Inventory control method, electronic device, and machine-readable storage medium

Abstract

This application discloses an inventory control method, electronic device, and machine-readable storage medium, relating to the field of supply chain management technology. The method includes: acquiring multi-source heterogeneous data of a target inventory object; generating a causal-enhanced state representation by performing causal enhancement processing on the original state representation corresponding to the multi-source heterogeneous data; enabling the decision-making process to have risk perception capabilities by acquiring a measure of the demand uncertainty of the target inventory object within a preset future period; fusing the causal-enhanced state representation, the demand uncertainty measure, and real-time inventory state data to construct an inventory decision feature set, achieving the organic integration of multi-dimensional information; and outputting replenishment decision values ​​for the target inventory object through a pre-trained inventory decision model based on the inventory decision feature set, thereby improving the generalization ability, risk adaptability, and decision quality of the inventory control method, and reducing inventory costs while ensuring service levels.

Description

Technical Field

[0001] This application relates to the field of supply chain management technology, and more specifically to an inventory control method, electronic device, and machine-readable storage medium. Background Technology

[0002] Inventory management is a core component of a company's supply chain operations, and accurately predicting inventory level changes is crucial for ensuring service levels and reducing operating costs. However, inventory changes in the modern supply chain environment are influenced by a variety of factors, including historical sales patterns, promotional activities, holiday effects, and supplier delivery cycles. These factors often have complex interactions and time-varying characteristics, making accurate inventory level prediction a significant challenge.

[0003] Traditional inventory management methods primarily rely on statistical time-series forecasting models or replenishment strategies based on empirical rules. In recent years, some research has attempted to introduce machine learning models for demand forecasting. However, these methods still have significant shortcomings in application: the forecasting and decision-making processes are disconnected, leading to amplified cumulative errors; the models mainly learn correlation patterns from historical data, making it difficult to cope with structural changes in the market environment; and the lack of quantitative consideration of forecast uncertainty prevents the making of decisions with appropriate risk. Therefore, how to effectively integrate multi-source heterogeneous data, accurately capture the key factors affecting inventory changes, and formulate reasonable replenishment strategies accordingly has become a pressing technical problem for those skilled in the art. Summary of the Invention

[0004] In view of the above-mentioned shortcomings of the prior art, the purpose of this application is to provide an inventory control method, an electronic device, and a machine-readable storage medium.

[0005] To achieve the above objectives, the first aspect of this application provides an inventory control method, comprising: Acquire multi-source heterogeneous data of the target inventory object, including historical time-series data, event data, and supply chain data; Causal augmentation processing is performed on the original state representations corresponding to multi-source heterogeneous data to generate causal augmented state representations; Obtain the demand uncertainty measure of the target inventory object within a future preset period, where the demand uncertainty measure is used to characterize the probability distribution characteristics of the demand quantity within the future preset period; By fusing causal-enhanced state representation, demand uncertainty measurement, and real-time inventory status data, an inventory decision feature set is constructed. Based on the inventory decision feature set, the replenishment decision value of the target inventory object is output through a pre-trained inventory decision model.

[0006] In this embodiment of the application, causal enhancement processing is performed on the original state representation corresponding to multi-source heterogeneous data to generate a causal enhanced state representation, including: Based on preset inventory control business rules, key variables related to inventory changes are extracted from multi-source heterogeneous data to form a set of key variables; Based on the conditional independence test method, the causal relationships among the key variables in the key variable set are determined, and a causal relationship diagram is generated; Causal intervention is performed on multi-source heterogeneous data based on causal relationship graphs to generate causal-enhanced state representations.

[0007] In this embodiment of the application, causal intervention is performed on multi-source heterogeneous data based on a causal relationship graph to generate a causal-enhanced state representation, including: Based on the causal relationship diagram, identify the treatment variables and outcome variables among the key variables; Based on the pointing relationships between key variables in the causal relationship diagram, target key variables that simultaneously point to both processing variables and outcome variables are identified as confounding factors. The confounding factors are adjusted through intervention to generate post-intervention features, which are used to characterize the causal effect of the treatment variable on the outcome variable after removing the influence of the confounding factors. By fusing the original state representations corresponding to multi-source heterogeneous data with the post-intervention features, a causal-enhanced state representation is obtained.

[0008] In this embodiment of the application, based on the inventory decision feature set, a pre-trained inventory decision model outputs the replenishment decision value of the target inventory object, including: The inventory decision feature set is input into the pre-trained inventory decision model. The inventory decision model aims to maximize long-term cumulative returns and outputs the replenishment decision value of the target inventory object.

[0009] In this embodiment of the application, the inventory control method further includes: Based on replenishment decision values, real-time inventory status data, and demand uncertainty measures, the Monte Carlo simulation method is used to generate the inventory change trajectory within a future preset period. The probability of stockouts for a target inventory item within a future preset period is determined based on inventory change trajectories.

