Electronic commerce inventory dynamic scheduling management method and system based on artificial intelligence

Abstract

The application relates to the field of artificial intelligence, and discloses an e-commerce inventory dynamic scheduling management method and system based on artificial intelligence, multi-source heterogeneous data of an e-commerce platform is collected, a supply chain heterogeneous graph including a warehouse node, a supplier node and a distribution center node is constructed based on the multi-source heterogeneous data; a graph attention network is used to perform feature extraction on the supply chain heterogeneous graph to obtain dynamic representation vectors of the nodes; the dynamic representation vectors of the nodes are input into a TCN-Transformer model to generate a multi-step predicted sales sequence; the dynamic representation vectors of the nodes and the multi-step predicted sales sequence are input into a reinforcement learning framework including an upper-layer global coordination agent and a lower-layer node-level agent to make an inventory scheduling decision for a current node; the inventory scheduling decision is issued to each inventory node for execution, and the inventory state changes and order fulfillment results after the execution are collected in real time to make iterations; and the application improves inventory turnover efficiency.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to an artificial intelligence-based method and system for dynamic scheduling and management of e-commerce inventory. Background Technology

[0002] With the rapid development of the e-commerce industry, marketing scenarios such as live-streaming sales, flash sales, and holiday promotions have become commonplace. E-commerce orders are characterized by explosive growth, fragmentation, and volatility, which places extremely high demands on the real-time, accurate, and dynamic nature of inventory scheduling. Current e-commerce inventory management methods severely restrict the operational efficiency of the e-commerce supply chain. Existing technologies mostly adopt a static threshold control model, which pre-sets fixed inventory warning values, replenishment quantities, and allocation rules, and cannot adjust the thresholds dynamically in real time. Summary of the Invention

[0003] The purpose of this invention is to solve the above problems by designing an artificial intelligence-based method and system for dynamic scheduling and management of e-commerce inventory.

[0004] The first aspect of this invention provides an artificial intelligence-based method for dynamic scheduling and management of e-commerce inventory, the method comprising the following steps: Collect multi-source heterogeneous data from e-commerce platforms, and construct a supply chain heterogeneous graph including warehouse nodes, supplier nodes, and distribution center nodes based on the multi-source heterogeneous data; A graph attention network is used to extract features from the heterogeneous supply chain graph to obtain the dynamic representation vector of each node; The dynamic representation vectors of each node are input into the TCN-Transformer model. First, short-term temporal dependency features are extracted through the TCN layer, and then long-term correlation features across cycles are captured through the Transformer encoder. After feature fusion, a multi-step predicted sales sequence is generated. The dynamic representation vectors of each node and the multi-step predicted sales sequence are input into a reinforcement learning framework that includes an upper-level global coordinating agent and a lower-level node-level agent to make inventory scheduling decisions for the current node. A causal graph structure is introduced to describe the causal relationships between inventory nodes. The inventory scheduling decision is sent to each inventory node for execution, and the changes in inventory status and order fulfillment results are collected in real time for iteration.

[0005] Optionally, in a first implementation of the first aspect of the present invention, the step of collecting multi-source heterogeneous data from an e-commerce platform and constructing a supply chain heterogeneous graph including warehouse nodes, supplier nodes, and distribution center nodes based on the multi-source heterogeneous data includes: Collect multi-source heterogeneous data from e-commerce platforms, including at least historical sales time-series data, price and promotional activity data, and product-related purchase data for each inventory node; The collected raw multi-source heterogeneous data is cleaned, and the categorized data is converted by one-hot encoding to obtain the initial feature vector of each node. The data cleaning includes removing outliers, filling missing values, and normalization. Based on the physical topology of the e-commerce inventory system, each independent inventory management unit is defined as a node, where each node includes at least a warehouse node, a supplier node, and a distribution center node. Using all nodes as vertices, initial edge connections are established based on actual logistics paths and supply chain relationships. For each edge, a weight is calculated based on logistics timeliness factors, sales coupling relationships, and transportation efficiency. The calculated weights are multiplied by preset adjustable coefficients to obtain the final weight of the edge. The weights of all edges are then used to construct a weighted adjacency matrix. The node set, the weighted edge set, and the initial feature vectors of each node are integrated to form a supply chain heterogeneous graph.

[0006] Optionally, in a second implementation of the first aspect of the present invention, the step of using a graph attention network to extract features from the heterogeneous supply chain graph to obtain dynamic representation vectors for each node includes: The heterogeneous graph of the supply chain is input into the graph attention network, and the initial feature vector of each node is linearly transformed to obtain the transformed feature vector. For each attention head, the attention coefficient is calculated for the central node and its neighboring nodes connected by any edge in the supply chain heterogeneous graph, and the attention coefficient is normalized using softmax. Multiply the transformed feature vector of each neighbor node by the corresponding normalized attention coefficient to obtain the weighted neighbor features; Sum the weighted features of all neighbors and add the features of the center node to obtain the new feature vector of the current center node; The operation is repeated with 8 attention heads, and the feature vectors output by each attention head are concatenated to obtain the final dynamic representation vector of the current center node. The operations of a single-layer graph attention network are stacked into three layers, with the input of each layer being the output of the previous layer. Through multiple stacking layers, the final dynamic feature vectors of all nodes are output after passing through the multi-layer graph attention network, thus obtaining the dynamic representation vector of each node.

