Supply chain multi-level warehousing intelligent scheduling and collaboration method and system

Abstract

The present invention provides a supply chain multi-level warehousing intelligent scheduling and collaboration method and system, which relates to the field of intelligent warehousing technology. The method comprises the following steps: based on historical order data, a long short-term memory network is used to perform hierarchical modeling and prediction of the short-term and long-term demands of warehouses at all levels to obtain inventory demand; each warehouse node is set as an independent intelligent agent, and an inventory allocation strategy is generated through iterative optimization of strategies between intelligent agents; the inventory allocation strategy is converted into a scheduling instruction containing the allocation object, allocation quantity and allocation time based on a rule knowledge base; the scheduling instruction is received and sequenced for execution in the edge computing unit of each warehouse node; when an inventory anomaly or resource conflict is detected, an emergency collaboration request is initiated to the adjacent warehouse node; the warehouse node that receives the emergency collaboration request returns response information based on its own resource status, and based on the response information, an emergency treatment plan is determined through local negotiation between nodes.

Description

Technical Field

[0001] The present invention relates to intelligent warehousing technology, and in particular to a method and system for intelligent scheduling and collaboration of multi-level warehousing in a supply chain. Background Art

[0002] Traditional multi-tiered supply chain warehousing management relies primarily on manual experience and static rules for inventory allocation, making it difficult to adapt to complex and volatile market demands. This approach suffers from low demand forecasting accuracy, uneven inventory allocation, and inefficient collaboration, leading to increased overall supply chain operating costs and decreased customer satisfaction.

[0003] Existing technologies typically use centralized decision-making models to handle inventory management, resulting in high computational complexity and slow response times, making it difficult to respond to emergencies in real time. Furthermore, due to the lack of effective inter-node coordination mechanisms, each warehouse often operates independently, resulting in low overall resource utilization and the paradoxical coexistence of inventory backlogs and shortages.

[0004] With the development of the Internet of Things and artificial intelligence (AI), intelligent warehouse management has become possible. However, existing solutions still suffer from complex model training, disconnected decision-making and execution, and weak emergency response capabilities. In particular, the system lacks a flexible response mechanism in emergencies, making it difficult to achieve intelligent collaboration and optimal resource allocation across warehouse nodes, seriously impacting the stability and resilience of the supply chain. Summary of the Invention

[0005] In response to the deficiencies in the prior art, the present invention provides a supply chain multi-level warehousing intelligent scheduling and collaboration method and system, which can solve the problems in the prior art.

[0006] The present invention provides a supply chain multi-level warehousing intelligent scheduling and coordination method, including:

[0007] Obtain real-time inventory data and historical order data for warehouses at all levels in a multi-level warehousing network;

[0008] Based on the historical order data, a long short-term memory network is used to perform hierarchical modeling and prediction of short-term and long-term demand of warehouses at all levels to obtain inventory demand;

[0009] A multi-agent reinforcement learning model is constructed, with each warehouse node set as an independent agent. The state space of the multi-agent reinforcement learning model includes the real-time inventory data, inventory demand, and available resources of each warehouse node, and the action space includes the inventory allocation quantity and allocation timing. The inventory allocation strategy is generated through inter-agent strategy iteration and optimization.

[0010] Converting the inventory allocation strategy into a scheduling instruction including allocation objects, allocation quantities and allocation time based on a rule knowledge base;

[0011] In the edge computing unit of each warehouse node, the scheduling instructions are received and sorted for execution. When inventory anomalies or resource conflicts are detected, an emergency collaboration request is initiated to the adjacent warehouse node. The warehouse node that receives the emergency collaboration request returns response information based on its own resource status. Based on the response information, the emergency handling plan is determined through local consultation between nodes.

[0012] Optionally,

[0013] Based on the historical order data, a long short-term memory network is used to perform hierarchical modeling and prediction of short-term and long-term demand of warehouses at all levels. The steps of obtaining inventory demand include:

[0014] Dividing the historical order data into daily order sequences and weekly order sequences;

[0015] An adaptive hierarchical forecasting network is constructed, comprising a short-term forecasting layer and a long-term forecasting layer. The short-term forecasting layer receives daily order sequences and dynamically adjusts the forecast time window based on inventory categories and seasonal characteristics. The long-term forecasting layer receives weekly order sequences and captures long-term patterns using a memory enhancement module. The output error gradient of the short-term forecasting layer is passed to the parameter update process of the long-term forecasting layer, embedding the trend characteristics of the long-term forecasting layer into the input of the short-term forecasting layer.

[0016] The adaptive hierarchical prediction network is trained using a multi-scale loss function, wherein the multi-scale loss function includes a short-term prediction loss term, a long-term prediction loss term, and a regularization term, and the loss terms are dynamically weighted based on the time scale characteristics of each prediction layer;

[0017] The forecast results are integrated to obtain an initial forecast value; the initial forecast value is calibrated based on the storage capacity constraint, and combined with the forecast standard deviation to generate an inventory demand forecast result with a confidence interval.

[0018] Optionally,

[0019] A multi-agent reinforcement learning model is constructed, with each warehouse node set as an independent agent. The state space of the multi-agent reinforcement learning model includes the real-time inventory data, inventory demand, and available resources of each warehouse node, and the action space includes the inventory allocation quantity and allocation timing. The steps of generating an inventory allocation strategy through inter-agent strategy iterative optimization include:

[0020] Each warehouse node is set as an independent intelligent agent, and a state vector of the independent intelligent agent is constructed. The state vector includes real-time inventory data, inventory demand, and available resources. An adjacency matrix is ​​constructed based on the logistics channels between each warehouse node. The adjacency matrix is ​​used to represent the connection relationship between each warehouse node.

[0021] Constructing an action space for an independent agent, the action space including a transfer quantity matrix and a transfer timing vector, wherein the transfer quantity matrix is ​​used to represent the inventory transfer quantity between warehouse nodes, and the transfer timing vector is used to represent the transfer execution time;

[0022] Constructing a multi-objective reward function, which includes a local reward function and a global reward function. The local reward function calculates the reward value of each independent agent based on inventory balance, turnover efficiency, and scheduling cost, and the global reward function calculates the global collaborative reward value based on the inventory level difference of adjacent warehouse nodes;

[0023] Independent agents are trained using a proximal policy optimization algorithm. During the training process, a hierarchical experience buffer is constructed to store state transition samples. The hierarchical experience buffer includes a local buffer for storing the local experience of the independent agents and a global buffer for storing global state information. A parameter sharing mechanism is used to train the policy network for independent agents of the same type.

[0024] Based on the trained policy network, an inventory allocation strategy is generated according to the real-time status of each warehouse node.

[0025] Optionally,

[0026] The steps of constructing a multi-objective reward function, which includes a local reward function and a global reward function, wherein the local reward function calculates the reward value of each independent agent based on inventory balance, turnover efficiency, and scheduling cost, and the global reward function calculates the global collaborative reward value based on the inventory level difference of adjacent warehouse nodes include:

[0027] Items are classified based on historical order data for each warehouse node, and target inventory levels are determined. Based on the target inventory levels, the standardized inventory deviation for each warehouse node is calculated, and the product of the standardized inventory deviation and the time decay factor is determined as the inventory balance reward value.

[0028] The daily, weekly, and monthly turnover rates are calculated based on the outbound volume, beginning inventory, and ending inventory of each warehouse node, and the weighted sum of the results with the preset weight coefficient is used as the turnover efficiency bonus value.

[0029] Determine the dispatch cost bonus value based on the weighted sum of storage cost, transportation cost and operation cost;

[0030] Calculating the inventory level difference ratio between adjacent warehouses to construct a difference matrix, calculating the distance weight coefficient based on the logistics distance between adjacent warehouses, and determining the global collaboration reward value as the product of the distance weight coefficient, the difference matrix, the warehouse hierarchical collaboration coefficient, and the congestion penalty factor;

[0031] The performance improvement rates of the inventory balance reward value, turnover efficiency reward value, scheduling cost reward value, and global synergy reward value are calculated within a preset evaluation period. A multivariate linear regression model is constructed based on the performance improvement rates to calculate a correlation coefficient matrix. The contribution of each reward value is calculated according to the correlation coefficient matrix. The difference between the contribution and the current weight is multiplied by the learning rate to obtain a weight adjustment amount. The weight coefficient of each reward value is updated under the weight adjustment constraint to obtain a multi-objective reward function.

[0032] Optionally,

[0033] The proximal policy optimization algorithm is used to train the independent agent. The steps of constructing a hierarchical experience buffer to store state transition samples during the training process include:

[0034] The hierarchical experience buffer includes a local experience buffer for storing local state transition samples of each independent agent, a collaborative experience buffer for storing interaction samples of adjacent agents, and a global experience buffer for storing global state samples;

[0035] Calculate the importance score of the state transition sample, which is determined based on the reward value, state transition frequency, state impact range, and strategy difference, and store the state transition sample in the experience buffer of the corresponding level according to the importance score;

[0036] Calculating a sampling probability of a sample based on sample timeliness and the number of strategy updates, and selecting a training sample from the hierarchical experience buffer according to the sampling probability;

[0037] Update the policy network parameters of a single agent based on samples from the local experience buffer, update the policy network parameters of a group of adjacent agents based on samples from the collaborative experience buffer, and update the policy network parameters of all agents based on samples from the global experience buffer;

[0038] Constructing a knowledge sharing weight matrix, and calculating element values ​​of the knowledge sharing weight matrix based on the historical decision data of the intelligent agent and the network topology structure;

[0039] The objective function of the near-end policy optimization algorithm is divided into immediate optimization items, periodic optimization items and long-term optimization items. The weight coefficient of each optimization item is calculated based on the knowledge sharing weight matrix and system state information, and the policy network parameters are optimized and updated.

[0040] Optionally,

[0041] The steps of converting the inventory allocation strategy into a scheduling instruction including an allocation object, an allocation quantity, and an allocation time based on a rule knowledge base include:

[0042] Constructing a rule knowledge base including physical constraint rules, business constraint rules, and priority rules, and parsing the inventory allocation strategy into an initial scheduling plan including warehouse node pairs, allocation quantities, and allocation time windows;

[0043] Based on the rule knowledge base, the initial scheduling plan is constraint verified and a rule matching score function is constructed. The rule matching score function calculates a matching score based on the degree of physical constraint satisfaction, business goal achievement and resource utilization, and automatically corrects scheduling plans that fail constraint verification or have a matching score below a threshold;

[0044] Based on the scheduling delay rate, resource conflict rate and inventory fluctuation rate in the historical scheduling execution data, the execution risk probability of the revised scheduling plan in different time windows is calculated, and the execution priority considering risk aversion is generated. The sorted scheduling plan is converted into a scheduling instruction containing the allocation object, allocation quantity and allocation time.

