Intelligent inventory management method, device and equipment and storage medium

Abstract

The invention provides an intelligent inventory management method, device and equipment and a storage medium, and the method comprises the steps: carrying out the data collection and preprocessing of an enterprise resource planning system, a supplier management system and a point-of-sale system of a target enterprise, obtaining an integrated data set, constructing a plurality of time-varying autoregression demand prediction models, and carrying out the prediction of the demands of the target enterprise. Combining a plurality of prediction results for constructing a plurality of time-varying autoregression demand prediction models in combination with an ensemble learning method to obtain a dynamic demand prediction result; according to the deep Q network model, evaluating an inventory replenishment decision in combination with the dynamic demand prediction result and the integrated data set to obtain an inventory optimization strategy; and performing multi-target Pareto optimization on the inventory management of the target enterprise according to the inventory optimization strategy to obtain and execute an intelligent inventory management scheme. According to the method, the comprehensiveness and the dynamism of demand prediction are remarkably improved by comprehensively utilizing data of all systems in an enterprise and applying an advanced time-varying prediction model and an integrated learning technology.

Description

Technical Field

[0001] The present invention relates to the field of inventory management, and in particular to an intelligent inventory management method, device, equipment and storage medium. Background Art

[0002] In the modern business environment, inventory management, as an important part of supply chain management, has a key impact on the operational efficiency and cost control of enterprises. Traditional inventory management methods mainly rely on information systems such as enterprise resource planning (ERP) systems, supplier management systems and point of sale (POS) systems to monitor and adjust inventory levels by analyzing historical sales data, supply chain information and real-time sales data. However, these traditional methods generally have a major flaw, that is, the accuracy of demand forecasting is insufficient.

[0003] Traditional inventory management methods mostly use static demand forecasting models, such as simple time series analysis, moving average or linear regression models. These models are often unable to cope with complex and changing market demands, and it is difficult to capture potential demand fluctuations and seasonal changes. In addition, with the rapid changes in the market environment and the increasing diversification of consumer behavior, single-band forecasting models cannot effectively integrate data sources from different systems, such as inventory data in the ERP system, supply information in the supplier management system, and real-time sales data in the POS system, resulting in a lack of comprehensiveness and dynamism in the forecast results. Due to inaccurate demand forecasts, companies often face the problem of overstocking or out-of-stock. Overstocking not only occupies the company's funds and storage space, but may also cause inventory goods to become obsolete or damaged, increasing the company's operating costs; while out-of-stock directly affects customer satisfaction and purchasing experience, thereby damaging the company's market competitiveness and brand reputation. Summary of the invention

[0004] The main purpose of the present invention is to solve the technical problem that the accuracy of demand forecasting in existing inventory management methods is insufficient, resulting in excess inventory or shortage of inventory; A first aspect of the present invention provides an intelligent inventory management method, the intelligent inventory management method comprising: Collect and preprocess data from the target enterprise's enterprise resource planning system, supplier management system, and point-of-sale system to obtain an integrated data set; According to the integrated data set, multiple time-varying autoregressive demand forecasting models are constructed, and the forecast results of multiple time-varying autoregressive demand forecasting models are combined with an integrated learning method to obtain a dynamic demand forecasting result including a point forecast and a forecast interval; According to the pre-trained deep Q network model, the inventory replenishment decision is evaluated in combination with the dynamic demand forecast results and the integrated data set to obtain an inventory optimization strategy; The inventory management of the target enterprise is subjected to multi-objective Pareto optimization according to the inventory optimization strategy, and an intelligent inventory management solution is obtained and executed.

[0005] Optionally, in a first implementation of the first aspect of the present invention, constructing multiple time-varying autoregressive demand forecasting models based on the integrated data set, and combining the prediction results of multiple time-varying autoregressive demand forecasting models constructed by an integrated learning method to obtain a dynamic demand forecasting result including a point prediction and a prediction interval includes: Performing time series decomposition on the integrated data set to obtain trend components, seasonal components and random components; Based on the trend, seasonality and random components, multiple time-varying autoregressive models with different autoregressive orders and exogenous variable combinations are constructed; The Kalman filter algorithm is used to estimate and update the parameters of each time-varying autoregressive model to obtain the prediction results of each time-varying autoregressive model; The prediction results of each model are weighted averaged to obtain a comprehensive prediction result. The Monte Carlo simulation method is used to perform multiple sampling and calculations on the comprehensive prediction result to obtain a dynamic demand prediction result that includes point prediction and prediction interval.

