machine learning-based dynamic inventory management system for traded goods as described in claim 1, characterized in that, include: Obtain the inventory structure status data corresponding to each commodity in the trading system, and construct a corresponding inventory structure status vector for each commodity based on the inventory structure status data. Using each product as a graph node and the inventory structure state vector of each product as a node attribute, structural connection edges between products are constructed based on the structural reachability relationship of the inventory structure state vectors between different products, thereby generating a dynamic product graph with inventory structure awareness. Obtain historical demand time series data for each commodity, and aggregate structural features of the adjacency relationships of each commodity node based on the commodity dynamic graph to obtain structural aggregation features that characterize the supply availability of the commodity under the current inventory structure. The improved iTransformer prediction model takes historical demand time series data, structural aggregation features, and inventory structure state vectors of each commodity as input data, performs attention modeling based on commodity variable dimensions and feature fusion processing that introduces commodity dynamic graph structure constraints, and outputs demand prediction results and demand plasticity parameters for each commodity. Based on demand forecast results, demand plasticity parameters, and preset inventory control constraints, inventory decision calculations are performed to jointly determine inventory replenishment, transfer, and release operations for each commodity at multiple time points, and output the corresponding inventory adjustment trajectory. Based on the inventory adjustment trajectory, corresponding inventory control instructions are generated and sent to the inventory execution system to complete the inventory management operation. The inventory execution results are collected and the inventory structure status data is updated.

3. The method for dynamic inventory management of traded commodities based on machine learning according to claim 2, characterized in that, The inventory structure status data includes the replenishment cycle parameters, minimum replenishment interval parameters, and replenishment frequency parameters corresponding to the product; warehouse node information and sales area identification information covered by the product; the number of supply paths, supply path levels, and supply path switching status information corresponding to the product; inventory transfer feasibility identification and transfer constraint information between different warehouse nodes; and inventory lock status, sales restriction status, or quota control status information corresponding to the product.

4. The method for dynamic inventory management of traded commodities based on machine learning according to claim 2, characterized in that, The construction of the inventory structure state vector includes: Standardize the inventory structure status data to convert inventory structure status data of different dimensions into structural feature values ​​of a unified dimension; According to the preset inventory structure feature dimension order, the standardized replenishment cycle parameters, warehouse coverage information, supply path information, inventory transfer information and inventory restriction control information are mapped to the corresponding vector dimensions respectively; The structural feature values ​​of each vector dimension are concatenated and combined to generate an inventory structure state vector corresponding to a single product.

5. The method for dynamic inventory management of traded commodities based on machine learning according to claim 2, characterized in that, The generation of the inventory structure-aware product dynamic map includes: The inventory structure state vector of each commodity is parsed, and the replenishment cycle information, warehouse coverage information, supply path information, inventory transfer information and inventory restriction information are mapped to the corresponding structural feature fields respectively; Based on structural feature fields, rule-based matching is performed on the degree of overlap in warehouse coverage, the degree of matching in supply routes, and the degree of similarity in replenishment cycles for any two commodities to obtain candidate association weights; The candidate association weights are filtered using inventory transfer feasibility indicators, and only product pairs that can achieve demand transfer under the current inventory rules are retained. A set of structural connection attributes, including association weight, substitution priority and transfer cost level, is established for each retained product pair. Read the feedback factors generated based on historical inventory adjustment results, perform adaptive correction processing on the set of structural connection attributes, and obtain the target connection attribute set that can be automatically updated with changes in inventory structure. Using a set of goods as the set of nodes and pairs of goods with target connection attributes as the set of edges, multi-attribute product relationship edges are constructed according to the association weight, substitution priority, and allocation cost level in the edge set, generating a dynamic product graph with inventory structure awareness.

6. The method for dynamic inventory management of traded commodities based on machine learning according to claim 2, characterized in that, The structural aggregation features that characterize the supply availability of goods under the current inventory structure include: Read the inventory structure state vector of each product and the historical demand time series of each product, slice the historical demand time series according to the preset time granularity, and generate the time sequence code input of each product; Read the dynamic graph of goods perceived by the inventory structure, determine the set of adjacency relationships for each goods, sort and stratify the set of adjacency relationships according to the association weight, substitution priority and transfer cost level, and remove adjacency relationships that do not meet the reachability conditions according to the transfer feasibility indicator. For each product, structural feature aggregation processing is performed on the set of adjacency relationships after sorting and layering. Using the association weight and substitution priority of each layer as control parameters, the temporal coding results of adjacent products within the layer are fused within the layer, and the corresponding layer aggregation features are output. Under the conditions of satisfying the preset reachability rules and cost limits, structural feature aggregation processing is performed on each product based on the product dynamic graph. Path-level fusion is performed on the path composed of two continuous structural connecting edges to generate path aggregation features. The feedback factor is used as the masking condition to perform masking processing on the path that does not meet the stability requirements. The contribution of the masked path is set to invalid to form the effect of structural attention masking. The product's structural aggregate features are generated by merging its own time-series encoding results, layer aggregation features, path aggregation features, and inventory structure state vector.