[0045] Optionally,

[0046] The warehouse node that receives the emergency coordination request returns a response message based on its own resource status. The steps of determining an emergency handling plan through local negotiation between nodes based on the response message include:

[0047] Calculate the logistics relevance based on the historical logistics frequency and current logistics intensity of the warehouse node, calculate the resource response capability based on the ratio of the requested resources to the available resources, and generate response information including response time, response cost and resource status based on the logistics relevance and resource response capability;

[0048] Obtain the task completion rate and resource stability of the warehouse nodes participating in the negotiation, calculate the node credibility based on the response information, generate an influence weight based on the number of hops between nodes, and the influence weight decreases exponentially with the increase of the hop number. Determine the initial negotiation weight based on the node credibility and influence weight;

[0049] Receive the negotiation solutions of neighboring warehouse nodes, calculate the weighted average of the solutions based on the initial negotiation weights, combine the weighted average with the current optimal solution to form an updated local solution, and calculate the variance of the updated local solution within the negotiation group as a consistency indicator;

[0050] When the update amplitude of the plan is continuously lower than the preset amplitude threshold, the cost, time and resource consumption weights in the negotiation plan are adjusted according to the historical execution results, and a random perturbation value positively correlated with the duration of the stagnation is generated to restart the negotiation; when the consistency index meets the preset conditions, an emergency treatment plan is generated;

[0051] Monitor the heartbeat information and response delay of the warehouse nodes participating in the negotiation. When it is detected that the heartbeat information is interrupted or the response delay exceeds the preset value, obtain the unfinished negotiation tasks of the heartbeat abnormal node, calculate the task reception fitness based on the resource status, response time and logistics distance of the remaining warehouse nodes, and reallocate the unfinished negotiation tasks based on the task reception fitness while satisfying the resource constraints.

[0052] A second aspect of an embodiment of the present invention provides a supply chain multi-level warehousing intelligent scheduling and collaboration system, including:

[0053] The first unit is used to obtain real-time inventory data and historical order data of warehouses at all levels in the multi-level warehousing network;

[0054] The second unit is configured to perform hierarchical modeling and prediction of short-term and long-term demands of warehouses at all levels using a long short-term memory network based on the historical order data to obtain inventory demand;

[0055] The third unit is used to build a multi-agent reinforcement learning model, setting each warehouse node as an independent agent. The state space of the multi-agent reinforcement learning model includes the real-time inventory data, inventory demand, and available resources of each warehouse node. The action space includes the inventory allocation quantity and allocation timing. The inventory allocation strategy is generated through inter-agent policy iterative optimization.

[0056] The fourth unit is configured to convert the inventory allocation strategy into a scheduling instruction including an allocation object, an allocation quantity, and an allocation time based on a rule knowledge base;

[0057] The fifth unit is used to receive the scheduling instructions and sort them for execution in the edge computing unit of each warehouse node. When inventory anomalies or resource conflicts are detected, an emergency collaboration request is initiated to the adjacent warehouse node. The warehouse node that receives the emergency collaboration request returns response information based on its own resource status. Based on the response information, the emergency handling plan is determined through local consultation between nodes.

[0058] According to a third aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0059] The present invention realizes hierarchical demand forecasting through long-short-term memory networks, significantly improving the accuracy of demand forecasting and effectively reducing inventory fluctuations and out-of-stock risks; the multi-agent reinforcement learning model enables each warehouse node to autonomously learn the optimal allocation strategy, thereby improving inventory balance, increasing turnover efficiency and reducing operating costs, and optimizing the overall supply chain resource allocation efficiency.

[0060] The proposed policy transformation mechanism, based on a rule knowledge base, ensures the enforceability and business compliance of decisions, significantly reducing the rate of execution anomalies. The edge computing architecture localizes decision-making, significantly shortens response time, improves system throughput, reduces computing pressure and communication overhead on central nodes, and makes the system more efficient and scalable.

[0061] The distributed emergency coordination mechanism of the present invention enables the system to have self-healing capabilities when facing abnormal situations. Through local consultation between nodes, emergency response plans are quickly formed, which shortens the abnormal processing time and improves the abnormal resolution rate. It significantly enhances the resilience and stability of the supply chain system, effectively responds to market fluctuations and supply interruption risks, and provides enterprises with more reliable supply chain guarantees. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 Schematic diagram of the process of the multi-level warehousing intelligent scheduling and collaboration method for supply chain according to an embodiment of the present invention;

[0063] Figure 2 This is a diagram showing the impact of gradual iteration of various technical components of the present invention on system performance. DETAILED DESCRIPTION

[0064] The technical solutions in the embodiments of the present invention will be described below in conjunction with the drawings in the embodiments of the present invention. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0065] Figure 1 This is a flow chart of the multi-level warehousing intelligent scheduling and collaboration method for the supply chain of the present invention, as shown in FIG. Figure 1 As shown, the method includes:

[0066] Obtain real-time inventory data and historical order data for warehouses at all levels in a multi-level warehousing network;

[0067] Based on the historical order data, a long short-term memory network is used to perform hierarchical modeling and prediction of short-term and long-term demand of warehouses at all levels to obtain inventory demand;

[0068] A multi-agent reinforcement learning model is constructed, with each warehouse node set as an independent agent. The state space of the multi-agent reinforcement learning model includes the real-time inventory data, inventory demand, and available resources of each warehouse node, and the action space includes the inventory allocation quantity and allocation timing. The inventory allocation strategy is generated through inter-agent strategy iteration and optimization.

[0069] Converting the inventory allocation strategy into a scheduling instruction including allocation objects, allocation quantities and allocation time based on a rule knowledge base;

[0070] In the edge computing unit of each warehouse node, the scheduling instructions are received and sorted for execution. When inventory anomalies or resource conflicts are detected, an emergency collaboration request is initiated to the adjacent warehouse node. The warehouse node that receives the emergency collaboration request returns response information based on its own resource status. Based on the response information, the emergency handling plan is determined through local consultation between nodes.

[0071] Optionally,

[0072] Based on the historical order data, a long short-term memory network is used to perform hierarchical modeling and prediction of short-term and long-term demand of warehouses at all levels. The steps of obtaining inventory demand include:

[0073] Dividing the historical order data into daily order sequences and weekly order sequences;

[0074] An adaptive hierarchical forecasting network is constructed, comprising a short-term forecasting layer and a long-term forecasting layer. The short-term forecasting layer receives daily order sequences and dynamically adjusts the forecast time window based on inventory categories and seasonal characteristics. The long-term forecasting layer receives weekly order sequences and captures long-term patterns using a memory enhancement module. The output error gradient of the short-term forecasting layer is passed to the parameter update process of the long-term forecasting layer, embedding the trend characteristics of the long-term forecasting layer into the input of the short-term forecasting layer.

[0075] The adaptive hierarchical prediction network is trained using a multi-scale loss function, wherein the multi-scale loss function includes a short-term prediction loss term, a long-term prediction loss term, and a regularization term, and the loss terms are dynamically weighted based on the time scale characteristics of each prediction layer;

[0076] The forecast results are integrated to obtain an initial forecast value; the initial forecast value is calibrated based on the storage capacity constraint, and combined with the forecast standard deviation to generate an inventory demand forecast result with a confidence interval.

[0077] In this embodiment, historical order data from warehouses at each level of a multi-level warehousing network is obtained. This data typically includes fields such as order date, order quantity, product category, and customer information. The obtained historical order data is preprocessed, including data cleaning, outlier processing, and data normalization. During the data cleaning process, missing values ​​are detected and processed. For example, missing order data for a particular day can be filled using the average of the adjacent days. Outlier processing identifies and adjusts data points that significantly deviate from the normal range, such as sudden order surges caused by promotional events or holidays. Data normalization converts data of different magnitudes to the same scale. The preprocessed historical order data is divided into daily and weekly order series. The daily order series records the daily order quantity and is primarily used for short-term demand forecasting; the weekly order series summarizes the weekly order quantity and is used for long-term trend forecasting. For example, for the food category in a warehouse, the daily series records the daily order volume, such as "Date 1: 150 pieces, Date 2: 165 pieces...", while the weekly series records "Week 1: 1050 pieces, Week 2: 1120 pieces..."

[0078] An adaptive hierarchical forecasting network is constructed, consisting of short-term and long-term forecasting layers, with an information exchange mechanism between the two layers. The short-term forecasting layer utilizes a long short-term memory (LSTM) network with an attention mechanism, primarily processing daily order series to capture short-term fluctuations. The long-term forecasting layer utilizes a bidirectional LSTM network with a memory-enhanced module to process weekly order series to capture long-term trends. The short-term forecasting layer dynamically adjusts the forecast window based on inventory category and seasonality. For example, for clothing products with strong seasonality, the forecast window is set to 7-14 days; for daily necessities with stable demand, the forecast window is set to 3-7 days. In the implementation, volatility indicators such as the coefficient of variation (CV) of historical data for each product category are analyzed, and the appropriate forecast window size is dynamically determined based on these indicators. A smaller window yields higher forecast accuracy but a shorter timeframe. A larger window allows for predictions further into the future, but with reduced accuracy.

[0079] The memory-enhanced module is a structure consisting of multiple memory cells, each of which stores historical pattern information over a specific time span. For example, for products with significant seasonality, a memory cell will focus specifically on sales patterns from the same period last year. For products affected by economic cycles, a memory cell will capture trends over a longer time span. This design enables the model to "remember" and leverage long-term historical patterns, improving its ability to predict future long-term trends.

[0080] A bidirectional information flow mechanism is established between the two prediction layers. The output error gradient of the short-term prediction layer is passed to the parameter update process of the long-term prediction layer, allowing the long-term prediction to be adjusted based on the results of the short-term prediction. Simultaneously, the trend characteristics of the long-term prediction layer are embedded in the input of the short-term prediction layer, allowing the short-term prediction to be fine-tuned based on the long-term trend. This design enables the two prediction layers to complement and correct each other, resulting in more accurate prediction results.