[0006] Optionally, in a second implementation of the first aspect of the present invention, constructing multiple time-varying autoregressive models with different autoregressive orders and exogenous variable combinations according to the trend, seasonality and random components includes: Perform polynomial fitting on the trend component to obtain the trend function that describes the long-term trend. Perform Fourier series expansion on the seasonal component to obtain the seasonal function that describes the periodic change. Perform autocorrelation analysis on the random component to determine the maximum autoregressive order. Selecting exogenous variables from the integrated data set, and constructing a preliminary autoregressive model based on the trend function, seasonal function and autoregressive terms, wherein the exogenous variables include economic indicators, market data and supply chain characteristics; Each exogenous variable is added to the preliminary autoregressive model one by one to form multiple extended models, and each extended model is evaluated using an information criterion to select an optimal subset of exogenous variables; A grid search method is used to generate all combinations of different autoregressive orders and different exogenous variables in the exogenous variable subset, and for each combination, a corresponding time-varying autoregressive model is constructed.

[0007] Optionally, in a third implementation of the first aspect of the present invention, the prediction results of each model are weighted averaged to obtain a comprehensive prediction result, and the comprehensive prediction result is sampled and calculated multiple times using a Monte Carlo simulation method to obtain a dynamic demand prediction result including a point prediction and a prediction interval, including: The root mean square of the historical forecast error of each time-varying autoregressive model during the parameter estimation and updating process is calculated, and the weight of each time-varying autoregressive model is calculated according to the inverse of the root mean square of each time-varying autoregressive model; By calculating the weights, weighted averaging is performed on the prediction results of the corresponding time-varying autoregressive models to obtain a comprehensive prediction result, and a normality test is performed on the error term of the comprehensive prediction result to obtain an error distribution parameter; Using the Monte Carlo method, samples are drawn from the error distribution parameters multiple times, and the drawn samples are added to the comprehensive prediction result each time to obtain multiple simulated prediction values; The average value of the simulated prediction values ​​is calculated as the point prediction, and the quantiles at a specific confidence level are calculated as the upper and lower bounds of the prediction interval to obtain a dynamic demand forecast result including the point prediction and the prediction interval.

[0008] Optionally, in a fourth implementation of the first aspect of the present invention, the inventory replenishment decision is evaluated based on the pre-trained deep Q network model in combination with the dynamic demand forecast result and the integrated data set to obtain the inventory optimization strategy, including: encoding the dynamic demand forecast result and the current inventory level, supplier reliability index and inventory holding cost in the integrated data set into a state vector; Input the state vector into a pre-trained deep Q network model to obtain Q values ​​of various possible replenishment decisions, and select the optimal replenishment decision based on the Q values ​​and a preset greedy strategy; The selected replenishment decision is simulated and executed, the expected inventory cost and service level are calculated, and based on the simulation results, the replenishment decision is fine-tuned to obtain the inventory optimization strategy.

[0009] Optionally, in a fifth implementation of the first aspect of the present invention, inputting the state vector into a pre-trained deep Q network model to obtain Q values ​​of various possible replenishment decisions, and selecting the optimal replenishment decision according to the Q value and a preset greedy strategy includes: Input the state vector into the pre-trained deep Q network model, normalize the input state vector, and obtain a standardized state representation; Input the standardized state representation into the input layer of the deep Q network, perform forward propagation calculation through a multi-layer convolutional neural network and a fully connected layer, and obtain the Q value of each possible replenishment decision; The Q values ​​are sorted, and according to a preset ε-greedy strategy, a replenishment decision with the highest Q value is selected with a probability of 1-ε, and a replenishment decision is randomly selected with a probability of ε to obtain a candidate replenishment decision; Action noise is applied to the candidate replenishment decisions to obtain the optimal replenishment decision by adding Gaussian noise with zero mean and standard deviation decaying over time.