[0007] Optionally, in a third implementation of the first aspect of the present invention, the step of inputting the dynamic representation vectors of each node into the TCN-Transformer model, first extracting short-term temporal dependency features through the TCN layer, then capturing long-term correlation features across periods through the Transformer encoder, and generating a multi-step predicted sales sequence after feature fusion includes: The dynamic representation vector of each node is arranged in chronological order to form an input sequence window, which is then input into the TCN-Transformer model. The TCN module consists of multiple stacked dilated causal convolutional layers. Each convolutional layer is processed by an activation function and batch normalization. By stacking convolutional layers with different dilation rates, it captures short-term fluctuation patterns and outputs short-term time series features. The input sequence with added position encoding is fed into the Transformer encoder, which consists of multiple identical layers stacked together, each layer including a multi-head self-attention sublayer and a feedforward neural network sublayer; The multi-head self-attention sublayer calculates attention weights for each time step in the input sequence, capturing long-term dependencies across cycles. The feedforward neural network sublayer performs a non-linear transformation on the attention output at each time step, outputting long-term global features. Short-term temporal features and long-term global features are fused to obtain fused features. The input is then processed by a fully connected layer, which maps the features to sales forecasts for multiple future time steps, and integrates them to obtain a multi-step predicted sales sequence.

[0008] Optionally, in a fourth implementation of the first aspect of the present invention, the step of inputting the dynamic representation vectors of each node and the multi-step predicted sales sequence into a reinforcement learning framework including an upper-level global coordinating agent and lower-level node-level agents to make inventory scheduling decisions for the current node includes: The upper-level global coordinating agent receives the dynamic representation vectors and multi-step predicted sales sequences from all nodes, forming a global state space, and outputs global constraints based on the MAPPO algorithm. The lower-level node-level agent receives global constraints and local observation states of each node, and uses the DDPG algorithm to output the decision actions of each node. The local observation states include at least the dynamic representation vector, multi-step predicted sales sequence, current inventory, in-transit replenishment quantity, and order backlog. The global constraints output by the upper-level global coordinating agent and the decision actions output by all lower-level node-level agents are combined to form the final inventory scheduling decision.

[0009] Optionally, in a fifth implementation of the first aspect of the present invention, the upper-layer global coordination agent receives the dynamic representation vectors of all nodes and the multi-step predicted sales sequence to form a global state space, and outputs global constraints based on the MAPPO algorithm, including: In the current global state space, the policy network of the upper-level global coordinating agent executes the action of the global constraint parameter vector according to the input and output global constraint parameter vector. The environment returns the reward value and stores the tuple in the experience replay buffer. Experience data is sampled from the experience replay buffer. For each sampled experience data, the value of the current state and the value of the next state are evaluated using a value network. The temporal difference error is calculated. The temporal difference errors of multiple time series are weighted and summed using a generalized dominance estimation algorithm to obtain the dominance function value for each time step. Obtain the policy parameters used when sampling each empirical data point, calculate the probability of taking an action in a given state using the old policy, calculate the probability under the current new policy parameters, and calculate the probability ratio. Calculate the KL divergence between the new and old policies, construct the MAPPO objective function based on the probability ratio, and update the policy network parameters using the Adam optimizer; The loss function of the value network is calculated, and the parameters of the value network are updated by gradient descent. After multiple training iterations, the policy network converges and the global constraints are directly applied.

[0010] Optionally, in a sixth implementation of the first aspect of the present invention, the lower-level node-level agent receives global constraints and the local observation states of each node, and outputs the decision actions of each node using the DDPG algorithm, including: Obtain historical inventory level sequences, demand sequences, and scheduling action sequences for each node, use the PC algorithm to perform conditional independence tests, and construct a cause-effect graph; During the policy update process of lower-level node agents, it is calculated whether the current action violates the causal dependency relationship in the causal graph. If violated, the causal consistency loss is calculated, multiplied by the regularization coefficient, and then added to the policy gradient. During gradient descent updates, this causes the policy network to favor causal consistent actions.

[0011] A second aspect of the present invention provides an artificial intelligence-based e-commerce inventory dynamic scheduling and management system, the system comprising: The module is used to collect multi-source heterogeneous data from e-commerce platforms and construct a supply chain heterogeneous graph including warehouse nodes, supplier nodes, and distribution center nodes based on the multi-source heterogeneous data. The feature extraction module is used to extract features from the heterogeneous supply chain graph using a graph attention network to obtain the dynamic representation vector of each node. The prediction module is used to input the dynamic representation vector of each node into the TCN-Transformer model. First, the TCN layer extracts short-term temporal dependency features, and then the Transformer encoder captures long-term correlation features across cycles. After feature fusion, a multi-step predicted sales sequence is generated. The decision module is used to input the dynamic representation vectors of each node and the multi-step predicted sales sequence into a reinforcement learning framework that includes an upper-level global coordinating agent and a lower-level node-level agent to make inventory scheduling decisions for the current node. A causal graph structure is introduced to describe the causal relationships between inventory nodes. The execution module is used to distribute inventory scheduling decisions to each inventory node for execution, and to collect and iterate the changes in inventory status and order fulfillment results after execution in real time.

[0012] A third aspect of the present invention provides an AI-based e-commerce inventory dynamic scheduling and management device, the AI-based e-commerce inventory dynamic scheduling and management device comprising a memory and at least one processor, the memory storing instructions; the at least one processor invokes the instructions in the memory to cause the AI-based e-commerce inventory dynamic scheduling and management device to perform the various steps of the AI-based e-commerce inventory dynamic scheduling and management method as described in any of the preceding claims.

[0013] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the steps of the AI-based e-commerce inventory dynamic scheduling and management method as described in any of the preceding claims.