[0081] During the network training phase, a multi-scale loss function is used to train the adaptive hierarchical forecasting network. This loss function includes a short-term forecast loss term, a long-term forecast loss term, and a regularization term. The short-term forecast loss term primarily measures the difference between the daily forecast result and the actual value, the long-term forecast loss term measures the difference between the weekly forecast result and the actual value, and the regularization term is used to prevent model overfitting. The loss terms are dynamically weighted based on the timescale characteristics of each forecast layer. For example, the short-term forecast loss term is weighted more heavily for products with high volatility, while the long-term forecast loss term is weighted more heavily for products with significant trends.

[0082] After training is complete, the forecast results from the short-term and long-term forecast layers are integrated to produce an initial forecast. This integration method uses a weighted average approach, with weights dynamically adjusted based on historical forecast accuracy. For example, if the short-term forecast layer performs better in the recent period, its weight is increased accordingly; conversely, if the long-term forecast layer captures important seasonal trends, its weight is increased. This dynamic adjustment mechanism ensures that the integrated results balance short-term accuracy with long-term trends. Finally, the initial forecast is calibrated based on storage capacity constraints. If the forecasted demand exceeds the warehouse's maximum storage capacity, the forecast is adjusted based on historical overflow handling experience to ensure that the forecast result is feasible in practice. Furthermore, the forecast standard deviation is calculated based on the uncertainty in the forecast process, generating an inventory demand forecast with a confidence interval. For example, a result such as "Forecast demand for the next 7 days is 1000 ± 50 units" is generated, where ±50 units represents the uncertainty range of the forecast, helping inventory managers formulate more reasonable inventory strategies.

[0083] This method achieves accurate forecasting of multi-level warehousing demand by constructing an adaptive hierarchical forecasting network, offering significant advantages over traditional methods. The bidirectional information exchange mechanism between the short-term and long-term forecast layers effectively integrates characteristics of different time scales. The dynamically adjusted forecast time window adapts to the demand characteristics of different commodities, and the multi-scale loss function training method improves the model's generalization capabilities. The forecast calibration and confidence interval generation based on warehousing constraints make the forecast results more practical and reliable, providing accurate data support for inventory management decisions, effectively reducing inventory costs, and improving supply chain responsiveness and customer satisfaction.

[0084] Optionally,

[0085] A multi-agent reinforcement learning model is constructed, with each warehouse node set as an independent agent. The state space of the multi-agent reinforcement learning model includes the real-time inventory data, inventory demand, and available resources of each warehouse node, and the action space includes the inventory allocation quantity and allocation timing. The steps of generating an inventory allocation strategy through inter-agent strategy iterative optimization include:

[0086] Each warehouse node is set as an independent intelligent agent, and a state vector of the independent intelligent agent is constructed. The state vector includes real-time inventory data, inventory demand, and available resources. An adjacency matrix is ​​constructed based on the logistics channels between each warehouse node. The adjacency matrix is ​​used to represent the connection relationship between each warehouse node.

[0087] Constructing an action space for an independent agent, the action space including a transfer quantity matrix and a transfer timing vector, wherein the transfer quantity matrix is ​​used to represent the inventory transfer quantity between warehouse nodes, and the transfer timing vector is used to represent the transfer execution time;

[0088] Constructing a multi-objective reward function, which includes a local reward function and a global reward function. The local reward function calculates the reward value of each independent agent based on inventory balance, turnover efficiency, and scheduling cost, and the global reward function calculates the global collaborative reward value based on the inventory level difference of adjacent warehouse nodes;

[0089] Independent agents are trained using a proximal policy optimization algorithm. During the training process, a hierarchical experience buffer is constructed to store state transition samples. The hierarchical experience buffer includes a local buffer for storing the local experience of the independent agents and a global buffer for storing global state information. A parameter sharing mechanism is used to train the policy network for independent agents of the same type.

[0090] Based on the trained policy network, an inventory allocation strategy is generated according to the real-time status of each warehouse node.

[0091] In this embodiment, each warehouse node in a multi-level warehousing network is configured as an independent agent. In practical applications, for example, a large e-commerce company's supply chain network consists of five central warehouses, 25 regional warehouses, and 100 forward warehouses. Each warehouse node is configured as an independent agent. A state vector is constructed for each agent, which includes three key pieces of information: real-time inventory data, inventory demand, and available resources. Real-time inventory data records the current inventory level of each commodity, for example, "Commodity A: 500 pieces, Commodity B: 320 pieces." Inventory demand is the demand forecast for different time periods in the future, obtained using the aforementioned forecasting method, for example, "Commodity A demand in the next three days: 150 pieces, Commodity A demand in the next seven days: 380 pieces." Available resources include the remaining warehouse storage space, the number of dispatchable vehicles, and the number of available loading and unloading workers, for example, "Remaining storage space: 2000 cubic meters, dispatchable vehicles: 15 vehicles."

[0092] An adjacency matrix is ​​constructed based on the logistics channels between warehouse nodes. If a direct logistics channel exists between two warehouse nodes, the corresponding matrix element is 1; otherwise, it is 0. For example, a 130×130 adjacency matrix (corresponding to 130 warehouse nodes) is constructed, in which the central warehouse is connected to all regional warehouses, which in turn are connected to forward warehouses within their coverage area. Forward warehouses in different regions are generally not directly connected to each other. This adjacency matrix plays three key roles in the subsequent agent training and decision-making process: it limits the communication range between agents, ensuring that each agent can only exchange information with directly connected nodes; it defines the boundary conditions under which deployment operations can be executed, ensuring that agents can only initiate deployments to adjacent nodes; and it serves as a weighting factor in the calculation of the global reward function, giving greater weight to the inventory balance between adjacent nodes. This constraint mechanism, based on the topology of the actual logistics network, ensures that the generated deployment strategy is physically feasible.

[0093] The dispatch quantity matrix describes the quantities of various items dispatched from one warehouse to other adjacent warehouses. For example, the dispatch quantity matrix for a regional warehouse might include information such as "Distribute 50 pieces of item A to forward warehouse 1, and 30 pieces of item B to forward warehouse 2." The dispatch timing vector describes the time points at which these dispatch operations will be executed, such as "Distribute 1 execution time: T+1 day, Distribute 2 execution time: T+3 days." To discretize the action space, the dispatch quantities are divided into several percentage levels (e.g., 0%, 20%, 40%, 60%, 80%, and 100%), and the dispatch timings are divided into several time periods (e.g., immediately, within 1 day, within 3 days, and within 7 days). This design makes the action space manageable while retaining sufficient flexibility.

[0094] A multi-objective reward function is constructed, including local and global reward functions. Independent agents are trained using a proximal policy optimization algorithm. The core concept of this algorithm is to ensure training stability by limiting the policy update step size. To improve training efficiency and generalization performance, a parameter sharing mechanism is used to train the policy network for independent agents of the same type. Specifically, warehouse nodes at the same level (e.g., all regional warehouses) share policy network parameters. This allows the experience of each agent to be used to update the shared policy, significantly increasing the training sample size. For example, in one actual deployment, 25 regional warehouses shared a set of policy network parameters, allowing the experience of each regional warehouse to be leveraged by other regional warehouses, accelerating training convergence and improving the robustness of the policy.

[0095] During training, a phased training strategy is employed. First, each agent performs local training based on samples from its local buffer, learning basic inventory allocation strategies. Then, collaborative training is performed based on samples from the global buffer, learning allocation strategies that consider overall efficiency. In actual training, pre-training is performed using historical data, followed by fine-tuning in a simulated environment. Finally, small-scale testing is conducted in the real environment, with the application scope gradually expanded.

[0096] Real-time inventory data, demand forecasts, and resource status from each warehouse node are fed into a trained policy network. The network outputs a probability distribution of allocation quantities and allocation timings, selecting the action with the highest probability as the final allocation decision. For example, the policy network might issue the decision "Allocate 200 units of item X from the central warehouse to regional warehouse A, execution time: 8:00 AM tomorrow." These decisions are combined to form the inventory allocation policy for the entire warehousing network, guiding actual inventory scheduling.

[0097] The present invention realizes intelligent inventory allocation in warehousing networks by constructing a multi-agent reinforcement learning model. Each warehouse is set as an independent agent to have distributed decision-making capabilities. The design of the multi-objective reward function takes into account both local optimization and global coordination. The proximal policy optimization algorithm and the hierarchical experience buffer improve training efficiency and policy quality. The parameter sharing mechanism enhances the generalization ability of the model, making it adaptable to different warehouse scenarios. Through the iterative optimization of policies among agents, the optimal allocation strategy can be generated to balance inventory distribution, improve turnover efficiency, and reduce operating costs, providing an effective solution for the intelligent management of multi-level warehousing networks.

[0098] Optionally,

[0099] The steps of constructing a multi-objective reward function, which includes a local reward function and a global reward function, wherein the local reward function calculates the reward value of each independent agent based on inventory balance, turnover efficiency, and scheduling cost, and the global reward function calculates the global collaborative reward value based on the inventory level difference of adjacent warehouse nodes include:

[0100] Items are classified based on historical order data for each warehouse node, and target inventory levels are determined. Based on the target inventory levels, the standardized inventory deviation for each warehouse node is calculated, and the product of the standardized inventory deviation and the time decay factor is determined as the inventory balance reward value.

[0101] The daily, weekly, and monthly turnover rates are calculated based on the outbound volume, beginning inventory, and ending inventory of each warehouse node, and the weighted sum of the results with the preset weight coefficient is used as the turnover efficiency bonus value.

[0102] Determine the dispatch cost bonus value based on the weighted sum of storage cost, transportation cost and operation cost;

[0103] Calculating the inventory level difference ratio between adjacent warehouses to construct a difference matrix, calculating the distance weight coefficient based on the logistics distance between adjacent warehouses, and determining the global collaboration reward value as the product of the distance weight coefficient, the difference matrix, the warehouse hierarchical collaboration coefficient, and the congestion penalty factor;

[0104] The performance improvement rates of the inventory balance reward value, turnover efficiency reward value, scheduling cost reward value, and global synergy reward value are calculated within a preset evaluation period. A multivariate linear regression model is constructed based on the performance improvement rates to calculate a correlation coefficient matrix. The contribution of each reward value is calculated according to the correlation coefficient matrix. The difference between the contribution and the current weight is multiplied by the learning rate to obtain a weight adjustment amount. The weight coefficient of each reward value is updated under the weight adjustment constraint to obtain a multi-objective reward function.