[0010] In this embodiment of the application, obtaining a measure of the demand uncertainty of the target inventory object within a future preset period includes: Using a pre-trained probabilistic prediction model, a measure of the demand uncertainty of a target inventory object within a future preset period is generated based on multi-source heterogeneous data. The measure of demand uncertainty includes the expected value and standard deviation of the demand within the future preset period.

[0011] In this embodiment of the application, the causal enhanced state representation, the demand uncertainty metric, and the real-time inventory status data are fused to construct an inventory decision feature set, including: Based on causal enhanced state representation, a sliding window sequence of target inventory objects is constructed according to a preset time window; Temporal features are extracted from the sliding window sequence and fused with the causal augmented state representation to obtain the temporal feature tensor; By fusing time-series feature tensors, demand uncertainty metrics, and real-time inventory status data, an inventory decision feature set is constructed.

[0012] In this embodiment of the application, before the step of performing causal enhancement processing on the original state representation corresponding to multi-source heterogeneous data to generate a causal enhanced state representation, the method further includes: Preprocessing of multi-source heterogeneous data includes at least one of data cleaning, missing value imputation, and outlier detection and replacement.

[0013] A second aspect of this application provides an electronic device, comprising: The memory is configured to store instructions; The processor is configured to retrieve instructions from memory and, when executing the instructions, to implement the inventory control method as described in the above embodiments.

[0014] A third aspect of this application provides a machine-readable storage medium storing instructions that cause a machine to perform the inventory control method as described in the above embodiments.

[0015] By employing the aforementioned technical solutions, multi-source heterogeneous data of the target inventory object is obtained, providing a comprehensive information foundation for subsequent processing to characterize the factors influencing inventory, thus solving the problems of single data sources and incomplete information in traditional methods. Causal enhancement processing is applied to the original state representations corresponding to the multi-source heterogeneous data to generate causal enhanced state representations, effectively removing spurious correlations caused by confounding factors and addressing the issues of existing models' inability to cope with environmental changes and the tendency for historical correlation-based learning to fail. By obtaining the demand uncertainty measure of the target inventory object within a preset future period, the decision-making process gains risk perception capabilities, solving the problem that traditional deterministic prediction cannot provide a basis for risk-based decision-making. The causal enhanced state representations, demand uncertainty measures, and real-time inventory state data are integrated to construct an inventory decision feature set, achieving organic integration of multi-dimensional information and solving the problems of fragmented decision information and lack of unified representation. Based on the inventory decision feature set, a pre-trained inventory decision model outputs replenishment decision values ​​for the target inventory object, achieving end-to-end optimization from data to decision, and solving the problems of separation between prediction and decision-making and error accumulation amplification in traditional two-stage methods. This embodiment improves the generalization ability, risk adaptability, and decision quality of the inventory control method, reducing inventory costs while ensuring service levels.

[0016] Other features and advantages of the embodiments of this application will be described in detail in the following detailed description section. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the following detailed description to explain the embodiments of this application, but do not constitute a limitation on the embodiments of this application. In the drawings: Figure 1 The illustration shows a schematic flowchart of an inventory control method according to an embodiment of this application; Figure 2 A schematic block diagram of an electronic device according to an embodiment of this application is shown. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for illustration and explanation of the embodiments of this application and are not intended to limit the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0019] It should be noted that the acquisition, transmission, storage, use, and processing of data in the technical solution of this application all comply with relevant laws and regulations. In the embodiments of this application, certain existing industry solutions such as software, components, and models may be mentioned. These should be considered exemplary, intended only to illustrate the feasibility of implementing the technical solution of this application, and do not imply that the applicant has already used or necessarily used such solutions.

[0020] It should be noted that if the embodiments of this application involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.

[0021] Furthermore, if the embodiments of this application involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed in this application.

[0022] Figure 1 A schematic flowchart illustrating an inventory control method according to an embodiment of this application is shown. Figure 1 As shown in the figure, this application provides an inventory control method, which may include the following steps: Step 100: Obtain multi-source heterogeneous data of the target inventory object, wherein the multi-source heterogeneous data includes historical time-series data, event data, and supply chain data; In this embodiment, it should be noted that the target inventory object refers to the specific material unit that needs to be managed, such as a specific SKU (Stock Keeping Unit). Multi-source heterogeneous data refers to datasets from different dimensions, types, and with different data structures. High-quality model input is constructed through multi-source data fusion and feature engineering, laying the data foundation for subsequent safety stock calculations. It should be noted that in a real supply chain environment, factors affecting demand go far beyond historical sales figures. Events such as promotional activities, holiday arrangements, and weather changes, as well as supply chain status such as inventory levels and in-transit supplies, all significantly impact demand.