[0010] Optionally, in a sixth implementation of the first aspect of the present invention, performing multi-objective Pareto optimization on the inventory management of the target enterprise according to the inventory optimization strategy to obtain and execute the intelligent inventory management solution includes: According to extracting the key decision parameters of the inventory optimization strategy, wherein the key decision parameters include replenishment point, replenishment quantity and safety stock level; Using the decision parameters, a multi-objective optimization model including minimizing inventory cost, maximizing service level and minimizing supply chain risk is constructed; An improved multi-objective particle swarm algorithm is applied to the multi-objective optimization model to generate a series of non-dominated solutions and form the Pareto front; Perform fuzzy C-means clustering on the solutions on the Pareto front to obtain multiple inventory management solutions, and use the TOPSIS method combined with the evaluation results of the inventory optimization strategy to comprehensively evaluate and rank the multiple inventory management solutions to determine the intelligent inventory management solution; The intelligent inventory management solution is converted into specific inventory control parameters and operation instructions, and distributed to relevant departments for execution through the target enterprise's enterprise resource planning system.

[0011] A second aspect of the present invention provides an intelligent inventory management device, the intelligent inventory management device comprising: The data integration module is used to collect and preprocess data from the target enterprise's enterprise resource planning system, supplier management system, and point of sale system to obtain an integrated data set; A demand forecasting module is used to construct multiple time-varying autoregressive demand forecasting models based on the integrated data set, and combine the forecast results of multiple time-varying autoregressive demand forecasting models in combination with an integrated learning method to obtain a dynamic demand forecasting result including a point forecast and a forecast interval; A replenishment decision module, used to evaluate the inventory replenishment decision based on the pre-trained deep Q network model, combined with the dynamic demand forecast results and the integrated data set, to obtain an inventory optimization strategy; The inventory optimization module is used to perform multi-objective Pareto optimization on the inventory management of the target enterprise according to the inventory optimization strategy, and obtain and execute an intelligent inventory management solution.

[0012] The third aspect of the present invention provides an intelligent inventory management device, comprising: a memory and at least one processor, wherein instructions are stored in the memory, and the memory and the at least one processor are interconnected via lines; the at least one processor calls the instructions in the memory so that the intelligent inventory management device executes the steps of the above-mentioned intelligent inventory management method.

[0013] A fourth aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the steps of the above-mentioned intelligent inventory management method.

[0014] The above-mentioned intelligent inventory management method, device, equipment and storage medium obtain an integrated data set by collecting and preprocessing data from the target enterprise's enterprise resource planning system, supplier management system and point of sale system, and then construct multiple time-varying autoregressive demand forecasting models, and combine the prediction results of multiple time-varying autoregressive demand forecasting models with an integrated learning method to obtain dynamic demand forecasting results; according to the deep Q network model, the inventory replenishment decision is evaluated in combination with the dynamic demand forecasting results and the integrated data set to obtain an inventory optimization strategy; according to the inventory optimization strategy, the inventory management of the target enterprise is multi-objective Pareto optimized to obtain and execute an intelligent inventory management plan. The present invention significantly improves the comprehensiveness and dynamism of demand forecasting by comprehensively utilizing the data of various systems within the enterprise and applying advanced time-varying prediction models and integrated learning technologies.

[0015] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0016] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A schematic diagram of a first embodiment of an intelligent inventory management method in an embodiment of the present invention; Figure 2 A schematic diagram of an embodiment of an intelligent inventory management device in an embodiment of the present invention; Figure 3 Schematic diagram of an embodiment of an intelligent inventory management device in an embodiment of the present invention. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] The terms "including" and "having" and any variations thereof mentioned in the embodiments of the present invention are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device end including a series of steps or units is not limited to the listed steps or units, but may optionally include other steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products or device ends.