[0014] The technical solution provided by this invention involves collecting multi-source heterogeneous data from e-commerce platforms, constructing a supply chain heterogeneous graph including warehouse nodes, supplier nodes, and distribution center nodes based on the multi-source heterogeneous data, extracting features from the supply chain heterogeneous graph using a graph attention network to obtain dynamic representation vectors for each node, inputting the dynamic representation vectors of each node into a TCN-Transformer model, first extracting short-term temporal dependency features through the TCN layer, then capturing long-term correlation features across cycles through the Transformer encoder, and generating a multi-step predicted sales sequence after feature fusion; inputting the dynamic representation vectors of each node and the multi-step predicted sales sequence into a reinforcement learning framework including an upper-layer global coordinating agent and a lower-layer node-level agent to make inventory scheduling decisions for the current node; distributing the inventory scheduling decisions to each inventory node for execution, and collecting real-time changes in inventory status and order fulfillment results for iteration; this invention effectively improves sales forecast accuracy, significantly reduces inventory costs and stockout rates, improves inventory turnover efficiency, and is adaptable to various marketing and seasonal scenarios in e-commerce. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0016] Figure 1 A flowchart illustrating an AI-based e-commerce inventory dynamic scheduling and management method provided in an embodiment of the present invention; Figure 2 A schematic diagram of the structure of an AI-based e-commerce inventory dynamic scheduling and management system provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an AI-based e-commerce inventory dynamic scheduling and management device provided in an embodiment of the present invention. Detailed Implementation

[0017] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” or “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, apparatus, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0018] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 The flowchart of the AI-based e-commerce inventory dynamic scheduling and management method provided in this embodiment of the invention includes the following steps: Step 101: Collect multi-source heterogeneous data from e-commerce platforms and construct a supply chain heterogeneous graph including warehouse nodes, supplier nodes, and distribution center nodes based on the multi-source heterogeneous data; In this embodiment, multi-source heterogeneous data is synchronously acquired from multiple business systems of the e-commerce platform through a data interface or log collection module. Historical sales time-series data records the sales quantity of each SKU at each inventory node at a daily or hourly granularity. Price and promotional activity data includes the product's base price, discount level, spending threshold rules, and major promotion markers. Product-related purchase data is obtained based on order co-occurrence analysis, reflecting the probability of bundled purchases between different products and the strength of association rules. In addition, supplementary data such as holiday calendars, regional weather, and competitor prices can be added as needed. The raw data undergoes standardized preprocessing: First, outliers are removed, such as negative sales, inventory records exceeding the warehouse's maximum capacity, or outliers clearly caused by system errors. Then, missing sales or inventory data is filled in using the mean or linear interpolation method between preceding and following time points. Sequences with a large number of consecutive missing data are marked as invalid and discarded. Next, continuous numerical values ​​such as sales, price, and inventory are normalized by scaling the values ​​to the [0,1] interval to eliminate the influence of units. Finally, categorized data such as product category, warehouse type, and promotion type are converted using one-hot encoding, mapping each category to a binary vector to obtain the initial feature vector for each node. Based on the actual supply chain network topology, each independent inventory management unit is defined as a graph node. Nodes are mainly divided into three categories: warehouse nodes, supplier nodes, and distribution center nodes. Each node is assigned a unique identifier and its static attributes are recorded, such as geographical coordinates, maximum storage capacity, unit warehousing cost, and daily processing capacity limit. Using all nodes as vertices, initial edge connections are established based on actual logistics routes and supply chain relationships: if there is a direct goods transfer route between two warehouses, an undirected edge is established between them; if there is a procurement relationship between a warehouse and a supplier, a directed edge is established from the supplier to the warehouse; if there is a distribution relationship between a warehouse and a distribution center, a bidirectional edge is established. For each edge, the weight is calculated comprehensively from three dimensions: first, based on the logistics timeliness factor, the reciprocal of the geographical distance between nodes is calculated; second, based on the sales coupling relationship, the Pearson correlation coefficient between the historical sales sequences of the regions to which the two nodes belong is calculated; third, based on transportation efficiency, the contribution value of logistics timeliness is calculated using an exponential decay function. The calculation results of the three dimensions are multiplied by three adjustable hyperparameters and then summed to obtain the final weight of the edge. The weights of all edges are organized into a weighted adjacency matrix according to node indices. Each element in the matrix represents the connection strength between the corresponding pair of nodes; if two nodes are not directly connected, the corresponding element is zero. The resulting set of nodes, the set of weighted edges, and the initial feature vectors of each node are then integrated. During integration, the index correspondence between node identifiers and feature vectors must be maintained, and the weighted adjacency matrix is ​​used as a description of the adjacency relationships of the graph. The final output is a structured heterogeneous supply chain graph.

[0019] Step 102: Use a graph attention network to extract features from the heterogeneous graph of the supply chain to obtain the dynamic representation vector of each node; In this embodiment, the original feature vector of each node is used as the initial input to the graph attention network. The original feature vector contains the node's sales statistics, inventory level, spatial location encoding, and product association features. All node features are stacked row-wise to form a feature matrix. The initial feature vector of each node is linearly transformed by a learnable weight matrix, mapping the original feature dimensions to a higher-dimensional hidden space, resulting in a transformed feature vector. This weight matrix is ​​continuously updated through backpropagation during training. For any two nodes connected by an edge in the graph, the attention coefficient between the center node and its neighboring nodes is calculated. Specifically, the transformed feature vectors of the two nodes are concatenated, and the concatenated result is multiplied by a learnable attention weight vector, then passed through the LeakyReLU activation function to obtain an unnormalized attention score. This process is repeated to calculate the attention scores between the center node and all its neighboring nodes. For all neighboring nodes of the central node, the attention scores obtained in step three are softmax normalized so that the sum of the attention coefficients of all neighboring nodes equals 1. The normalized coefficients represent the importance of each neighboring node when aggregating neighbor information. The transformed feature vector of each neighboring node is multiplied by its corresponding normalized attention coefficient to obtain the weighted neighbor features. Then, the weighted features of all neighbors are summed and the central node's own features are added to obtain a new feature vector for the central node. This process is repeated 8 times, each time using independent attention weight parameters, to obtain node feature vectors from multiple different perspectives. These multi-head output feature vectors are concatenated to obtain the final dynamic representation vector of the node. The operations of the single-layer graph attention network are stacked into 3 layers, with the input of each layer being the output of the previous layer. Through multi-layer stacking, each node can aggregate information from higher-order neighbors, expanding its receptive field. After passing through the multi-layer graph attention network, the final feature vectors of all nodes are output as the dynamic representation of each inventory node in subsequent steps.