[0023] Specifically, historical time-series data is used to capture historical patterns of demand changes, such as sales records at each point in time over the past year or three years. This data reflects the cyclical, trend-based, and seasonal characteristics of demand. Examples include historical daily sales data, historical weekly purchase volume data, and historical monthly outbound volume data, which visually demonstrate the patterns of demand changes over time. Event data is used to introduce the potential impact of known future events on demand, such as upcoming promotional plans, statutory holiday arrangements, or large-scale marketing events. These events often lead to abnormal fluctuations in demand in the short term. Supply chain data reflects the real-time dynamics of the supply side, such as current inventory levels, the quantity of inventory in transit, and estimated delivery times. While this information does not directly determine demand, it affects actual salable inventory and thus indirectly impacts the ability to meet demand. For example, in a retail scenario, time-series data for a product might show increased sales every summer, event data might indicate a promotional event next week, and supply chain data might show a replenishment shipment arriving before the promotion begins. These three elements together constitute a complete input for predicting future demand for that product.

[0024] Understandably, in real-world enterprise environments, multi-source heterogeneous data typically originates from multiple business systems, such as ERP (Enterprise Resource Planning), WMS, SRM, TMS, and CRM (Customer Relationship Management). These systems record business data across different dimensions, including purchase orders, inventory status, supplier performance, in-transit shipments, and promotional plans, collectively forming the core source of multi-source heterogeneous data. Furthermore, external environmental data, such as holiday schedules and weather forecasts, can also be obtained through APIs (Application Programming Interfaces) as supplementary input for event data.

[0025] This embodiment integrates relevant data from multiple dimensions and sources, breaking down the information barriers of single data sources. This allows subsequent inventory decisions to be based on more comprehensive information that is more in line with actual business operations, effectively avoiding decision-making errors caused by incomplete data, and providing complete and effective data support for subsequent status processing and model decision-making.

[0026] Step 200: Perform causal augmentation processing on the original state representations corresponding to multi-source heterogeneous data to generate causal augmented state representations; It should be noted that traditional inventory control directly uses raw data for decision-making. The raw state representation contains a large number of spurious correlations caused by confounding factors, which can easily lead the model to learn incorrect patterns and make unreasonable inventory decisions. Therefore, it is necessary to perform causal enhancement processing on the raw state representation to uncover the true causal relationships behind the data.

[0027] Specifically, the original state representation is a set of features used to characterize the original state of inventory objects after preliminary processing of multi-source heterogeneous data; it is a direct feature representation of multi-source heterogeneous data. Causal enhancement processing is a process of optimizing the original state representation through causal analysis techniques. Its core is to eliminate spurious correlations and retain true causal relationships. This process can be based on Bayesian structural learning methods to mine causal relationships between variables and then intervene and adjust; it can also be based on causal graphical models to identify confounding factors and then correct features; or it can be based on propensity score matching methods to eliminate confounding biases between variables and then reconstruct state features. The causal enhanced state representation is a set of features that, after causal enhancement processing, eliminates spurious correlations and accurately reflects the true causal relationships between the state variables of the inventory object. This feature set retains the effective information in the original state representation and also purifies the state features through causal processing.

[0028] Traditional methods rely on correlation for learning, which can easily fail when the environment changes. This embodiment uses causal reinforcement processing to free the feature basis of inventory decisions from the interference of spurious correlations, enabling subsequent decisions based on these features to align with the true causal logic of the business. This effectively improves the scientificity and rationality of subsequent inventory decisions, allowing the model to learn the core laws that truly affect inventory changes and providing reliable feature support for accurate inventory decisions.

[0029] Step 300: Obtain the demand uncertainty measure of the target inventory object in the future preset period, wherein the demand uncertainty measure is used to characterize the probability distribution characteristics of the demand in the future preset period. It should be noted that traditional inventory control methods rely solely on single-point demand forecasts for decision-making, ignoring the volatile nature of future demand. When actual demand deviates from the forecast, it can easily lead to stockouts or inventory backlogs. Therefore, it is necessary to obtain a measure of demand uncertainty and incorporate the volatile characteristics of demand into inventory decision-making.

[0030] Specifically, the future preset period is a pre-set demand forecast time range based on the actual business needs of enterprise inventory management. This can be 7 days, 15 days, or 30 days ahead, or it can be a future procurement cycle or a future peak sales season cycle. The demand uncertainty measure is an indicator used to quantify the degree of fluctuation and probability distribution characteristics of demand within the future preset period. The demand uncertainty measure is not limited to a single numerical form and can be characterized by multiple indicators. The demand uncertainty measure can be obtained by analyzing multi-source heterogeneous data through a probability prediction model, by statistically calculating based on the distribution characteristics of historical demand data, or by comprehensively evaluating industry demand fluctuation patterns and the enterprise's own business characteristics. In practical applications, the demand uncertainty measure can be a combination of the expected value and standard deviation of demand, a combination of the median and interquartile range of demand, or a confidence interval of demand under different probabilities.