[0020] To facilitate understanding of this embodiment, firstly, a smart inventory management method disclosed in an embodiment of the present invention is described in detail. Figure 1 As shown, the method comprises the following steps: 101. Collect and preprocess data from the target enterprise's enterprise resource planning system, supplier management system, and point of sale system to obtain an integrated data set; In one embodiment of the present invention, data collection and data preprocessing are performed on the target enterprise's enterprise resource planning (ERP) system, supplier management system, and point of sale (POS) system to generate an integrated data set. In the specific implementation process, key data, including financial statements, inventory records, production plans, and purchase orders, are first extracted from the ERP system through a secure API interface or data export tool. These data reflect the internal resource allocation, inventory status, and production operations of the enterprise. At the same time, data such as supplier performance indicators, order fulfillment, delivery time, and supply chain interruption records are obtained from the supplier management system. These data are used to evaluate the reliability and performance of suppliers. Another data source, the point of sale system, provides detailed sales transaction records, customer purchase behavior, product return information, and real-time inventory consumption. These data can reflect market demand and sales dynamics. After completing data collection, enter the data preprocessing stage. First, data cleaning is performed to delete redundant and duplicate records, process missing values, and ensure the integrity and accuracy of the data through interpolation, mean filling, or deletion of missing data. Then, the data collected in different systems are formatted and standardized. For example, the inventory quantity units in the ERP system are unified with the sales units in the POS system to ensure the consistency of data in subsequent analysis. In addition, through data conversion, unstructured data is converted into a structured format to facilitate subsequent storage and analysis. Subsequently, the cleaned and converted data is further processed using ETL (extraction, transformation, loading) tools. The extraction phase is responsible for extracting data from different systems, and the conversion phase includes data normalization and consistency processing, such as the unification of time formats and the conversion of currency units, to ensure that data from different sources can be seamlessly integrated on the same platform. The loading phase imports the processed data into a centralized data warehouse or data lake to establish a unified data storage environment. This process requires the definition of data mapping and association rules to ensure that data from different systems can be correctly associated, such as through key fields such as order numbers, product IDs or supplier IDs, to achieve accurate matching and association between data. In the data integration process, data fusion technology is used to integrate data from the ERP system, supplier management system and POS system to form a comprehensive integrated data set. By establishing a multidimensional data model, complex query and analysis requirements are supported to ensure the relevance and consistency of data in different dimensions. Ultimately, the resulting integrated data set contains resource allocation and inventory information from the ERP system, performance and order fulfillment data from the supplier management system, and sales and inventory consumption records from the POS system. These data are stored in the data warehouse in a unified format and structure.

[0021] 102. Based on the integrated data set, multiple time-varying autoregressive demand forecasting models are constructed, and the forecast results of multiple time-varying autoregressive demand forecasting models are combined with the integrated learning method to obtain dynamic demand forecasting results including point forecasts and forecast intervals; In one embodiment of the present invention, multiple time-varying autoregressive demand forecasting models are constructed based on the integrated data set, and the prediction results of multiple time-varying autoregressive demand forecasting models are combined in combination with an integrated learning method to obtain a dynamic demand forecasting result including a point prediction and a prediction interval, including: performing time series decomposition on the integrated data set to obtain a trend component, a seasonal component and a random component; constructing multiple time-varying autoregressive models with different autoregressive orders and exogenous variable combinations based on the trend, seasonality and random components; using the Kalman filtering algorithm to estimate and update the parameters of each time-varying autoregressive model to obtain the prediction results of each time-varying autoregressive model; performing weighted averaging on the prediction results of each model to obtain a comprehensive prediction result, and using the Monte Carlo simulation method to sample and calculate the comprehensive prediction result multiple times to obtain a dynamic demand forecasting result including a point prediction and a prediction interval.