[0020] Step 103: Input the dynamic representation vector of each node into the TCN-Transformer model. First, extract short-term temporal dependency features through the TCN layer, and then capture long-term correlation features across cycles through the Transformer encoder. After feature fusion, generate a multi-step predicted sales sequence. In this embodiment, for each node, its dynamic representation vector over a past period is arranged chronologically to form an input sequence window. The input at each time step includes the node representation vector at that time step and the node's historical sales value. The input sequence is fed into the TCN module. The TCN module consists of multiple stacked dilated causal convolutional layers. The first convolutional layer uses a kernel with a dilation rate of 1 to convolve the sequence, ensuring that the convolution operation only uses information from the current and previous time steps. The second layer uses a kernel with a dilation rate of 2, the third layer uses a kernel with a dilation rate of 4, and so on. Each convolutional layer is followed by an activation function and batch normalization. By stacking convolutional layers with different dilation rates, the receptive field of the TCN continuously expands, enabling it to capture short-term fluctuation patterns within a few recent time steps. Finally, the TCN outputs a short-term temporal feature tensor. The temporal position information in the input sequence is encoded to generate a position encoding vector with the same length as the input sequence. The position encoding is added to the original features of the input sequence, enabling the model to perceive the temporal order. The input sequence with added positional encoding is fed into the Transformer encoder. The encoder consists of multiple stacked identical layers, each containing two sub-layers: a multi-head self-attention layer and a feedforward neural network layer. The multi-head self-attention layer operates as follows: for each time step in the sequence, it calculates the attention weights between that time step and all other time steps in the sequence, thereby capturing long-term dependencies across cycles, such as associations between seasonally similar periods. The feedforward neural network layer performs a non-linear transformation on the attention output of each time step. Residual connections and layer normalization are applied after each layer. After repeating multiple layers, a long-term global feature tensor is output. The short-term temporal features output by the TCN are then compared with the input of the Transformer encoder. The long-term global features are adaptively fused. The fusion method is as follows: a learnable weight coefficient, β, is calculated for the short-term features and the long-term features respectively. The short-term features are multiplied by β, and the long-term features are multiplied by (1-β). Then, the two weighted feature tensors are added element by element. The β parameter automatically learns the optimal balance during model training. The fused feature tensor is input into a fully connected prediction network, usually two fully connected layers. This network maps the features to the predicted sales values ​​for multiple future time steps. The number of neurons in the output layer is equal to the prediction step size, and each neuron outputs the predicted sales for a future day. The multi-step predicted sales results of each node are organized into sequence data according to the node identifier and the prediction time window.

[0021] Step 104: Input the dynamic representation vectors of each node and the multi-step predicted sales sequence into a reinforcement learning framework that includes an upper-level global coordinating agent and a lower-level node-level agent to make inventory scheduling decisions for the current node. A causal graph structure is introduced to describe the causal relationships between inventory nodes. In this embodiment, a hierarchical reinforcement learning framework is constructed: a global coordinating agent is set at the upper layer, and a node-level agent is set at the lower layer for each inventory node; the global coordinating agent is responsible for perceiving the state of the entire supply chain network and outputting global constraints; each node-level agent makes its own inventory scheduling action based on its local state under global constraints; the global coordinating agent receives the dynamic representation vectors of all nodes and the multi-step predicted sales sequence to form a global state space; the global coordinating agent integrates this information into a global state vector. The global coordinating agent makes decisions based on the MAPPO algorithm and outputs a set of global constraints, including: the lower bound of total inventory in each region, the upper limit of total quota for cross-regional transfers, the replenishment priority coefficient of each node, and the safety stock multiple adjustment factor of each node. The specific algorithm process for the constraint output is as follows: the policy network calculates the probability distribution of each constraint parameter based on the current global state, and then samples or takes the mean from the distribution to obtain the specific value. In order to ensure training stability, confidence interval reduction is used to limit the KL divergence between the new policy and the old policy to not exceed a preset threshold. Each node-level agent receives the following information as its local state: the dynamic representation vector of this node, the predicted sales volume of this node in the next multiple steps, the current inventory of this node, the replenishment volume in transit, the number of backlogged orders, and the part of the global constraints received from the global coordinating agent that is relevant to this node. Each node-level agent uses the DDPG algorithm for decision-making. The specific process of the algorithm is as follows: the policy network takes the local state vector as input and outputs an action vector, which contains multiple dimensions: the quantity ordered from the supplier, the quantity requested from the upper-level warehouse, the quantity shipped to the lower-level warehouse, and the proportion of inventory available for sale at this node. In order to ensure that the action is within a reasonable range, the output layer uses the tanh activation function to scale the action to [-1,1], and then maps it to the actual quantity range according to the business rules. At the same time, the network input local state and action are evaluated, and the value assessment Q value of the action is output to guide the update of the policy network. A causal graph is pre-constructed to describe the causal relationships between inventory nodes. The construction process involves collecting historical inventory level sequences, demand sequences, and scheduling action sequences for each node. A PC algorithm is used to test the conditional independence of each pair of variables. If the change in variable X causes a change in variable Y independently of other variables, a directed edge is added from X to Y. The causal graph may contain causal relationships such as "inventory level → sales volume" and "replenishment action → inventory level." During the policy update process of the node-level agent, it is calculated whether the current action violates the causal dependencies specified in the causal graph. The specific algorithm involves processing the actions generated by the policy network through... The differential method simulates the impact of actions on downstream variables such as inventory levels and sales, and calculates whether the direction of this impact is consistent with the direction of the edges in the causal graph. If they are inconsistent, a causal consistency loss is calculated. For example, actions that violate the causal direction are given a positive penalty. This loss is multiplied by a regularization coefficient and then added to the policy gradient, thereby forcing the policy network to favor causal consistent actions during gradient descent updates. The global constraints output by the global coordinating agent and the actions output by all node-level agents are combined into a complete scheduling decision scheme, including: the replenishment schedule of each node, the replenishment quantity, the transfer instructions between warehouses, and the transportation route selection suggestions.