[0105] In this embodiment, items are classified based on the historical order data of each warehouse node and the target inventory level is determined. Item classification is usually carried out according to sales speed and demand volatility. For example, goods are divided into Class A (high-frequency fast-moving consumer goods), Class B (medium-frequency conventional goods) and Class C (low-frequency long-tail goods). For a regional warehouse, shampoo with an average daily sales volume of more than 100 pieces and a fluctuation coefficient of less than 0.3 is classified as Class A goods, skin care products with an average daily sales volume of between 20-100 pieces and a fluctuation coefficient of between 0.3-0.6 are classified as Class B goods, and special seasonal goods with an average daily sales volume of less than 20 pieces or a fluctuation coefficient greater than 0.6 are classified as Class C goods. Different methods are used to determine the target inventory level for different categories of goods. Category A goods usually adopt the method of demand forecast value plus safety stock. For example, the 7-day demand forecast of a Category A product is 700 pieces, the fluctuation standard deviation is 50 pieces, and its target inventory is set at 850 pieces (forecast demand plus 3 times the standard deviation); Category B goods adopt the method of historical average consumption plus buffer stock; Category C goods adopt the method of combining minimum order quantity and periodic replenishment.

[0106] Based on the target inventory level, the standardized inventory deviation for each warehouse node is calculated. For each product, the target inventory level is subtracted from the actual inventory level, and then divided by the target inventory level to obtain the relative deviation value. For example, if the actual inventory of Category A product in a warehouse is 900 pieces and the target inventory level is 850 pieces, the relative deviation is (900 - 850) / 850 = 0.059, indicating a slight overstock. The absolute value of the deviations for all products is taken and weighted by sales volume to obtain the warehouse's overall standardized inventory deviation. The smaller the deviation, the closer the actual inventory level is to the target level, indicating more accurate inventory management. The standardized inventory deviation is multiplied by the time decay factor to determine the inventory balance reward. The time decay factor is a value that decreases over time, for example, it can be set to e raised to the power of negative t (where t is the number of days after the inventory adjustment). This design encourages inventory balance to be reached quickly and maintained over time. If inventory balance improves immediately after an adjustment but then relapses after a few days, the reward value will be reduced due to the time decay factor. The final inventory balance reward value is equal to 1 minus the normalized inventory deviation after decay, so that the smaller the deviation, the greater the reward.

[0107] The turnover rate is calculated by dividing the shipment volume during a specific period by the average inventory level. The average inventory level is equal to the beginning inventory level plus the ending inventory level, divided by 2. For example, if a warehouse ships 600 items per day, has an beginning inventory of 1,200 items, and an ending inventory of 1,000 items, the daily turnover rate is 600 / ((1,200 + 1,000) / 2) = 0.55, indicating that 55% of the inventory is turned over on average each day. Weekly and monthly turnover rates are calculated using the same principles, but with different statistical periods. Turnover rates for different periods reflect inventory flow efficiency over different timescales and, when combined, provide a comprehensive assessment of inventory turnover. For example, weights could be set to 0.2 for daily turnover, 0.3 for weekly turnover, and 0.5 for monthly turnover, reflecting a greater emphasis on long-term turnover performance. If a warehouse's daily, weekly, and monthly turnover rates are 0.55, 3.5, and 14.0, respectively, then the weighted turnover efficiency is 0.55 × 0.2 + 3.5 × 0.3 + 14.0 × 0.5 = 8.15. This value is normalized (for example, divided by the expected maximum value) to obtain a turnover efficiency bonus value between 0 and 1.

[0108] The dispatch cost bonus is determined based on the weighted sum of storage costs, transportation costs, and operating costs. Storage costs include warehouse rent, equipment depreciation, energy consumption, and other expenses allocated to each commodity; transportation costs include vehicle usage fees, fuel costs, labor costs, etc.; and operating costs include operational costs such as picking, packaging, and loading and unloading. For example, a dispatch operation involving 500 items incurs a storage cost of 1,000 yuan, a transportation cost of 3,000 yuan, and an operating cost of 2,000 yuan, for a total cost of 6,000 yuan. This cost is compared with a benchmark cost (such as a historical average cost), and the cost reduction ratio is calculated. This ratio is then converted into a dispatch cost bonus value between 0 and 1. The lower the cost, the higher the bonus value.

[0109] Calculate the inventory level difference ratios between adjacent warehouses to construct a difference matrix. For each pair of adjacent warehouses, calculate the inventory level difference ratio, which is the absolute value of the difference between the two warehouses' standardized inventory levels (actual inventory divided by target inventory). For example, if Warehouse A's standardized inventory level is 1.1 (slightly overstocked), and Warehouse B's is 0.85 (slightly understocked), then the difference ratio between them is |1.1 - 0.85| = 0.25. The difference ratios of all pairs of adjacent warehouses are combined to form a difference matrix. A distance weight coefficient is calculated based on the logistics distance between adjacent warehouses. The closer the distance between warehouse pairs, the more important their inventory balancing coordination is, and the larger the weight coefficient. For example, the weight coefficient can be set as a baseline value (e.g., 1.0) divided by the standardized distance, where the standardized distance is the actual distance divided by the average warehouse distance in the network. For example, if the distance between two warehouses is 50 kilometers and the average network distance is 100 kilometers, then the standardized distance is 0.5, and the weight coefficient is 1.0 / 0.5 = 2.0. The global collaboration reward is determined by multiplying the distance weight coefficient, the difference matrix, the warehouse-level collaboration coefficient, and the congestion penalty factor. The warehouse-level collaboration coefficient reflects the importance of collaboration between warehouses at different levels. For example, the collaboration coefficient between a central warehouse and regional warehouses is higher than the collaboration coefficient between a regional warehouse and a forward warehouse. The congestion penalty factor reduces the reward when the logistics channel is congested, avoiding the generation of scheduling strategies that exacerbate congestion. The final global collaboration reward is the weighted sum of the collaboration rewards for all adjacent warehouse pairs, normalized to a value between 0 and 1.

[0110] Within a preset evaluation period (e.g., every 30 days), calculate the performance improvement rate for inventory balance rewards, turnover efficiency rewards, scheduling cost rewards, and global synergy rewards. The performance improvement rate is the percentage change in the average reward value during the current evaluation period compared to the previous evaluation period. For example, if the average inventory balance reward during the current period is 0.75 and the previous period was 0.7, the improvement rate is (0.75 - 0.7) / 0.7 = 7.14%. Based on these performance improvement rates, a multivariate linear regression model is constructed, with overall performance improvement as the dependent variable and the improvement rate of each individual reward as the independent variable. Regression analysis is used to calculate the correlation coefficient matrix. The correlation coefficient reflects the contribution of each reward to overall performance. Based on the correlation coefficient matrix, the contribution of each reward is calculated as the absolute value of each coefficient divided by the sum of the absolute values ​​of all coefficients. The weight adjustment is calculated by multiplying the difference between the contribution and the current weight by the learning rate. For example, if the contribution of the inventory balance reward is 0.4, the current weight is 0.3, and the learning rate is set to 0.2, the adjustment amount is (0.4-0.3)×0.2=0.02. Under the weight adjustment constraints (such as the sum of all weights is 1 and the individual weight is no less than 0.1), the weight coefficients of each reward value are updated to form an adaptive multi-objective reward function.

[0111] This invention effectively guides multi-agent warehouse scheduling by constructing a multi-objective reward function that includes local and global rewards. Inventory balance rewards ensure that inventory levels approach the optimal target, turnover efficiency rewards promote efficient resource flow, scheduling cost rewards reduce operating costs, and global collaboration rewards optimize overall network collaboration. An adaptive weight adjustment mechanism based on performance improvement rate enables the reward function to be dynamically optimized based on actual results, better adapting to different warehouse environments and business stages. This multi-dimensional, adaptive reward mechanism provides agents with a clear learning direction, effectively improving inventory management quality and overall supply chain efficiency.

[0112] Optionally,

[0113] The proximal policy optimization algorithm is used to train the independent agent. The steps of constructing a hierarchical experience buffer to store state transition samples during the training process include:

[0114] The hierarchical experience buffer includes a local experience buffer for storing local state transition samples of each independent agent, a collaborative experience buffer for storing interaction samples of adjacent agents, and a global experience buffer for storing global state samples;

[0115] Calculate the importance score of the state transition sample, which is determined based on the reward value, state transition frequency, state impact range, and strategy difference, and store the state transition sample in the experience buffer of the corresponding level according to the importance score;

[0116] Calculating a sampling probability of a sample based on sample timeliness and the number of strategy updates, and selecting a training sample from the hierarchical experience buffer according to the sampling probability;

[0117] Update the policy network parameters of a single agent based on samples from the local experience buffer, update the policy network parameters of a group of adjacent agents based on samples from the collaborative experience buffer, and update the policy network parameters of all agents based on samples from the global experience buffer;

[0118] Constructing a knowledge sharing weight matrix, and calculating element values ​​of the knowledge sharing weight matrix based on the historical decision data of the intelligent agent and the network topology structure;

[0119] The objective function of the near-end policy optimization algorithm is divided into immediate optimization items, periodic optimization items and long-term optimization items. The weight coefficient of each optimization item is calculated based on the knowledge sharing weight matrix and system state information, and the policy network parameters are optimized and updated.

[0120] In this embodiment, the hierarchical experience buffer consists of three levels of buffers: a local experience buffer, a collaborative experience buffer, and a global experience buffer. The local experience buffer is used to store local state transition samples for each independent agent, primarily including the agent's state, action, reward, and next-state information. For example, for a regional warehouse agent, its local experience buffer stores samples such as "State: Current inventory level 75%, forecasted demand 120 pieces / day, 8 vehicles available; Action: Dispatch 30 pieces of item X to forward warehouse A; Reward: Local reward value 0.65; Next state: Inventory level 70%, forecasted demand 118 pieces / day, 7 vehicles available."