[0031] This embodiment obtains a measure of demand uncertainty, enabling inventory decisions to fully consider the risk of future demand fluctuations. This changes the limitations of traditional single-point forecasting and allows subsequent inventory replenishment decisions to be adjusted according to the degree of demand uncertainty. This effectively improves the risk response capability of inventory decisions and reduces losses from stockouts or inventory backlogs caused by demand fluctuations.

[0032] Step 400: The causal enhancement state representation, demand uncertainty metric, and real-time inventory status data are fused to construct an inventory decision feature set; It should be noted that the causal reinforcement state representation, demand uncertainty measurement value and real-time inventory status data reflect relevant information of inventory objects from different dimensions. Using any one dimension of information alone cannot form a complete basis for decision-making. Therefore, it is necessary to integrate multi-dimensional information to construct an inventory decision feature set to provide comprehensive and integrated feature support for inventory decision-making.

[0033] Specifically, real-time inventory status data represents the current actual inventory status of an inventory object. This can include current on-hand inventory quantity, available inventory quantity, current inventory in transit quantity, inventory awaiting shipment quantity, inventory age, and capital tied up in inventory. Fusion is the process of integrating and processing feature information of different dimensions and types to form a unified feature set. This fusion process can be a simple concatenation of feature dimensions, or it can involve reconstructing features by mining the correlation information between features of different dimensions using feature cross-validation methods, or it can use dimensionality reduction algorithms to fuse and compress multi-dimensional features to form core features. The inventory decision feature set is a set of features that, after fusing multi-dimensional information, comprehensively reflects the causal state, demand fluctuation characteristics, and real-time inventory status of the inventory object. This inventory decision feature set is the direct input to the subsequent inventory decision model, and the feature dimensions and feature types of the inventory decision feature set can be adaptively adjusted according to the needs of the inventory decision model.

[0034] This embodiment integrates historical causal patterns, future demand risks, and current actual status of inventory objects through the fusion of multi-dimensional information, forming a comprehensive and complete set of inventory decision features. This allows subsequent inventory decision models to reason based on features that are more closely aligned with actual decision-making scenarios, effectively improving the effectiveness and comprehensiveness of the model's input features and laying a solid feature foundation for the model to output accurate replenishment decision values.

[0035] Step 500: Based on the inventory decision feature set, output the replenishment decision value of the target inventory object through the pre-trained inventory decision model.

[0036] It should be noted that manual inventory replenishment decisions are easily influenced by subjective experience and cannot quickly process complex inventory decision feature sets. Traditional inventory decision models mostly aim to minimize single-step prediction errors, making it difficult to optimize long-term inventory benefits. Therefore, it is necessary to make intelligent replenishment decisions based on pre-trained inventory decision models to improve the efficiency and scientific nature of the decisions.

[0037] Specifically, the inventory decision model is an intelligent model pre-trained with a large amount of historical data, capable of outputting reasonable replenishment decisions based on an inventory decision feature set. This inventory decision model can be a deep reinforcement learning model, a combined model integrating machine learning and operations research optimization, or a deep learning model based on attention mechanisms. The training objective of the inventory decision model can be set according to the enterprise's business needs, such as maximizing long-term cumulative revenue, minimizing long-term cumulative costs, or optimizing service levels. Pre-training means that the inventory decision model has been trained with historical multi-source heterogeneous data, causal augmented state representations, and inventory decision feature sets before actual business application. The parameters of the inventory decision model have been optimized and can be directly applied to actual inventory decision scenarios. The replenishment decision value is a decision indicator output by the inventory decision model based on the inventory decision feature set, used to guide the enterprise in inventory replenishment operations. It can be a specific replenishment quantity, a replenishment quantity range combined with replenishment timing suggestions, or a replenishment ratio combined with supplier selection suggestions.

[0038] In this embodiment, an intelligent inventory replenishment decision-making process is achieved through a pre-trained inventory decision-making model, replacing traditional manual decision-making and simple model-based decision-making. This model can quickly and accurately process complex inventory decision feature sets. At the same time, the inventory decision-making model makes decisions based on preset optimization objectives, which can achieve long-term benefit optimization of inventory management. This effectively improves the efficiency and accuracy of inventory replenishment decisions, allowing enterprises to better align their inventory control with the actual needs of business development and achieve a balanced optimization of inventory costs and service levels.