[0022] In this embodiment, the global coordinating agent receives two types of input data: one is the dynamic representation vector of all nodes, and the other is the multi-step predicted sales sequence of each node. These two types of data are concatenated in node order to form a global state vector. This state vector contains the current inventory characteristics, sales forecast information, and inter-node relationship information of all nodes in the entire supply chain network. Two neural networks are established: a policy network and a value network. The policy network outputs the action probability distribution, and the value network evaluates the value of the current state. Both networks adopt a multi-layer fully connected structure. Network parameters are initialized, and an experience replay buffer is set to store historical interaction data. A hyperparameter for the confidence interval constraint is defined: the KL divergence threshold, which limits the degree of difference between the new and old policies during each policy update. Conventional training parameters such as the maximum update steps and learning rate are set. In the current global state, the policy network outputs an action based on the input. This action includes a vector of global constraint parameters, such as: the lower bound of total inventory in each region, the upper limit of total cross-regional allocation quotas, the replenishment priority coefficient for each node, and the safety stock multiple adjustment factor for each node. After executing this action, the environment returns a reward value and the next global state, storing the tuple in the experience replay buffer. A batch of experience data is sampled from the experience replay buffer. For each sampled experience, the value network is used to evaluate the value of the current state and the value of the next state. The temporal difference error is calculated, and then the weighted sum of the temporal difference errors from multiple time series is obtained using a generalized dominance estimation method to obtain the dominance function value for each time step. The dominance function represents the degree of advantage of taking an action in a given state compared to the average level. For each sampled experience, the policy parameters used when that experience was collected are recorded. The probability of taking an action in a given state is calculated using the old policy, and the same probability is calculated under the new policy parameters. The probability ratio is then calculated, and the KL divergence between the new and old policies is calculated. The KL divergence measures the degree of difference between the two probability distributions. The KL divergence is compared to a preset threshold. The MAPPO objective function introduces a penalty term with confidence interval constraints: if the KL divergence exceeds the threshold, a penalty proportional to the KL divergence is subtracted from the objective function, and a penalty coefficient is defined and dynamically adjusted, increasing as the KL divergence continues to exceed the threshold. The gradient of the improved objective function with respect to the policy network parameters is calculated, and the policy network parameters are updated using gradient ascent or the Adam optimizer to increase the objective function. Mini-batch gradient descent is used during updates, with multiple iterations to fully learn the current batch of data. The loss function of the value network is calculated, and the network parameters are updated using gradient descent. After each policy update, the relationship between the actual KL divergence and the threshold is checked: if the KL divergence exceeds 1.5 times the threshold, the penalty coefficient is increased; if the KL divergence is less than 0.5 times the threshold, it is decreased; otherwise, the penalty coefficient remains unchanged. This mechanism ensures that the policy update remains stable within a preset confidence interval. After multiple training iterations, the policy network converges. In the actual scheduling and execution phase, the current global state is input into the policy network, which directly outputs actions, i.e., a global constraint parameter vector. These parameter values ​​are converted into business-understandable constraints, such as: total inventory in East China should not be less than 5,000 units, and the total amount of cross-regional transfers should not exceed 2,000 units, etc., and passed to the lower-level node-level agents. During online operation, the policy network is retrained periodically using newly accumulated experience data, enabling the global coordinating agent to adapt to changes in market demand.