[0121] The collaborative experience buffer stores interaction samples between adjacent agents, primarily including collaborative states, joint actions, and shared rewards involving multiple adjacent agents. For example, a collaborative sample between a regional warehouse and its three covered forward warehouses might include information such as "Collaborative state: Regional warehouse inventory is 20% overstock, forward warehouse A is 15% out of stock, forward warehouse B is 10% out of stock, and forward warehouse C has normal inventory; Joint action: Regional warehouse allocates 15% inventory to warehouse A and 10% to warehouse B; Shared reward: Global collaborative reward of 0.75."

[0122] The global experience buffer stores global state samples for the entire warehouse network, primarily including state information, global performance metrics, and reward data. For example, a global sample might include information such as "global state: total inventory level 85%, inter-regional inventory variance coefficient 0.15, network congestion 0.25; performance: order fill rate 92%, average delivery time 1.8 days; reward: global reward value 0.82." This hierarchical design enables learning optimization strategies at multiple levels, from local to global.

[0123] An importance score is calculated for each state transition sample to determine which buffer layer to store the sample in. The importance score is calculated based on four key metrics: reward value, state transition frequency, state impact range, and policy divergence. Reward value refers to the size of the reward received by the sample; higher rewards indicate more important samples. State transition frequency refers to how often this type of state transition has occurred historically; lower frequencies indicate rarer and therefore more important states. State impact range refers to the number of agents affected by the state transition; the wider the impact range, the higher the importance. Policy divergence refers to the degree to which the sample's action differs from the action generated by the current policy; the greater the difference, the more new information the sample contains, and therefore, the more important it is. The importance score is calculated by weighting these four metrics and summing them together. For example, if a sample has a reward value of 0.9 (normalized), a state transition frequency of 0.1 (low frequency), a state influence range of 0.7 (affecting multiple agents), a policy divergence of 0.6 (significantly different from the current policy), and the weights of the indicators are 0.3, 0.2, 0.3, and 0.2, respectively, then the sample's importance score is 0.9 × 0.3 + 0.1 × 0.2 + 0.7 × 0.3 + 0.6 × 0.2 = 0.67. Samples are stored in the corresponding experience buffer based on their importance scores. Typically, two thresholds, T1 and T2, are set (e.g., T1 = 0.3 and T2 = 0.7). Samples with importance scores below T1 are stored in the local experience buffer, samples between T1 and T2 are stored in the collaborative experience buffer, and samples with scores above T2 are stored in the global experience buffer. This way, samples with greater global significance are stored in higher-level buffers, allowing them to be learned by a wider range of agents.

[0124] During the training process, sample timeliness refers to the freshness of the sample, which is usually expressed as the inverse of the sample age (that is, the number of training rounds after the sample is stored in the buffer). For example, the timeliness of the sample stored 10 rounds ago is 1 / 10=0.1, while the timeliness of the most recently stored sample is 1. The number of policy updates refers to the number of times the policy network parameters have been updated since the sample was generated. The more updates, the worse the representativeness of the sample under the current policy, and the lower the sampling probability should be. The function of sample timeliness and the number of policy updates is used as the sampling probability. For example, the sampling probability can be set equal to timeliness multiplied by e to the power of negative λ times the number of policy updates, where λ is a coefficient that controls the decay rate (such as λ=0.05). For a sample with a timeliness of 0.5 and that has undergone 10 policy updates, its sampling probability is 0.5×e (-0.05×10) =0.5×0.61=0.31. Training samples are selected from the stratified experience buffer according to the calculated sampling probability, ensuring that new and representative samples have a higher chance of being selected.

[0125] The policy network parameters of the agents within the corresponding scope are updated based on samples from buffers at different levels. Based on samples from the local experience buffer, only the policy network parameters of a single agent are updated; based on samples from the collaborative experience buffer, the policy network parameters of a cluster of adjacent agents (i.e., a group of interconnected agents) are updated; and based on samples from the global experience buffer, the policy network parameters of all agents are updated. For example, in a training iteration, samples are selected from the local buffer of regional warehouse A to update only the policy network of the agent in that warehouse; samples are selected from the collaborative buffer that includes regional warehouse A and the cluster of preceding warehouses it covers to update the policy networks of all agents in this cluster; and samples are selected from the global buffer to update the policy networks of all agents in the entire network. This hierarchical update mechanism ensures both local optimization and global coordination.

[0126] A knowledge-sharing weight matrix is ​​constructed, and its elements are calculated based on the agents' historical decision data and network topology. The knowledge-sharing weight matrix is ​​an N×N matrix (N is the number of agents), and the matrix element (i, j) represents the weight of the knowledge transferred from agent i to agent j. The weight is determined based on two factors: historical decision similarity and network topological distance. Historical decision similarity is calculated by comparing the decisions made by two agents under similar conditions. The more similar the decisions, the higher the similarity. For example, if two regional warehouses both tend to allocate inventory to the forward warehouse when faced with excess inventory, their decision similarity is high; if one tends to allocate while the other tends to maintain inventory, the similarity is low. Network topological distance refers to the number of hops connecting two agents in the warehouse network. The fewer hops, the closer the distance, and the higher the weight.

[0127] The function of historical decision similarity and network topological distance is used as the weight value. For example, the weight value can be set equal to the decision similarity multiplied by e minus d times the topological distance, where d is the coefficient that controls the influence of distance (such as d = 0.5). For two agents with a decision similarity of 0.8 and a topological distance of 2, the weight value is 0.8×e (-0.5×2) =0.8×0.37=0.30. This design enables stronger knowledge sharing between agents with similar decision-making patterns and close network positions.

[0128] The objective function of the proximal policy optimization algorithm is divided into three optimization terms: immediate optimization, periodic optimization, and long-term optimization. The immediate optimization term focuses on maximizing the immediate reward at the current time step, the periodic optimization term focuses on medium-term cumulative rewards, and the long-term optimization term focuses on long-term stability and performance. Each agent's knowledge influence index is extracted from the knowledge sharing weight matrix. This is calculated by summing and normalizing the values ​​of the elements in the matrix corresponding to that agent. Agents with higher influence receive a higher weight for the long-term optimization term. Current state characteristics are also analyzed, including inventory pressure (the difference between current and target inventory), demand volatility (the coefficient of variation of recent demand forecasts), and resource saturation (the ratio of available resources to required resources). The weight of the immediate optimization term is increased when inventory pressure is high; the weight of the periodic optimization term is increased when demand is stable and resources are sufficient; and the weight of the long-term optimization term is increased during critical business cycles (such as quarter-end or annual planning periods). These state-based weight adjustment factors are combined with the knowledge sharing-based weight adjustment factors to generate the final weight coefficients for the three optimization terms. For example, if a warehouse agent's knowledge influence is 0.6, current inventory pressure is 0.8, demand volatility is 0.3, and resource saturation is 0.5, then the calculated weights for the immediate optimization item are 0.5, the periodic optimization item is 0.3, and the long-term optimization item is 0.2. The proximal policy optimization algorithm serves as the core algorithm for agent training. Its key feature is to maintain a manageable range of policy changes during policy updates. The policy gradient is calculated using state transition samples sampled from the hierarchical experience buffer, and excessive update steps are avoided by limiting the difference between the old and new policies. For example, in inventory allocation decisions, the updated allocation policy is ensured to be moderately different from the original policy, preventing extreme over-allocation or under-allocation. This stable policy learning approach, combined with the hierarchical experience buffer and knowledge sharing mechanism, enables each agent to gradually improve its inventory allocation strategy while maintaining decision stability.

[0129] This invention significantly improves the training efficiency and policy quality of multi-agent reinforcement learning by constructing a hierarchical experience buffer and optimizing the policy update mechanism. The hierarchical experience buffer design enables simultaneous focus on local optimization and global coordination, while the importance score calculation mechanism ensures the full utilization of key samples. A sampling strategy based on timeliness and the number of policy updates improves the representativeness of training samples. This solution innovatively divides the update process into two dimensions: "which agents to update" and "how to update these parameters." The former addresses the update scope through a hierarchical buffer, while the latter addresses the update quality through a knowledge sharing matrix and multi-timescale optimization. The knowledge sharing weight matrix facilitates experience transfer between agents, and the three-tiered optimization objective division achieves good performance across different timescales. These innovative mechanisms work together to provide an efficient and reliable learning method for intelligent scheduling in multi-echelon warehousing networks, significantly improving the intelligence level and operational efficiency of inventory management.

[0130] Optionally,

[0131] The steps of converting the inventory allocation strategy into a scheduling instruction including an allocation object, an allocation quantity, and an allocation time based on a rule knowledge base include:

[0132] Constructing a rule knowledge base including physical constraint rules, business constraint rules, and priority rules, and parsing the inventory allocation strategy into an initial scheduling plan including warehouse node pairs, allocation quantities, and allocation time windows;

[0133] Based on the rule knowledge base, the initial scheduling plan is constraint verified and a rule matching score function is constructed. The rule matching score function calculates a matching score based on the degree of physical constraint satisfaction, business goal achievement and resource utilization, and automatically corrects scheduling plans that fail constraint verification or have a matching score below a threshold;

[0134] Based on the scheduling delay rate, resource conflict rate and inventory fluctuation rate in the historical scheduling execution data, the execution risk probability of the revised scheduling plan in different time windows is calculated, and the execution priority considering risk aversion is generated. The sorted scheduling plan is converted into a scheduling instruction containing the allocation object, allocation quantity and allocation time.

[0135] In this embodiment, a rule knowledge base is constructed that includes physical constraints, business constraints, and priority rules. Physical constraints primarily describe hard logistics restrictions, such as "maximum load of a single truck is 3 tons," "Class A hazardous materials and food cannot be mixed," and "transportation time for cold chain products must not exceed 4 hours." These rules are typically related to physical facilities, transportation tools, and product characteristics and are inviolable hard constraints. Business constraints primarily describe management requirements and business process regulations for enterprise operations, such as "high-value product allocations must be approved by the supervisor," "promotional product inventory levels must not fall below 120% of forecasted demand," and "allocated quantities must be an integer multiple of the standard packaging unit." These rules are typically related to enterprise management processes, business objectives, and customer service standards and are crucial for ensuring the normal operation of the business. Priority rules primarily describe the execution order of different allocation tasks, such as "out-of-stock risk allocations take precedence over inventory balance allocations," "allocations related to Class A customer orders take precedence over Class B customer orders," and "allocations for products approaching their expiration dates take precedence over allocations for new products." These rules are typically related to business importance, customer level, and timeliness requirements, and serve as the basis for rationally allocating tasks within resource constraints. These rules are stored in a structured manner. For example, the physical constraint rule "Maximum load capacity of a single truck is 3 tons" can be stored as "Rule Type: Physical Constraint; Rule Object: Transport Vehicle; Rule Condition: Truck Model XYZ; Constraint Content: Maximum Load; Constraint Value: 3 Tons; Constraint Severity: High." This structured storage facilitates quick retrieval and application of relevant rules.