[0022] Specifically, first, the integrated data set is decomposed into time series to extract trend components, seasonal components, and random components. In the specific implementation process, an additive or multiplicative time series decomposition model is used to decompose the complex demand data into three basic components. The additive model assumes that the time series can be expressed as a simple addition of trend, seasonality, and random fluctuations, while the multiplicative model assumes that these components are multiplied. The selection of an appropriate decomposition model depends on the characteristics of the demand data. For example, if the fluctuation amplitude of demand increases over time, a multiplicative model may be more appropriate. Using the moving average method or exponential smoothing method, long-term trend components can be effectively extracted to reflect the overall upward or downward trend of demand. At the same time, seasonal components with fixed periodicity are identified and separated through Fourier transform or seasonal decomposition algorithms. These components reveal the regular fluctuations of demand in different seasons or cycles. The remaining parts are classified as random components, representing unpredictable short-term fluctuations and noise. Through this decomposition, the different influencing factors in the demand data can be clearly understood and quantified, providing a clear structural basis for the subsequent construction of demand forecasting models. In addition, time series decomposition also helps identify abnormal points and abnormal patterns in the data, ensuring data quality and model accuracy in the subsequent modeling process. The decomposed components can not only be analyzed separately, but also play an important role in comprehensive modeling, ensuring that the demand forecasting model can fully capture the dynamic changing characteristics of demand.

[0023] Subsequently, multiple time-varying autoregressive models with different autoregressive orders and exogenous variable combinations are constructed based on the trend, seasonal and random components obtained by decomposition. By adjusting the autoregressive order p and selecting different exogenous variable combinations, multiple TVAR models are generated. The TVAR model is an extended autoregressive model that allows the model coefficients to change over time, thereby better capturing the dynamic changes in demand data. Specifically, different TVAR model structures are first designed based on the trend, seasonal and random components obtained by decomposition. By adjusting the autoregressive order, that is, considering different numbers of historical demand data points, multiple models can be constructed to adapt to different demand patterns and data characteristics. At the same time, appropriate exogenous variables, such as economic indicators, market trend data and supply chain characteristics, are selected and incorporated into the model to enhance the model's sensitivity and adaptability to changes in external factors. These exogenous variables provide additional sources of information by affecting various aspects of demand, helping the model to more accurately predict future demand. After determining different exogenous variable combinations and autoregressive orders for each TVAR model, a series of diversified models are generated to cover multiple possibilities of demand changes. This multi-model construction approach not only improves the flexibility and accuracy of predictions, but also reduces the bias and error that may be caused by a single model by integrating diverse model results.

[0024] Specifically, after constructing multiple time-varying autoregressive models, the Kalman filter algorithm is used to estimate and update the parameters of each TVAR model to obtain the prediction results of each model. Kalman filtering is a recursive algorithm that can estimate the state parameters of the system in real time in the presence of noise and uncertainty. Specifically, for each TVAR model, the prior distribution of the model parameters is first initialized, and then the posterior distribution of the parameters is updated at each time step based on the new observation data and model prediction. The Kalman filter process includes a prediction step and an update step: in the prediction step, the current parameter estimate is used to predict the demand value of the next time step; in the update step, the predicted value is corrected in combination with the actual observation value, and the parameter estimate is updated. This process enables each TVAR model to dynamically adjust its coefficients to adapt to the real-time changes in demand data. By applying Kalman filtering to each model, accurate prediction results of each model at different time points are obtained, ensuring that the model can reflect the latest dynamics of demand in a timely manner. Kalman filtering not only improves the real-time and adaptability of the model, but also reduces prediction errors and deviations by gradually updating parameters. Ultimately, the forecast results of each TVAR model represent the demand forecast value of the model at a specific point in time. These results will serve as the basic data for subsequent weighted averaging and Monte Carlo simulation to ensure that the final demand forecast results are highly accurate and reliable.

[0025] Specifically, after obtaining the prediction results of each time-varying autoregressive model, these results are integrated by weighted average method to obtain a comprehensive prediction result, and the dynamic demand prediction result containing point prediction and prediction interval is generated by Monte Carlo simulation method. First, according to the prediction error of each TVAR model on historical data, its weight is calculated, and the inverse of the root mean square error (RMSE) is usually used as the weight. Then, the prediction results of each model are weighted averaged according to the weight to obtain a comprehensive prediction value. In order to further quantify the uncertainty of the prediction, the Monte Carlo simulation method is used to sample and calculate the comprehensive prediction results multiple times. First, according to the error distribution parameters of the comprehensive prediction results, multiple simulation samples are generated. Each simulation sample adds a random error term to the comprehensive prediction result, and through the generation of a large number of simulation samples, the average value of each sample is calculated as the point prediction value, and the upper and lower bounds of the prediction interval are determined by the quantile of the sample.