[0023] In this embodiment, each node-level agent receives two types of input information: first, the global constraints output by the global coordinating agent, including the lower bound of the local inventory, the upper limit of the allocation quota, and the replenishment priority coefficient; second, the local observation state of the local node, including the local node's dynamic representation vector, the future multi-step predicted sales sequence, the current actual inventory, the replenishment quantity in transit, the number of unprocessed backlogged orders, and the logistics timeliness fluctuation factor. These two types of information are concatenated into a local state vector. Four neural networks are established for each node-level agent: the current policy network for outputting deterministic actions, the current evaluation network for evaluating the value of actions, the target policy network, and the target evaluation network. All network parameters are initialized, with the initial parameters of the target network being the same as those of the current network. An experience replay buffer is set up to store historical interaction samples. Load a pre-built causal graph structure, which describes the causal relationships between inventory nodes and between inventory nodes and demand. For example, the current node's inventory level → the current node's sales volume, the current node's replenishment action → the current node's inventory level, and the upstream node's transfer action → the current node's inventory in transit. When making decisions, the agent needs to be able to query the directed edges related to the current node in the causal graph and their causal directions. The causal graph is stored in the form of an adjacency list or matrix. The decision action of the node-level agent is a multi-dimensional continuous vector, which includes: the quantity ordered from the supplier, the quantity requested from the upstream warehouse, the quantity shipped to the downstream warehouse, and the proportion of the current node's available inventory release. The output range of all action dimensions is limited by activation functions and business mapping functions to ensure that the output values ​​are within reasonable business boundaries. During online decision-making, the current local state vector is input into the current policy network, which directly outputs a deterministic action vector. To increase exploration, noise related to the training time can be added to the output action, with the noise amplitude decreasing as the number of training steps increases. The final action is the inventory scheduling instruction for this node. After executing the action, the environment returns a reward value and the next local state. The reward is equal to the negative inventory cost, including warehousing cost, stockout cost, replenishment cost, and transshipment cost. The tuple is stored in the experience replay buffer. When the buffer is full, old samples are overwritten according to the first-in, first-out principle. During the training phase, a batch of samples is randomly drawn from the experience replay buffer. Random sampling breaks the temporal correlation between samples, improving training stability. For each sample, the following operations are performed: input the next state from the sample into the target policy network to output the next action; then input them together into the target evaluation network to output the target Q-value; finally, calculate the target Q-value. Using the target network instead of the current network to estimate the Q-value can mitigate oscillations during the learning process. Input the current state and action from the sample into the current evaluation network to obtain the current Q-value; calculate the error between the current Q-value and the target value; update the parameters of the current evaluation network using gradient descent to gradually approach the target value. The update objective of the policy network is to maximize the Q-value output by the evaluation network. Input the current state from the sample into the current policy network to output the action; then input it into the current evaluation network to obtain the Q-value; negate the Q-value because the optimization direction is to maximize the Q-value, while gradient descent usually minimizes the loss. Calculate the gradient of the policy network parameters and update the policy network parameters using gradient ascent or by minimizing -Q. When updating the current policy network, causal graph constraints are added to prevent policies from violating causal dependencies; the specific process is as follows: Causal forward simulation, for the current state in the sampled sample, uses the action output by the current policy network to simulate the changes in key variables of the current node and related nodes after the action is executed; the simulation is based on simplified inventory dynamics equations and the transitive relationships defined in the causal graph; To check causal consistency, for each directed edge in the causal graph, compare whether the direction of change of variable Y in the simulation results matches the causal expectation with the direction of change of X. For example, if the causal graph stipulates that replenishment action at this node → inventory level at this node, then an increase in replenishment action should lead to an increase in inventory level; if the simulation results show that an increase in replenishment action leads to a decrease in inventory level, then it is judged as a violation of causal consistency. Calculate the causal consistency loss; accumulate a penalty value for all edges that violate causal consistency; the magnitude of the penalty is proportional to the degree of violation, for example, based on the absolute value of the difference between the simulated effect sign and the expected sign; sum the penalties for each edge to obtain the causal consistency loss; The causal loss is superimposed on the policy update gradient. In the gradient update of the policy network, in addition to maximizing the Q value, the causal consistency loss is minimized at the same time. The causal loss is multiplied by a regularization coefficient, which can be dynamically adjusted. The initial value is a small value such as 0.01. Then, the weighted loss is superimposed on the loss function of the policy network. The two objectives are optimized at the same time through gradient descent, so that the policy network can learn high-reward actions while avoiding violating the causal graph constraints as much as possible. To ensure stable training, the parameters of the current network are not directly copied to the target network. Instead, a soft update method is used: the parameters of the target policy network and the target evaluation network are updated to the current network parameters at a small rate using a moving average. This allows the target network to change slowly, improving training stability. After training convergence or during online execution, the current local state is input into the current policy network, which directly outputs a deterministic action as the final inventory scheduling decision for this node. This decision includes the order quantity, allocation quantity, shipment quantity, and inventory release ratio, and is then sent to the execution system. During online operation, the node-level agent is retrained periodically using the most recently accumulated experience samples. The steps are repeated during retraining, and the causal graph may also be dynamically updated based on the latest data. Through continuous training, the node-level agent can adapt to market changes and shifts in demand patterns.

[0024] Step 105: Distribute the inventory scheduling decision to each inventory node for execution, and collect the changes in inventory status and order fulfillment results in real time for iteration.

[0025] In this embodiment, the scheduling decision plan is sent to the execution system of each inventory node through a data interface. The execution system includes a supplier order system, a warehouse management system, and a transportation management system. The distribution method can be a message queue or API call to ensure that each node receives a clear instruction. During the execution of the scheduling instructions at each inventory node, real-time status change data is collected, including: actual replenishment arrival time and quantity, actual transfer completion time and quantity, real-time inventory level change curves of each node, and order fulfillment status. Based on the collected execution results, the core indicators for each scheduling cycle are calculated: actual warehouse holdings, actual stockouts, actual replenishment costs, and actual inventory levels. The system calculates actual transportation costs; compares these indicators with the expected values ​​at the time of scheduling decision-making to form a deviation vector; stores the initial state, scheduling decision, execution results, and feedback indicators of this scheduling cycle as a complete time series sample in a multi-source heterogeneous database, covering historical sales data and inventory status data at the corresponding time points; this database is read in the next cycle; after completing all execution and data collection for the current scheduling cycle, the system automatically starts the next scheduling cycle, returns to step 101, and uses the updated multi-source heterogeneous data to reconstruct the graph, extract features, predict, make decisions, and execute; this cycle repeats continuously, forming a continuous closed loop of prediction-decision-scheduling-feedback; During the monitoring process, if a sudden abnormal event is detected, such as the sales volume of a certain node exceeding the normal fluctuation threshold in the past hour, or a supplier reporting stockouts or logistics trunk line interruptions, the emergency scheduling mode is immediately triggered: the normal scheduling process of the current cycle is suspended, the reinforcement learning framework in step 104 is directly called, but the optimization objective is temporarily adjusted to minimize stockout costs, and a rapid rescheduling plan is generated and issued for execution in a short time. The normal cycle scheduling is resumed after the abnormality subsides.