[0136] The inventory allocation policy is parsed into an initial scheduling plan consisting of warehouse node pairs, allocation quantities, and allocation time windows. For example, the policy instructs "Warehouse A to allocate 20% of the inventory of item X to Warehouse B, to be completed within the next three days." These abstract policy descriptions are converted into specific scheduling plans. For example, the reinforcement learning model receives the policy output: "The central warehouse allocates 25% of the inventory of item category A to regional warehouse 3, with an allocation time window of the next 48 hours." Querying the current inventory data reveals that the central warehouse has 8,000 items of item category A, of which 25% is 2,000 items. Therefore, the initial scheduling plan is parsed as "Allocation target: central warehouse → regional warehouse 3; Allocation item: category A; Allocation quantity: 2,000 items; Allocation time window: next 48 hours."

[0137] Constraint verification of the initial scheduling plan includes three aspects: physical constraint verification, business constraint verification, and resource conflict verification. Physical constraint verification checks whether the scheduling plan violates physical constraints, for example, checking whether the total weight of 2,000 items exceeds the load capacity of available transportation vehicles. Business constraint verification checks whether the scheduling plan complies with business regulations, for example, checking whether the remaining inventory in the central warehouse after allocation meets the minimum safety stock requirement. Resource conflict verification checks whether the scheduling plan conflicts with scheduled tasks, for example, checking whether the required transportation vehicle is already occupied by other tasks within the specified time window.

[0138] Continuing with the above example, verification revealed that the total weight of 2,000 items in Category A, totaling 5 tons, exceeded the 3-ton capacity of a single standard truck. After the allocation, 6,000 items remained in the central warehouse, exceeding the safety stock limit by 4,000 items and meeting business requirements. However, the number of trucks scheduled for use within the next 48 hours meant that the available capacity was insufficient to transport the 5 tons of goods in one go. Therefore, this initial scheduling plan failed both physical constraint verification and resource conflict verification.

[0139] A rule matching scoring function is constructed to calculate a matching score based on physical constraint satisfaction, business goal achievement, and resource utilization. Physical constraint satisfaction measures the degree to which the scheduling plan complies with physical constraints, for example, it can be calculated as the inverse ratio of "actual weight / maximum load capacity." Business goal achievement measures the scheduling plan's contribution to the business goal, for example, it can be calculated as the improvement in inventory balance after the dispatch is executed. Resource utilization measures the scheduling plan's efficiency in utilizing resources, for example, it can be calculated as the ratio of "actual load capacity / maximum load capacity."

[0140] Scheduling plans that meet both requirements (i.e., pass constraint validation and have a matching score above the threshold) are considered feasible and can proceed directly to the next step without automatic correction. Scheduling plans that fail constraint validation or have a matching score below a threshold (e.g., 0.75) are automatically corrected. Correction strategies include batching, adjusting time windows, and replacing resources. Batching breaks large batches into smaller batches to ensure that each batch complies with physical constraints; adjusting time windows schedules tasks during periods with fewer resource conflicts; and replacing resources by finding alternative transportation methods or routes to complete the task. In the example above, the dispatch of 2,000 items was automatically split into two batches of 1,000 items (2.5 tons) each, complying with the vehicle load limit. The first batch was scheduled to be completed within 24 hours, and the second batch within 24-48 hours to accommodate available transportation capacity. The revised scheduling plan becomes "Allocation object: Central Warehouse → Regional Warehouse 3; Allocation goods: Category A; Allocation quantity: 1,000 pieces for the first batch, 1,000 pieces for the second batch; Allocation time: within 24 hours for the first batch, within 24-48 hours for the second batch."

[0141] Based on historical scheduling execution data, the execution risk probability of the revised scheduling plan in different time windows is calculated. Historical data includes indicators such as scheduling delay rate, resource conflict rate, and inventory fluctuation rate. The scheduling delay rate reflects the deviation between the actual completion time and the planned time of similar scheduled tasks in the past; the resource conflict rate reflects the frequency of resource contention within the same time window in the past; and the inventory fluctuation rate reflects the difference between the actual and expected changes in inventory levels in the past.

[0142] We extracted historical records similar to the current scheduling plan from the historical database, for example, a similar-sized dispatch task on the "Central Warehouse → Regional Warehouse 3" route. Analysis revealed that within a 24-hour window, this route had a historical dispatch delay rate of 15%, a resource conflict rate of 10%, and an inventory volatility rate of 8%. In contrast, within a 24-48 hour window, the historical dispatch delay rate was 25%, the resource conflict rate was 18%, and the inventory volatility was 12%. This indicates that the execution risk for the second batch is significantly higher than that for the first batch.

[0143] Generate risk-averse execution priorities and convert the prioritized scheduling plans into scheduling instructions. Execution priorities are determined based on a comprehensive assessment of task importance, timeliness, and execution risk. Task importance is determined by business impact, timeliness by the deployment deadline, and execution risk by the aforementioned risk probabilities. Tasks with higher priorities receive higher priority in resource allocation and execution.

[0144] The sorted dispatch plan is converted into a dispatch instruction with specific execution details. Dispatch instructions include: dispatch object (source warehouse and destination warehouse), dispatched product (specific to the SKU level), dispatch quantity (accurate to the number of pieces), dispatch time (accurate to the hour and minute), resources to be used (specific vehicle number or personnel arrangement), and operational requirements (such as special handling precautions for special products).

[0145] In the above case, the final dispatch instruction is: "Instruction number: DS2023051001; Allocation object: Central warehouse CD001 → Regional warehouse RD003; Allocation products: Class A SKU10015-10025; Allocation quantity: 1000 pieces; Allocation time: 2023-05-10 08:00; Resources used: Logistics vehicle TK0058, operators OP108, OP109; Operation requirements: Standard Loading and unloading procedure B12" and "Instruction number: DS2023051002; Allocation object: Central warehouse CD001 → Regional warehouse RD003; Allocated goods: Class A SKU10015-10025; Allocation quantity: 1000 pieces; Allocation time: 2023-05-1110:00; Resources used: Logistics vehicle TK0062, operators OP112, OP115; Operation requirements: Standard loading and unloading procedure B12".

[0146] This invention achieves efficient transformation from abstract strategies to specific executable instructions by constructing a rule knowledge base and a multi-stage transformation process. The construction of the rule knowledge base ensures that scheduling instructions comply with physical constraints and business requirements. Constraint verification and automatic correction of the initial scheduling plan improve the feasibility of the plan. Risk assessment and priority sorting based on historical data enhance the reliability of scheduling execution. This rule-based strategy transformation mechanism effectively solves the "last mile" problem between AI decision-making and actual execution, significantly improving the implementation of inventory allocation strategies, reducing execution anomaly rates, and improving the overall operational efficiency of warehousing networks, providing key technical support for enterprises to achieve intelligent inventory management.

[0147] Optionally,

[0148] The warehouse node that receives the emergency coordination request returns a response message based on its own resource status. The steps of determining an emergency handling plan through local negotiation between nodes based on the response message include:

[0149] Calculate the logistics relevance based on the historical logistics frequency and current logistics intensity of the warehouse node, calculate the resource response capability based on the ratio of the requested resources to the available resources, and generate response information including response time, response cost and resource status based on the logistics relevance and resource response capability;

[0150] Obtain the task completion rate and resource stability of the warehouse nodes participating in the negotiation, calculate the node credibility based on the response information, generate an influence weight based on the number of hops between nodes, and the influence weight decreases exponentially with the increase of the hop number. Determine the initial negotiation weight based on the node credibility and influence weight;

[0151] Receive the negotiation solutions of neighboring warehouse nodes, calculate the weighted average of the solutions based on the initial negotiation weights, combine the weighted average with the current optimal solution to form an updated local solution, and calculate the variance of the updated local solution within the negotiation group as a consistency indicator;

[0152] When the update amplitude of the plan is continuously lower than the preset amplitude threshold, the cost, time and resource consumption weights in the negotiation plan are adjusted according to the historical execution results, and a random perturbation value positively correlated with the duration of the stagnation is generated to restart the negotiation; when the consistency index meets the preset conditions, an emergency treatment plan is generated;

[0153] Monitor the heartbeat information and response delay of the warehouse nodes participating in the negotiation. When it is detected that the heartbeat information is interrupted or the response delay exceeds the preset value, obtain the unfinished negotiation tasks of the heartbeat abnormal node, calculate the task reception fitness based on the resource status, response time and logistics distance of the remaining warehouse nodes, and reallocate the unfinished negotiation tasks based on the task reception fitness while satisfying the resource constraints.

[0154] In this embodiment, a scheduling execution module is deployed in the edge computing unit of each warehouse node. This module receives the aforementioned scheduling instructions and executes them according to priority. When an inventory anomaly or resource conflict is detected during execution, the module automatically initiates an emergency coordination request to adjacent warehouse nodes. Inventory anomalies include various situations, such as actual inventory lower than recorded (inventory loss), inventory quality issues (such as damaged goods), and insufficient inventory due to a sudden surge in orders. Resource conflicts include situations where transportation failures, personnel absences, warehouse equipment failures, and other factors prevent scheduling from being executed as planned. For example, when a forward warehouse discovers that the actual inventory of Class A goods is 30% less than recorded during a delivery task, resulting in an inability to meet the day's delivery needs, it automatically initiates an emergency coordination request to adjacent regional warehouses and other forward warehouses. The request includes the following content: "Missing item type: Class A; Missing quantity: 150 items; Required urgency: High; Required response time: Within 2 hours."