7. The method for dynamic inventory management of traded commodities based on machine learning according to claim 2, characterized in that, The output of demand forecast results and demand plasticity parameters for each commodity includes: The structural aggregation features, inventory structure state vector, and historical demand time series data of each product are sequentially written into the input cache. The historical demand time series is divided into fixed-length windows and identified by the product dimension. The data stream output by the cache is defined as a variable-level input token sequence. The historical demand fragments, structural aggregation features and inventory structure state vectors in the variable-level input token are concatenated in a preset order. Dense mapping is performed on the product identifier and a three-dimensional structural bias consisting of warehouse coverage identifier, supply path level and allocation constraint level is superimposed. Then, the corresponding time position code is added to obtain a unified dimension of embedded vector sequence. A sequence of embedding vectors of a uniform dimension is fed into an improved iTransformer prediction model. The improved iTransformer prediction model consists of a backbone composed of six variable-dimensional self-attention blocks and six feedforward blocks stacked alternately. An inventory structure adaptive gating module and a demand fluctuation suppression attention module are inserted after the third and fifth self-attention blocks, respectively. The inventory structure adaptive gating module dynamically adjusts the attention distribution according to the correlation weights between commodities, and the demand fluctuation suppression attention module dynamically suppresses high-frequency fluctuation features according to the historical demand variance. All attention blocks are loaded with a structural bias matrix to limit the effective range of cross-commodity information propagation. The output of the backbone network is fed into the dual-branch output head. The first branch uses a one-dimensional convolutional layer with linear mapping to generate a demand forecast sequence for multiple time steps in the future using a sliding window method. The second branch uses a two-layer fully connected network to generate a demand plasticity parameter vector that represents the demand response to changes in inventory structure. Using historical demand sequences and corresponding actual demand values ​​and inventory adjustment records as supervision signals, gradient training is performed on the improved iTransformer. The training objective is to minimize the comprehensive loss consisting of demand forecasting error, demand plasticity error, and structural consistency constraint. After training, the demand forecasting sequence and corresponding demand plasticity parameter vector for each commodity are output.

8. The method for dynamic inventory management of traded commodities based on machine learning according to claim 2, characterized in that, The inventory adjustment trajectory corresponding to the output includes: Read the demand forecast sequence and demand plasticity parameters of each commodity, read the commodity dynamic diagram and feedback factors of inventory structure perception, read the inventory structure state vector and inventory control constraints, set the planning cycle and time granularity and establish the decision time axis. Based on demand plasticity parameters and commodity dynamics, a plasticity constraint set and a structural compatibility constraint set are generated for each commodity at each point in time, forming a constraint list that includes allowable replenishment range, allowable transfer range, allowable release range and prohibited sections. The consistency of the constraint list is verified based on the upper limit of storage capacity, upper limit of capital occupation, supply lead time, minimum purchase batch, lower limit of demand fulfillment rate and transfer route permission. Under the premise of satisfying the constraint list, a set of candidate actions is generated for each product on the decision timeline. The set of candidate actions consists of discretized replenishment operations, transfer operations and release operations. For each candidate action, four evaluation values ​​are calculated: service level impact, inventory holding impact, transfer execution impact and stability impact, and a candidate action pool is constructed. Inventory decision calculation is performed on the candidate action pool. A two-stage path solving method is used to determine the inventory adjustment trajectory. In the first stage, an reachable path search is performed under structural compatibility constraints to obtain an initial trajectory that meets the service level lower limit and does not enter the prohibited section. In the second stage, local improvement and smoothing are performed under the constraints of storage capacity and capital occupation to reduce the fluctuation of the trajectory. Backtracking consistency checks are performed on each action to avoid triggering demand amplification. The action sequence after two-stage path solving and backtracking consistency check is combined in chronological order to form an inventory adjustment trajectory, forming an execution list that includes time points, warehouse nodes, product identifiers and quantity instructions.

9. A method for dynamic inventory management of traded commodities based on machine learning according to claim 2, characterized in that, The process of generating corresponding inventory control instructions based on inventory adjustment trajectories and sending these instructions to the inventory execution system to complete inventory management operations includes: Read the inventory adjustment trajectory, parse the time points, warehouse nodes, product identifiers and quantity instructions contained in the inventory adjustment trajectory into an execution list, and generate an inventory control instruction set according to the preset interface protocol and field format; The inventory control instruction set is sent to the inventory execution system, which then performs replenishment, transfer, and release operations on the corresponding warehouse nodes at the corresponding time points. During the execution process, the actual execution status, actual execution quantity, and abnormal information of each inventory control instruction are recorded. The inventory execution results, which correspond one-to-one with the inventory control instruction set, are collected from the inventory execution system. The actual execution status, actual execution quantity, and abnormal information are compared with the execution list to generate summary data of execution results, including execution deviations, unexecuted instructions, and partial execution instruction identifiers. The inventory structure status data is updated based on the summary data of the execution results. Changes in warehouse coverage, supply path, allocation feasibility and inventory restriction status caused by inventory execution are written into the new inventory structure status vector, and the feedback factors used by the commodity dynamic graph of inventory structure perception are updated.

Images