[0039] This embodiment acquires multi-source heterogeneous data of the target inventory object, providing a comprehensive information foundation for subsequent processing to characterize the factors influencing inventory, thus solving the problems of single data sources and incomplete information in traditional methods. By performing causal enhancement processing on the original state representations corresponding to the multi-source heterogeneous data, a causal enhanced state representation is generated, effectively removing spurious correlations caused by confounding factors, and solving the problems of existing models being unable to cope with environmental changes and the easy failure of learning based on historical correlations. By acquiring the demand uncertainty measure of the target inventory object in the future preset period, the decision-making process has risk perception capabilities, solving the problem that traditional deterministic prediction cannot provide a basis for risk decision-making. By fusing the causal enhanced state representation, demand uncertainty measure, and real-time inventory state data, an inventory decision feature set is constructed, realizing the organic integration of multi-dimensional information and solving the problems of fragmented decision information and lack of unified representation. Based on the inventory decision feature set, the replenishment decision value of the target inventory object is output through a pre-trained inventory decision model, realizing end-to-end optimization from data to decision, and solving the problems of separation between prediction and decision and error accumulation amplification in traditional two-stage methods. This embodiment improves the generalization ability, risk adaptability, and decision quality of inventory control methods, reducing inventory costs while ensuring service levels.

[0040] In one embodiment, causal augmentation processing is performed on the original state representation corresponding to multi-source heterogeneous data to generate a causal augmented state representation, including: Based on preset inventory control business rules, key variables related to inventory changes are extracted from multi-source heterogeneous data to form a set of key variables; Based on the conditional independence test method, the causal relationships among the key variables in the key variable set are determined, and a causal relationship diagram is generated; Causal intervention is performed on multi-source heterogeneous data based on causal relationship graphs to generate causal-enhanced state representations.

[0041] In this embodiment, it should be noted that the preset inventory control business rules are screening rules formulated based on the actual business scenarios and control needs of enterprise inventory management. These rules can be formulated based on the core influencing factors of inventory changes, the enterprise's replenishment strategy, inventory warning thresholds, and other business requirements, or in conjunction with industry-standard inventory management principles. Key variables are various data variables that have a direct or indirect impact on inventory changes. Extraction methods include manual screening through business rules or assisted screening through correlation analysis. The set of key variables can include variables related to inventory changes such as sales volume, promotion status, holidays, inventory quantity, and supplier delivery status. Conditional independence testing is a data analysis method used to determine whether a conditional independence relationship exists between variables. It can be a chi-square test, Fisher's Z test, or mutual information test. This conditional independence test can eliminate spurious associations between variables and determine the true causal relationship. A causal relationship diagram is a graphical representation of the causal relationship between key variables. It can be represented as a directed acyclic graph, where nodes represent key variables and directed edges represent the causal relationship between variables. Causal intervention is a feature adjustment operation based on causal relationship diagrams of the original data. Its core is to eliminate the interference caused by confounding factors and restore the true causal effect between variables.

[0042] This embodiment combines business rules and statistical tests to automatically discover the causal structure between variables, avoiding the subjectivity and incompleteness of manually specifying causal relationships, and providing a scientific basis for subsequent causal intervention.

[0043] In one embodiment, causal intervention is performed on multi-source heterogeneous data based on a causal relationship graph to generate a causal-enhanced state representation, including: Based on the causal relationship diagram, identify the treatment variables and outcome variables among the key variables; Based on the pointing relationships between key variables in the causal relationship diagram, target key variables that simultaneously point to both processing variables and outcome variables are identified as confounding factors. The confounding factors are adjusted through intervention to generate post-intervention features, which are used to characterize the causal effect of the treatment variable on the outcome variable after removing the influence of the confounding factors. By fusing the original state representations corresponding to multi-source heterogeneous data with the post-intervention features, a causal-enhanced state representation is obtained.

[0044] In this embodiment, it should be noted that the treatment variable is the variable that acts as an influencing factor in causal analysis. This can be a variable that affects inventory-related outcomes, such as promotional status, holidays, or price adjustments. The outcome variable is the variable that is affected in causal analysis. This can be an inventory-related variable that is affected by the treatment variable, such as sales volume or inventory depletion rate. The determination of these two variables can be based on the business objectives of inventory management and the causal relationship between the variables. Confounding factors are variables that simultaneously affect both the treatment and outcome variables, easily leading to spurious correlations. They can be identified through the direction relationships of variables in a causal relationship diagram. If a variable has directed edges pointing to both the treatment and outcome variables, then that variable is a confounding factor. For example, if the season points to both promotional status and sales volume, then the season is a confounding factor in this causal relationship. Intervention adjustment is a feature correction operation performed on confounding factors. This can be achieved through backdoor adjustment or propensity score weighting. The core is to remove the interference of confounding factors on the causal relationship between the treatment and outcome variables. The post-intervention features are those that, after adjustment, only reflect the true causal effect of the treatment variable on the outcome variable. Feature fusion can be achieved by concatenating columns according to feature dimensions, integrating the features of the original state representation and the features after intervention into a complete feature set, forming a causal enhanced state representation.