[0155] Warehouse nodes that receive emergency coordination requests evaluate their response capabilities based on two key metrics: logistics relevance and resource responsiveness. Logistics relevance is calculated based on historical logistics frequency and current logistics intensity. Historical logistics frequency refers to the number of logistics interactions between two warehouse nodes over a period of time (e.g., 30 days), while current logistics intensity refers to the number of logistics tasks currently being executed between the two nodes. For example, if two warehouse nodes have had 25 logistics interactions over the past 30 days and currently have three logistics tasks being executed, the logistics relevance can be calculated as "historical frequency / reference period + current intensity": 25 / 30 + 3 = 3.83. A higher logistics relevance indicates closer collaboration between the two nodes and higher collaborative efficiency. Resource responsiveness is calculated based on the ratio of resources required for the request to available resources. Resources include inventory, transportation, and human resources. For example, if a request requires 150 items of Category A, and the node currently has 200 items of Category A available, the inventory resource response ratio is 200 / 150 = 1.33, indicating that the request can be fully met. If a request requires two transport vehicles, and the node currently has only one available, the transport resource response ratio is 1 / 2 = 0.5, indicating that it can only be partially met. The minimum of the resource response ratios is taken as the overall resource response capacity, reflecting the "barrel principle."

[0156] Based on logistics relevance and resource responsiveness, the receiving node generates a response message containing response time, response cost, and resource status. Response time refers to the time required to complete the requested task, including preparation and logistics time; response cost includes resource allocation costs, transportation costs, and opportunity costs; and resource status details the quantity and quality of available resources. For example, a regional warehouse returns the following response message to the previously mentioned out-of-stock forward warehouse: "Available Class A product: 120 pieces; Response time: 90 minutes; Response cost: 2,000 yuan; Resource status: 1 transport vehicle available, 2 operators available."

[0157] The requesting node obtains the task completion rate and resource stability of each participating responding node and combines these responses to calculate the node's credibility. The task completion rate represents the percentage of collaborative tasks successfully completed by the node over the past period, while resource stability reflects the consistency of the node's resource supply. For example, if a responding node has a past task completion rate of 92% and a resource stability score of 0.85 (out of a maximum score of 1), combined with the promised completion time and resource availability in its response, its node credibility can be calculated as 0.92 × 0.85 × (Promised Resources / Requested Resources) = 0.92 × 0.85 × (120 / 150) = 0.626. The higher the node's credibility, the greater the reference value of its response.

[0158] The influence weight is generated based on the number of hops between nodes, and the influence weight decreases exponentially as the number of hops increases. The hop count refers to the shortest path length between two nodes in the network. The hop count of directly connected nodes is 1, the hop count of a node connected through an intermediate node is 2, and so on. The influence weight can be calculated by multiplying the base weight by the hop count power of the attenuation coefficient. For example, if the base weight is 1 and the attenuation coefficient is 0.7, the influence weight of a node with a hop count of 1 is 1×0.7. 1 =0.7, the influence weight of a node with 2 hops is 1×0.7 2 =0.49. This design gives nodes with closer physical proximity greater influence in negotiations, which aligns with the actual characteristics of logistics networks. The initial negotiation weight is the product of a node's credibility and its influence weight. For example, if a node has a credibility of 0.626 and an influence weight of 0.7, its initial negotiation weight is 0.626 × 0.7 = 0.438. The initial negotiation weights of each participating node are normalized so that the sum of all weights is 1, which serves as the final negotiation weight.

[0159] During the negotiation process, each node receives proposals from neighboring nodes and calculates a weighted average of these proposals based on negotiation weights. For example, if a node receives proposals from three neighboring nodes, namely "Product Allocation Quantity: 100, 80, and 120," with corresponding negotiation weights of "0.5, 0.3, and 0.2," the weighted average is 100 × 0.5 + 80 × 0.3 + 120 × 0.2 = 98 items. The node combines this weighted average with the current optimal proposal (e.g., the best locally feasible proposal) to form an updated local proposal. This combination can be performed using linear interpolation, for example, "Updated proposal = current optimal proposal × (1 - learning rate) + weighted average × learning rate," where the learning rate is a parameter between 0 and 1 that controls the aggressiveness of proposal updates.

[0160] The variance of the updated local plans within the negotiation group is calculated as a consistency indicator. A smaller variance indicates closer agreement between the plans of each node, and closer agreement between the two nodes. For example, if the updated plans of the three nodes are "Product Allocation Quantity: 95, 98, 102," the variance is small, indicating near-consistency. If the plans differ significantly, such as "80, 100, 120," the variance is large, indicating further negotiation.

[0161] When the amplitude of a proposal update continuously falls below a preset amplitude threshold, it indicates that the negotiation has reached a local stagnation. For example, if the amplitude of a proposal update is less than 2% for five consecutive negotiation rounds, while the preset threshold is 3%, the negotiation is considered to be stagnant. At this point, the weights for cost, time, and resource consumption in the negotiation proposal are adjusted based on historical performance. For example, if historical data shows that lowering the weight for cost and increasing the weight for time effectively breaks the stagnation in similar situations, the weights are adjusted accordingly. At the same time, a random perturbation value positively correlated with the duration of the stagnation is generated to restart the negotiation. The perturbation value can be designed as "base perturbation amplitude × (1 + number of stagnation rounds / 10)", so that the longer the stagnation, the greater the perturbation, increasing the probability of escaping the local optimum.

[0162] When the consistency indicator meets the preset conditions, a final emergency response plan is generated. The preset conditions might be "the variance value is less than a threshold and the variance is less than 1% for three consecutive rounds," indicating that all nodes have reached a stable and consistent plan. The final plan details the specific tasks for each participating node, including the type and quantity of resources provided, the execution time, and the responsible individuals. For example, the final plan might be "Regional Warehouse A provides 100 items of Category A goods, and Forward Warehouse B provides 50 items of Category A goods, to be delivered to the requesting node before 2:00 PM and 2:30 PM, respectively, with their respective logistics teams responsible for execution."

[0163] Real-time monitoring of the heartbeat information and response latency of warehouse nodes participating in the negotiation. Heartbeat information is a status notification sent periodically by a node, indicating normal operation. Response latency is the time interval between a node receiving a message and returning a response. If a node's heartbeat information is interrupted (e.g., no heartbeats received for three consecutive times) or its response latency exceeds a preset value (e.g., three times the normal value), the node is deemed faulty or has experienced a network outage. Unfinished negotiation tasks for the node with the abnormal heartbeat are obtained, and the task acceptance fitness is calculated based on the resource status, response time, and logistics distance of the remaining healthy nodes. Task acceptance fitness measures a node's ability to accept tasks from a faulty node and is calculated as the weighted average of "resource sufficiency × time response × distance suitability." For example, if a node's resource sufficiency is 0.8 (basically sufficient resources), its time response is 0.9 (capable of timely response), and its distance suitability is 0.7 (close but not the closest), its task acceptance fitness is 0.8 × 0.4 + 0.9 × 0.4 + 0.7 × 0.2 = 0.82 (assuming weights are 0.4, 0.4, and 0.2, respectively).

[0164] Based on the task's suitability for receiving, unfinished negotiated tasks are reallocated while meeting resource constraints. Resource constraints ensure that task reallocation does not overstrain the receiving node's resources. The principle of allocation is to prioritize nodes with high suitability and sufficient resources. If necessary, large tasks can be split across multiple nodes. For example, if a faulty node was originally scheduled to deliver 100 items, the task can be reallocated to "node C delivering 60 items and node D delivering 40 items," ensuring the reliable execution of the overall emergency plan.

[0165] Figure 2 This figure shows the impact of the gradual iteration of the various technical components of the present invention on system performance. With the addition of key technical components, negotiation convergence rounds were significantly reduced from 25 in the base version to 3 in the final version, improving negotiation efficiency. Resource utilization also increased from 65.2% to 85.1%, and solution consistency improved from 69.0% to 87.4%. In particular, the addition of stall detection and heartbeat monitoring mechanisms significantly improved system performance, significantly reducing convergence rounds while maintaining high resource utilization and solution consistency. This demonstrates that the technical solution of the present invention effectively balances resource utilization and solution quality while improving decision-making efficiency. By constructing a distributed emergency coordination mechanism, the present invention effectively addresses sudden inventory anomalies and resource conflicts in multi-tiered warehousing networks. The local response capabilities of edge computing enable the system to maintain operation even when the central system is unavailable. Comprehensive assessment of logistics relevance and resource responsiveness ensures the scientific nature of collaborative decision-making. Dynamic node credibility and influence weight calculation enhance negotiation efficiency and solution quality. Automatic detection and restart of negotiation stalls strengthens the system's ability to find the optimal solution. Heartbeat monitoring and task reallocation mechanisms improve the system's resilience to node failures. This distributed collaborative architecture significantly reduces exception processing time, improves resource utilization efficiency, enhances the overall resilience of the supply chain, and provides strong technical support for enterprises to cope with complex and changing market environments.

[0166] A second aspect of an embodiment of the present invention provides a supply chain multi-level warehousing intelligent scheduling and collaboration system, including:

[0167] The first unit is used to obtain real-time inventory data and historical order data of warehouses at all levels in the multi-level warehousing network;

[0168] The second unit is configured to perform hierarchical modeling and prediction of short-term and long-term demands of warehouses at all levels using a long short-term memory network based on the historical order data to obtain inventory demand;

[0169] The third unit is used to build a multi-agent reinforcement learning model, setting each warehouse node as an independent agent. The state space of the multi-agent reinforcement learning model includes the real-time inventory data, inventory demand, and available resources of each warehouse node. The action space includes the inventory allocation quantity and allocation timing. The inventory allocation strategy is generated through inter-agent policy iterative optimization.

[0170] The fourth unit is configured to convert the inventory allocation strategy into a scheduling instruction including an allocation object, an allocation quantity, and an allocation time based on a rule knowledge base;

[0171] The fifth unit is used to receive the scheduling instructions and sort them for execution in the edge computing unit of each warehouse node. When inventory anomalies or resource conflicts are detected, an emergency collaboration request is initiated to the adjacent warehouse node. The warehouse node that receives the emergency collaboration request returns response information based on its own resource status. Based on the response information, the emergency handling plan is determined through local consultation between nodes.