[0026] Please see Figure 2 The present invention provides a schematic diagram of the structure of an AI-based e-commerce inventory dynamic scheduling and management system, which includes: The module is used to collect multi-source heterogeneous data from e-commerce platforms and construct a supply chain heterogeneous graph including warehouse nodes, supplier nodes, and distribution center nodes based on the multi-source heterogeneous data. The feature extraction module is used to extract features from the heterogeneous graph of the supply chain using a graph attention network to obtain the dynamic representation vector of each node. The prediction module is used to input the dynamic representation vector of each node into the TCN-Transformer model. First, the TCN layer extracts short-term temporal dependency features, and then the Transformer encoder captures long-term correlation features across cycles. After feature fusion, a multi-step predicted sales sequence is generated. The decision module is used to input the dynamic representation vectors of each node and the multi-step predicted sales sequence into a reinforcement learning framework that includes an upper-level global coordinating agent and a lower-level node-level agent to make inventory scheduling decisions for the current node. A causal graph structure is introduced to describe the causal relationships between inventory nodes. The execution module is used to distribute inventory scheduling decisions to each inventory node for execution, and to collect and iterate on the changes in inventory status and order fulfillment results after execution in real time.

[0027] Figure 3This is a schematic diagram of the structure of an AI-based e-commerce inventory dynamic scheduling and management device 300 provided in an embodiment of the present invention. The AI-based e-commerce inventory dynamic scheduling and management device 300 can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 310 (e.g., one or more processors) and a memory 320, and one or more storage media 330 (e.g., one or more mass storage devices) for storing application programs 333 or data 332. The memory 320 and storage media 330 can be temporary or persistent storage. The program stored in the storage media 330 may include one or more modules (not shown in the diagram), each module may include a series of instruction operations on the AI-based e-commerce inventory dynamic scheduling and management device 300. Furthermore, the processor 310 may be configured to communicate with the storage media 330 and execute the series of instruction operations in the storage media 330 on the AI-based e-commerce inventory dynamic scheduling and management device 300 to implement the method provided in the above embodiment.

[0028] The AI-based e-commerce inventory dynamic scheduling and management device 300 may also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input / output interfaces 360, and / or one or more operating systems 331, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 3 The structure of the AI-based e-commerce inventory dynamic scheduling and management device shown does not constitute a limitation on the computer device provided by this invention. It may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0029] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the various steps of the artificial intelligence-based e-commerce inventory dynamic scheduling and management method provided in the above embodiments.

[0030] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the above-described equipment or apparatus / unit can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0031] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0032] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An artificial intelligence-based method for dynamic scheduling and management of e-commerce inventory, characterized in that, The method includes the following steps: Collect multi-source heterogeneous data from e-commerce platforms, and construct a supply chain heterogeneous graph including warehouse nodes, supplier nodes, and distribution center nodes based on the multi-source heterogeneous data; A graph attention network is used to extract features from the heterogeneous supply chain graph to obtain the dynamic representation vector of each node; The dynamic representation vectors of each node are input into the TCN-Transformer model. First, short-term temporal dependency features are extracted through the TCN layer, and then long-term correlation features across cycles are captured through the Transformer encoder. After feature fusion, a multi-step predicted sales sequence is generated. The dynamic representation vectors of each node and the multi-step predicted sales sequence are input into a reinforcement learning framework that includes an upper-level global coordinating agent and a lower-level node-level agent to make inventory scheduling decisions for the current node. A causal graph structure is introduced to describe the causal relationships between inventory nodes. The inventory scheduling decision is sent to each inventory node for execution, and the changes in inventory status and order fulfillment results are collected in real time for iteration.

2. The AI-based e-commerce inventory dynamic scheduling and management method as described in claim 1, characterized in that, The process involves collecting multi-source heterogeneous data from e-commerce platforms and constructing a supply chain heterogeneous graph based on this data, including warehouse nodes, supplier nodes, and distribution center nodes. Collect multi-source heterogeneous data from e-commerce platforms, including at least historical sales time-series data, price and promotional activity data, and product-related purchase data for each inventory node; The collected raw multi-source heterogeneous data is cleaned, and the categorized data is converted by one-hot encoding to obtain the initial feature vector of each node. The data cleaning includes removing outliers, filling missing values, and normalization. Based on the physical topology of the e-commerce inventory system, each independent inventory management unit is defined as a node, where each node includes at least a warehouse node, a supplier node, and a distribution center node. Using all nodes as vertices, initial edge connections are established based on actual logistics paths and supply chain relationships. For each edge, a weight is calculated based on logistics timeliness factors, sales coupling relationships, and transportation efficiency. The calculated weights are multiplied by preset adjustable coefficients to obtain the final weight of the edge. The weights of all edges are then used to construct a weighted adjacency matrix. The node set, the weighted edge set, and the initial feature vectors of each node are integrated to form a supply chain heterogeneous graph.

3. The AI-based e-commerce inventory dynamic scheduling and management method as described in claim 1, characterized in that, The graph attention network is used to extract features from the heterogeneous supply chain graph to obtain the dynamic representation vector of each node, including: The heterogeneous graph of the supply chain is input into the graph attention network, and the initial feature vector of each node is linearly transformed to obtain the transformed feature vector. For each attention head, the attention coefficient is calculated for the central node and its neighboring nodes connected by any edge in the supply chain heterogeneous graph, and the attention coefficient is normalized using softmax. Multiply the transformed feature vector of each neighbor node by the corresponding normalized attention coefficient to obtain the weighted neighbor features; Sum the weighted features of all neighbors and add the features of the center node to obtain the new feature vector of the current center node; The operation is repeated with 8 attention heads, and the feature vectors output by each attention head are concatenated to obtain the final dynamic representation vector of the current center node. The operations of a single-layer graph attention network are stacked into three layers, with the input of each layer being the output of the previous layer. Through multiple stacking layers, the final dynamic feature vectors of all nodes are output after passing through the multi-layer graph attention network, thus obtaining the dynamic representation vector of each node.