[0045] This embodiment generates a state representation that truly reflects causal relationships by identifying and intervening in confounding factors. This enables subsequent models to learn and make decisions based on causal mechanisms rather than spurious correlations, significantly improving the model's generalization ability in the face of environmental changes.

[0046] In one embodiment, based on an inventory decision feature set, a pre-trained inventory decision model outputs a replenishment decision value for the target inventory object, including: The inventory decision feature set is input into the pre-trained inventory decision model. The inventory decision model aims to maximize long-term cumulative returns and outputs the replenishment decision value of the target inventory object.

[0047] In this embodiment, it should be noted that the pre-trained inventory decision model can be a general term for various machine learning models. For example, it can use deep reinforcement learning models such as DDPG (Deep Deterministic Policy Gradient), or supervised learning models such as deep neural networks to directly fit the optimal replenishment quantity. This inventory decision model aims to maximize long-term cumulative returns. Long-term cumulative returns can be defined according to business needs, representing the comprehensive returns achieved by the enterprise through inventory control over a continuous operating cycle. These long-term cumulative returns can include the net profit after deducting inventory holding costs, stockout costs, and procurement costs from sales profits; they can also include the financial returns from improved inventory turnover efficiency; and they can be combined with the potential returns from improved customer service levels. The inventory decision feature set can be input into the pre-trained inventory decision model in several ways: it can be directly input after standardizing the feature set according to the input data format of the inventory decision model; it can be input after feature dimensionality reduction of the feature set; or it can be input after dividing the feature set into sequential data according to the time dimension. During the training phase, the inventory decision model learns the optimal replenishment strategy corresponding to different inventory decision feature sets based on historical data. The training process always aims to maximize long-term cumulative returns. During the inference phase, it outputs replenishment decision values ​​that can achieve this goal based on the input inventory decision feature set.

[0048] This embodiment avoids the suboptimal problem caused by the separation of prediction and decision-making in traditional methods through end-to-end optimization aimed at long-term cumulative benefits. It enables the inventory decision model to make replenishment decisions that take into account both current costs and future benefits from a global perspective. At the same time, once the inventory decision model is trained, it can achieve rapid inference, which meets the real-time requirements of actual business.

[0049] In one embodiment, the inventory control method further includes: Based on replenishment decision values, real-time inventory status data, and demand uncertainty measures, the Monte Carlo simulation method is used to generate the inventory change trajectory within a future preset period. The probability of stockouts for a target inventory item within a future preset period is determined based on inventory change trajectories.

[0050] In this embodiment, it should be noted that any predictive model has inherent uncertainties. Before implementing the replenishment decisions output by the inventory decision model, it is necessary to assess potential risks to enhance the interpretability and credibility of the decision. Monte Carlo simulation is a numerical analysis method that predicts the probability distribution of event outcomes through numerous random simulation experiments. Monte Carlo simulation simulates future inventory changes by repeatedly generating random data that conforms to the demand probability distribution. The inventory change trajectory refers to the trend of daily or time-point inventory changes within a preset future period, assuming the replenishment decision value is executed. This inventory change trajectory can be generated by simulating different demand scenarios, including both the expected inventory change value and the inventory fluctuation range. The stockout probability refers to the probability that the inventory quantity cannot meet actual demand within a preset future period. It can be determined by statistically analyzing the proportion of stockout scenarios occurring in the Monte Carlo simulation experiments to the total number of experiments. The number of simulation experiments can be set according to actual demand; the more experiments, the more accurate the stockout probability calculation. Simultaneously, the inventory change trajectory of each simulation can be recorded to calculate the expected value and confidence interval of the daily inventory quantity in the future, such as the interval formed by the 5th and 95th quantiles.

[0051] This embodiment uses Monte Carlo simulation to simulate the risks of replenishment decisions, providing inventory managers with a quantitative risk assessment for their decisions, making the decision-making process more transparent and credible. It also provides feedback information for model optimization, allowing for timely adjustment of decision parameters when simulation results indicate that the risk is too high.

[0052] In one embodiment, obtaining a measure of the demand uncertainty for a target inventory object over a future preset period includes: Using a pre-trained probabilistic prediction model, a measure of the demand uncertainty of a target inventory object within a future preset period is generated based on multi-source heterogeneous data. The measure of demand uncertainty includes the expected value and standard deviation of the demand within the future preset period.