[0172] According to a third aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

Claims

1. Supply chain multi-level warehousing intelligent scheduling and collaboration method, characterized by: include: Obtain real-time inventory data and historical order data for warehouses at all levels in a multi-level warehousing network; Based on the historical order data, a long short-term memory network is used to perform hierarchical modeling and prediction of short-term and long-term demand of warehouses at all levels to obtain inventory demand; A multi-agent reinforcement learning model is constructed, and each warehouse node is set as an independent agent. The state space of the multi-agent reinforcement learning model includes the real-time inventory data, inventory demand and available resources of each warehouse node, and the action space includes the inventory allocation quantity and allocation timing. The inventory allocation strategy is generated through policy iterative optimization between agents, including: using a proximal policy optimization algorithm to train independent agents, and constructing a hierarchical experience buffer to store state transition samples during the training process, including: the hierarchical experience buffer includes a local experience buffer for storing local state transition samples of each independent agent, a collaborative experience buffer for storing interaction samples of adjacent agents, and a global experience buffer for storing global state samples; calculating the importance score of the state transition sample, and the importance score is determined based on the reward value, state transition frequency, state influence range and strategy difference. Determine, store the state transition samples in the experience buffer of the corresponding level according to the importance score; calculate the sampling probability of the samples based on the timeliness of the samples and the number of strategy updates, and select training samples from the hierarchical experience buffer according to the sampling probability; update the strategy network parameters of a single agent based on the samples in the local experience buffer, update the strategy network parameters of the adjacent agent group based on the samples in the collaborative experience buffer, and update the strategy network parameters of all agents based on the samples in the global experience buffer; construct a knowledge sharing weight matrix, and calculate the element values ​​of the knowledge sharing weight matrix based on the historical decision data of the agent and the network topology structure; divide the objective function of the proximal strategy optimization algorithm into immediate optimization items, periodic optimization items and long-term optimization items, calculate the weight coefficient of each optimization item based on the knowledge sharing weight matrix and system state information, and optimize and update the strategy network parameters; Converting the inventory allocation strategy into a scheduling instruction including allocation objects, allocation quantities and allocation time based on a rule knowledge base; In the edge computing unit of each warehouse node, the scheduling instructions are received and sorted for execution. When inventory anomalies or resource conflicts are detected, an emergency collaboration request is initiated to the adjacent warehouse node. The warehouse node that receives the emergency collaboration request returns response information based on its own resource status. Based on the response information, the emergency handling plan is determined through local consultation between nodes.

2. The method according to claim 1, characterized in that Based on the historical order data, a long short-term memory network is used to perform hierarchical modeling and prediction of short-term and long-term demand of warehouses at all levels. The steps of obtaining inventory demand include: Dividing the historical order data into daily order sequences and weekly order sequences; An adaptive hierarchical forecasting network is constructed, comprising a short-term forecasting layer and a long-term forecasting layer. The short-term forecasting layer receives daily order sequences and dynamically adjusts the forecast time window based on inventory categories and seasonal characteristics. The long-term forecasting layer receives weekly order sequences and captures long-term patterns using a memory enhancement module. The output error gradient of the short-term forecasting layer is passed to the parameter update process of the long-term forecasting layer, embedding the trend characteristics of the long-term forecasting layer into the input of the short-term forecasting layer. The adaptive hierarchical prediction network is trained using a multi-scale loss function, wherein the multi-scale loss function includes a short-term prediction loss term, a long-term prediction loss term, and a regularization term, and the loss terms are dynamically weighted based on the time scale characteristics of each prediction layer; The forecast results are integrated to obtain an initial forecast value; the initial forecast value is calibrated based on the storage capacity constraint, and combined with the forecast standard deviation to generate an inventory demand forecast result with a confidence interval.

3. The method according to claim 1, characterized in that The steps for building a multi-agent reinforcement learning model, setting each warehouse node as an independent agent, and generating an inventory allocation strategy through inter-agent policy iterative optimization include: Each warehouse node is set as an independent intelligent agent, and a state vector of the independent intelligent agent is constructed. The state vector includes real-time inventory data, inventory demand, and available resources. An adjacency matrix is ​​constructed based on the logistics channels between each warehouse node. The adjacency matrix is ​​used to represent the connection relationship between each warehouse node. Constructing an action space for an independent agent, the action space including a transfer quantity matrix and a transfer timing vector, wherein the transfer quantity matrix is ​​used to represent the inventory transfer quantity between warehouse nodes, and the transfer timing vector is used to represent the transfer execution time; Constructing a multi-objective reward function, which includes a local reward function and a global reward function. The local reward function calculates the reward value of each independent agent based on inventory balance, turnover efficiency, and scheduling cost, and the global reward function calculates the global collaborative reward value based on the inventory level difference of adjacent warehouse nodes; Independent agents are trained using a proximal policy optimization algorithm. During the training process, a hierarchical experience buffer is constructed to store state transition samples. The hierarchical experience buffer includes a local buffer for storing the local experience of the independent agents and a global buffer for storing global state information. A parameter sharing mechanism is used to train the policy network for independent agents of the same type. Based on the trained policy network, an inventory allocation strategy is generated according to the real-time status of each warehouse node.

4. The method according to claim 3, characterized in that The steps to construct a multi-objective reward function include: Items are classified based on historical order data for each warehouse node, and target inventory levels are determined. Based on the target inventory levels, the standardized inventory deviation for each warehouse node is calculated, and the product of the standardized inventory deviation and the time decay factor is determined as the inventory balance reward value. The daily, weekly, and monthly turnover rates are calculated based on the outbound volume, beginning inventory, and ending inventory of each warehouse node, and the weighted sum of the results with the preset weight coefficient is used as the turnover efficiency bonus value. Determine the dispatch cost bonus value based on the weighted sum of storage cost, transportation cost and operation cost; Calculating the inventory level difference ratio between adjacent warehouses to construct a difference matrix, calculating the distance weight coefficient based on the logistics distance between adjacent warehouses, and determining the global collaboration reward value as the product of the distance weight coefficient, the difference matrix, the warehouse hierarchical collaboration coefficient, and the congestion penalty factor; The performance improvement rates of the inventory balance reward value, turnover efficiency reward value, scheduling cost reward value, and global synergy reward value are calculated within a preset evaluation period. A multivariate linear regression model is constructed based on the performance improvement rates to calculate a correlation coefficient matrix. The contribution of each reward value is calculated according to the correlation coefficient matrix. The difference between the contribution and the current weight is multiplied by the learning rate to obtain a weight adjustment amount. The weight coefficient of each reward value is updated under the weight adjustment constraint to obtain a multi-objective reward function.

5. The method according to claim 1, wherein The steps of converting the inventory allocation strategy into a scheduling instruction including an allocation object, an allocation quantity, and an allocation time based on a rule knowledge base include: Constructing a rule knowledge base including physical constraint rules, business constraint rules, and priority rules, and parsing the inventory allocation strategy into an initial scheduling plan including warehouse node pairs, allocation quantities, and allocation time windows; Based on the rule knowledge base, the initial scheduling plan is constraint verified and a rule matching score function is constructed. The rule matching score function calculates a matching score based on the degree of physical constraint satisfaction, business goal achievement and resource utilization, and automatically corrects scheduling plans that fail constraint verification or have a matching score below a threshold; Based on the scheduling delay rate, resource conflict rate and inventory fluctuation rate in the historical scheduling execution data, the execution risk probability of the revised scheduling plan in different time windows is calculated, and the execution priority considering risk aversion is generated. The sorted scheduling plan is converted into a scheduling instruction containing the allocation object, allocation quantity and allocation time.

6. The method according to claim 1, characterized in that The warehouse node that receives the emergency coordination request returns a response message based on its own resource status. The steps of determining an emergency handling plan through local negotiation between nodes based on the response message include: Calculate the logistics relevance based on the historical logistics frequency and current logistics intensity of the warehouse node, calculate the resource response capability based on the ratio of the requested resources to the available resources, and generate response information including response time, response cost and resource status based on the logistics relevance and resource response capability; Obtain the task completion rate and resource stability of the warehouse nodes participating in the negotiation, calculate the node credibility based on the response information, generate an influence weight based on the number of hops between nodes, and the influence weight decreases exponentially with the increase of the hop number. Determine the initial negotiation weight based on the node credibility and influence weight; Receive the negotiation solutions of neighboring warehouse nodes, calculate the weighted average of the solutions based on the initial negotiation weights, combine the weighted average with the current optimal solution to form an updated local solution, and calculate the variance of the updated local solution within the negotiation group as a consistency indicator; When the update amplitude of the plan is continuously lower than the preset amplitude threshold, the cost, time and resource consumption weights in the negotiation plan are adjusted according to the historical execution results, and a random perturbation value positively correlated with the duration of the stagnation is generated to restart the negotiation; when the consistency index meets the preset conditions, an emergency treatment plan is generated; Monitor the heartbeat information and response delay of the warehouse nodes participating in the negotiation. When it is detected that the heartbeat information is interrupted or the response delay exceeds the preset value, obtain the unfinished negotiation tasks of the heartbeat abnormal node, calculate the task reception fitness based on the resource status, response time and logistics distance of the remaining warehouse nodes, and reallocate the unfinished negotiation tasks based on the task reception fitness while satisfying the resource constraints.

7. A supply chain multi-level warehousing intelligent scheduling and coordination system, used to implement the method according to any one of claims 1 to 6, characterized in that: include: The first unit is used to obtain real-time inventory data and historical order data of warehouses at all levels in the multi-level warehousing network; The second unit is configured to perform hierarchical modeling and prediction of short-term and long-term demands of warehouses at all levels using a long short-term memory network based on the historical order data to obtain inventory demand; The third unit is used to build a multi-agent reinforcement learning model, setting each warehouse node as an independent agent. The state space of the multi-agent reinforcement learning model includes the real-time inventory data, inventory demand, and available resources of each warehouse node. The action space includes the inventory allocation quantity and allocation timing. The inventory allocation strategy is generated through inter-agent policy iterative optimization. The fourth unit is configured to convert the inventory allocation strategy into a scheduling instruction including an allocation object, an allocation quantity, and an allocation time based on a rule knowledge base; The fifth unit is used to receive the scheduling instructions and sort them for execution in the edge computing unit of each warehouse node. When inventory anomalies or resource conflicts are detected, an emergency collaboration request is initiated to the adjacent warehouse node. The warehouse node that receives the emergency collaboration request returns response information based on its own resource status. Based on the response information, the emergency handling plan is determined through local consultation between nodes.

8. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images