4. The AI-based e-commerce inventory dynamic scheduling and management method as described in claim 1, characterized in that, The dynamic representation vectors of each node are input into the TCN-Transformer model. First, short-term temporal dependency features are extracted through the TCN layer, and then long-term correlation features across periods are captured through the Transformer encoder. After feature fusion, a multi-step predicted sales sequence is generated, including: The dynamic representation vector of each node is arranged in chronological order to form an input sequence window, which is then input into the TCN-Transformer model. The TCN module consists of multiple stacked dilated causal convolutional layers. Each convolutional layer is processed by an activation function and batch normalization. By stacking convolutional layers with different dilation rates, it captures short-term fluctuation patterns and outputs short-term time series features. The input sequence with added position encoding is fed into the Transformer encoder, which consists of multiple identical layers stacked together, each layer including a multi-head self-attention sublayer and a feedforward neural network sublayer; The multi-head self-attention sublayer calculates attention weights for each time step in the input sequence, capturing long-term dependencies across cycles. The feedforward neural network sublayer performs a non-linear transformation on the attention output at each time step, outputting long-term global features. Short-term temporal features and long-term global features are fused to obtain fused features. The input is then processed by a fully connected layer, which maps the features to sales forecasts for multiple future time steps, and integrates them to obtain a multi-step predicted sales sequence.

5. The AI-based e-commerce inventory dynamic scheduling and management method as described in claim 1, characterized in that, The process of inputting the dynamic representation vectors of each node and the multi-step predicted sales sequence into a reinforcement learning framework that includes an upper-level global coordinating agent and lower-level node-level agents to make inventory scheduling decisions for the current node includes: The upper-level global coordinating agent receives the dynamic representation vectors and multi-step predicted sales sequences from all nodes, forming a global state space, and outputs global constraints based on the MAPPO algorithm. The lower-level node-level agent receives global constraints and local observation states of each node, and uses the DDPG algorithm to output the decision actions of each node. The local observation states include at least the dynamic representation vector, multi-step predicted sales sequence, current inventory, in-transit replenishment quantity, and order backlog. The global constraints output by the upper-level global coordinating agent and the decision actions output by all lower-level node-level agents are combined to form the final inventory scheduling decision.

6. The AI-based e-commerce inventory dynamic scheduling and management method as described in claim 5, characterized in that, The upper-layer global coordination agent receives the dynamic representation vectors and multi-step predicted sales sequences from all nodes, forming a global state space, and outputs global constraints based on the MAPPO algorithm, including: In the current global state space, the policy network of the upper-level global coordinating agent executes the action of the global constraint parameter vector according to the input and output global constraint parameter vector. The environment returns the reward value and stores the tuple in the experience replay buffer. Experience data is sampled from the experience replay buffer. For each sampled experience data, the value of the current state and the value of the next state are evaluated using a value network. The temporal difference error is calculated. The temporal difference errors of multiple time series are weighted and summed using a generalized dominance estimation algorithm to obtain the dominance function value for each time step. Obtain the policy parameters used when sampling each empirical data point, calculate the probability of taking an action in a given state using the old policy, calculate the probability under the current new policy parameters, and calculate the probability ratio. Calculate the KL divergence between the new and old policies, construct the MAPPO objective function based on the probability ratio, and update the policy network parameters using the Adam optimizer; The loss function of the value network is calculated, and the parameters of the value network are updated by gradient descent. After multiple training iterations, the policy network converges and the global constraints are directly applied.

7. The AI-based e-commerce inventory dynamic scheduling and management method as described in claim 5, characterized in that, The lower-level node-level agent receives global constraints and the local observation states of each node, and outputs the decision actions of each node using the DDPG algorithm, including: Obtain historical inventory level sequences, demand sequences, and scheduling action sequences for each node, use the PC algorithm to perform conditional independence tests, and construct a cause-effect graph; During the policy update process of lower-level node agents, it is calculated whether the current action violates the causal dependency relationship in the causal graph. If violated, the causal consistency loss is calculated, multiplied by the regularization coefficient, and then added to the policy gradient. During gradient descent updates, this causes the policy network to favor causal consistent actions.

8. An AI-based e-commerce inventory dynamic scheduling and management system, characterized in that: The system includes: The module is used to collect multi-source heterogeneous data from e-commerce platforms and construct a supply chain heterogeneous graph including warehouse nodes, supplier nodes, and distribution center nodes based on the multi-source heterogeneous data. The feature extraction module is used to extract features from the heterogeneous supply chain graph using a graph attention network to obtain the dynamic representation vector of each node. The prediction module is used to input the dynamic representation vector of each node into the TCN-Transformer model. First, the TCN layer extracts short-term temporal dependency features, and then the Transformer encoder captures long-term correlation features across cycles. After feature fusion, a multi-step predicted sales sequence is generated. The decision module is used to input the dynamic representation vectors of each node and the multi-step predicted sales sequence into a reinforcement learning framework that includes an upper-level global coordinating agent and a lower-level node-level agent to make inventory scheduling decisions for the current node. A causal graph structure is introduced to describe the causal relationships between inventory nodes. The execution module is used to distribute inventory scheduling decisions to each inventory node for execution, and to collect changes in inventory status and order fulfillment results in real time for iteration.

9. An AI-based e-commerce inventory dynamic scheduling and management device, characterized in that, The AI-based e-commerce inventory dynamic scheduling and management device includes a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the AI-based e-commerce inventory dynamic scheduling and management device to execute each step of the AI-based e-commerce inventory dynamic scheduling and management method as described in any one of claims 1-7.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement each step of the AI-based e-commerce inventory dynamic scheduling and management method as described in any one of claims 1-7.

Images