[0053] It should be noted that the generation of demand uncertainty metrics relies on scientific data analysis methods. Generating them solely through manual statistical analysis is susceptible to subjective influences and has low accuracy. Therefore, a pre-trained probabilistic prediction model is needed to generate demand uncertainty metrics based on multi-source heterogeneous data, explicitly including expected value and standard deviation, to improve the accuracy and objectivity of the metrics. In this embodiment, the probabilistic prediction model is a prediction model capable of outputting the probability distribution characteristics of the prediction results. It can be a Bayesian LSTM (Long Short-Term Memory) model, a Gaussian process regression model, or a random forest probabilistic prediction model. This probabilistic prediction model has been trained using historical multi-source heterogeneous data during the pre-training phase, enabling it to accurately uncover demand patterns within the data. The expected value is the average predicted value of demand within a preset future period, reflecting the overall level of future demand. The standard deviation is the degree to which demand deviates from the expected value within the preset future period, reflecting the magnitude of future demand fluctuations. The combination of these two values ​​comprehensively characterizes the probabilistic distribution characteristics of demand; a larger standard deviation indicates higher uncertainty in future demand. The process of generating a demand uncertainty metric using a probabilistic forecasting model involves taking demand-related data from multi-source heterogeneous data sources, such as historical sales sequences, promotional tags, and holiday tags, processing them according to the input requirements of the probabilistic forecasting model, and then inputting them into the model. The probabilistic forecasting model then infers and outputs the expected value and standard deviation of the demand within a preset future period. This process can be automated and batch-processed, and is suitable for demand analysis of multi-category target inventory objects.

[0054] This embodiment outputs the expected value and standard deviation of demand through a pre-trained probabilistic prediction model. This not only provides the most likely value of demand but also quantifies the degree of uncertainty in the prediction, providing key input information for subsequent risk perception decisions.

[0055] In one embodiment, causal augmentation state representation, demand uncertainty metric, and real-time inventory status data are fused to construct an inventory decision feature set, including: Based on causal enhanced state representation, a sliding window sequence of target inventory objects is constructed according to a preset time window; Temporal features are extracted from the sliding window sequence and fused with the causal augmented state representation to obtain the temporal feature tensor; By fusing time-series feature tensors, demand uncertainty metrics, and real-time inventory status data, an inventory decision feature set is constructed.

[0056] In this embodiment, it should be noted that the causal reinforcement state representation is static cross-sectional data, while inventory decisions often rely on historical change patterns, requiring the extraction of time-series features to capture dynamic information. The preset time window can be set according to business needs, such as the past 60 days. Constructing a sliding window sequence means that for each SKU, for each day, the causal reinforcement state representation of that SKU over the past 60 days is extracted to form a time window sequence. The time-series features extracted from the sliding window sequence can include various types, such as lag features like sales on day t-1 and day t-7, sliding window statistical features like the mean and standard deviation of sales over the past 7 days, periodic features like week-on-week and month-on-month comparisons, and calendar features like day of the week, month, and holiday identifiers. The extracted time-series features are fused with the original causal reinforcement state representation to obtain a time-series feature tensor, which is a three-dimensional structure, for example, with dimensions of SKU number × time step × feature dimension. Then, the time-series feature tensor is concatenated and fused with demand uncertainty metrics and real-time inventory status data to finally obtain the inventory decision feature set.

[0057] In this embodiment, by constructing a sliding window sequence and extracting temporal features, the static causal enhancement state is extended into a temporal feature tensor containing historical dynamic information, enabling the subsequent decision-making model to perceive the changing trends and fluctuation patterns of demand. Combined with uncertainty measurement and real-time status, a complete characterization of the current decision-making scenario is formed.

[0058] In one embodiment, before the step of performing causal augmentation processing on the original state representations corresponding to multi-source heterogeneous data to generate causal augmented state representations, the method further includes: Preprocessing of multi-source heterogeneous data includes at least one of data cleaning, missing value imputation, and outlier detection and replacement.

[0059] In this embodiment, it should be noted that the original multi-source heterogeneous data often has quality issues, and if not processed, it will directly affect the accuracy of subsequent causal analysis and model training. Data cleaning includes operations such as removing duplicate data, correcting erroneous data, and standardizing data formats. Missing value imputation refers to processing null values ​​in the data; for example, NULL values ​​in sales data can be filled with 0 to indicate no sales on that day, and missing values ​​in inventory snapshots can be filled with data from the previous day. Outlier detection can use statistical methods such as the 3-Sigma principle, that is, when the sales volume of a certain day exceeds the average of the past 30 days plus 3 times the standard deviation, it is marked as an outlier and replaced with the mean or median.

[0060] In this embodiment, by preprocessing the raw data, the interference of data quality issues on subsequent analysis is eliminated, providing a high-quality data foundation for causal augmentation processing and model training, thereby improving the reliability and accuracy of the entire technical solution.

[0061] Figure 2 A schematic block diagram of an electronic device according to an embodiment of this application is shown. Figure 2 As shown in the figure, this application provides an electronic device that may include: Memory 1000 is configured to store instructions; The processor 2000 is configured to retrieve instructions from memory 1000 and to implement the aforementioned inventory control method when executing instructions.

[0062] This application also provides a machine-readable storage medium storing instructions for causing a machine to perform the above-described inventory control method.

[0063] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0064] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0065] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0066] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0067] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0068] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0069] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0070] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0071] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. An inventory control method, characterized in that, include: Acquire multi-source heterogeneous data of the target inventory object, wherein the multi-source heterogeneous data includes historical time-series data, event data, and supply chain data; The original state representations corresponding to the multi-source heterogeneous data are subjected to causal enhancement processing to generate causal enhanced state representations; Obtain the demand uncertainty measure value of the target inventory object in the future preset period, wherein the demand uncertainty measure value is used to characterize the probability distribution characteristics of the demand in the future preset period; The causal enhancement state representation, the demand uncertainty metric, and the real-time inventory status data are fused to construct an inventory decision feature set; Based on the inventory decision feature set, the replenishment decision value of the target inventory object is output through a pre-trained inventory decision model.

2. The inventory control method according to claim 1, characterized in that, The step of performing causal augmentation processing on the original state representation corresponding to the multi-source heterogeneous data to generate a causal augmented state representation includes: Based on preset inventory control business rules, key variables related to inventory changes are extracted from the multi-source heterogeneous data to form a set of key variables; Based on the conditional independence test method, the causal relationships among the key variables in the set of key variables are determined, and a causal relationship diagram is generated; Based on the causal relationship graph, causal intervention is performed on the multi-source heterogeneous data to generate a causal-enhanced state representation.

3. The inventory control method according to claim 2, characterized in that, The step of performing causal intervention on the multi-source heterogeneous data based on the causal relationship graph to generate a causal-enhanced state representation includes: Based on the causal relationship diagram, the processing variables and outcome variables among the key variables are determined; Based on the pointing relationships between the key variables in the causal relationship diagram, a target key variable that simultaneously points to both the processing variable and the result variable is identified as a confounding factor. The confounding factors are adjusted through intervention to generate post-intervention features, wherein the post-intervention features are used to characterize the causal effect of the treatment variable on the outcome variable after removing the influence of the confounding factors; The original state representation corresponding to the multi-source heterogeneous data is fused with the post-intervention features to obtain a causal enhanced state representation.

4. The inventory control method according to claim 1, characterized in that, The step of outputting the replenishment decision value of the target inventory object based on the inventory decision feature set and through a pre-trained inventory decision model includes: The inventory decision feature set is input into the pre-trained inventory decision model, which aims to maximize long-term cumulative returns and outputs the replenishment decision value of the target inventory object.

5. The inventory control method according to claim 1, characterized in that, Also includes: Based on the replenishment decision value, the real-time inventory status data, and the demand uncertainty measure, the inventory change trajectory within the future preset period is generated using the Monte Carlo simulation method. The probability of stockout for the target inventory item within the future preset period is determined based on the inventory change trajectory.

6. The inventory control method according to claim 1, characterized in that, The step of obtaining the demand uncertainty measure of the target inventory object within a future preset period includes: Using a pre-trained probabilistic prediction model, a demand uncertainty measure for the target inventory object within a future preset period is generated based on the multi-source heterogeneous data. The demand uncertainty measure includes the expected value and standard deviation of the demand within the future preset period.

7. The inventory control method according to claim 1, characterized in that, The process of fusing the causal enhancement state representation, the demand uncertainty metric, and real-time inventory status data to construct an inventory decision feature set includes: Based on the causal enhanced state representation, a sliding window sequence of the target inventory object is constructed according to a preset time window; Temporal features are extracted from the sliding window sequence, and the temporal features are fused with the causal enhanced state representation to obtain a temporal feature tensor; The time-series feature tensor, the demand uncertainty metric, and the real-time inventory status data are fused to construct an inventory decision feature set.

8. The inventory control method according to claim 1, characterized in that, Before the step of performing causal augmentation processing on the original state representation corresponding to the multi-source heterogeneous data to generate a causal augmented state representation, the method further includes: The multi-source heterogeneous data is preprocessed, wherein the preprocessing includes at least one of data cleaning, missing value imputation, and outlier detection and replacement.

9. An electronic device, characterized in that, include: The memory is configured to store instructions; The processor is configured to retrieve the instructions from the memory and, when executing the instructions, to implement the inventory control method according to any one of claims 1 to 8.

10. A machine-readable storage medium, characterized in that, The machine-readable storage medium stores instructions for causing the machine to perform the inventory control method according to any one of claims 1 to 8.

Images