[0026] Furthermore, the construction of multiple time-varying autoregressive models with different autoregressive orders and exogenous variable combinations based on the trend, seasonal and random components includes: performing polynomial fitting on the trend component to obtain a trend function that describes the long-term change trend, performing Fourier series expansion on the seasonal component to obtain a seasonal function that describes cyclical changes, performing autocorrelation analysis on the random component to determine the maximum autoregressive order; selecting exogenous variables from the integrated data set, and constructing a preliminary autoregressive model based on the trend function, seasonal function and autoregressive terms, wherein the exogenous variables include economic indicators, market data and supply chain characteristics; adding each exogenous variable to the preliminary autoregressive model one by one to form multiple extended models, and using the information criterion to evaluate each extended model to select the optimal subset of exogenous variables; using the grid search method to generate all combinations of different autoregressive orders and different exogenous variables in the exogenous variable subset, and for each combination, constructing a corresponding time-varying autoregressive model.

[0027] Specifically, in the second step of the intelligent inventory management method, multiple time-varying autoregressive models with different autoregressive orders and exogenous variable combinations are constructed according to the trend, seasonality and random components. First, a polynomial fitting is performed on the trend component, and a trend region describing the long-term trend is obtained by fitting a quadratic or higher order polynomial function. For example, assuming that the trend component It can be expressed as: ; in is the fitting parameter, is a time variable. Polynomial fitting can effectively capture the long-term upward or downward trend in demand data, ensuring that the model can reflect the overall change direction and speed of demand. Subsequently, the seasonal component is expanded by Fourier series, and the seasonal fluctuation is expressed as a linear combination of a series of sine and cosine functions using Fourier transform, such as ; in is the cycle length, and are the Fourier coefficients, is the number of Fourier terms. The Fourier series expansion can accurately characterize the cyclical changes in demand data, capture seasonal fluctuations at different frequencies, and enhance the model's sensitivity to cyclical demand fluctuations. For random components, the maximum autoregressive order is determined by autocorrelation analysis. , that is, to select the order of the autoregressive model that can fully describe the random fluctuations. The autocorrelation function (ACF) and partial autocorrelation function (PACF) are used to identify the significant lags in the random component and thus determine the appropriate After processing the trend, seasonality and random components, exogenous variables are selected from the integrated data set, including economic indicators, market data and supply chain characteristics. These exogenous variables provide additional sources of information by affecting various aspects of demand, enhancing the model's adaptability and predictive ability to changes in external factors. Based on the trend function, seasonality function and autoregressive term, a preliminary autoregressive model is constructed: ; in is the demand, is the autoregression coefficient, is the error term. The preliminary autoregressive model integrates the influence of long-term trends, seasonal fluctuations, and random fluctuations, providing a basic framework. Subsequently, each exogenous variable is added to the preliminary autoregressive model one by one to form multiple extended models. For example, for each exogenous variable , the extended model can be expressed as ; Information criteria such as Akaike Information Criterion (AIC) or Bayesian Information Criterion (BIC) are used to evaluate the goodness of fit and complexity of each extended model, and the optimal subset of exogenous variables is selected. This process ensures that the model has both high prediction accuracy and avoids overfitting. By selecting the most explanatory exogenous variables, the stability and generalization ability of the model are improved. Finally, the grid search method is used to generate all possible combinations of different autoregressive orders and different exogenous variables in the selected exogenous variable subset. For each combination, the corresponding time-varying autoregressive model (TVAR) is constructed, which is in the form of: ; in and represents the coefficient that varies with time, is the number of exogenous variables. By systematically traversing all possible orders and variable combinations, a series of diverse TVAR models are generated. These models can fully cover various dynamic characteristics and external influencing factors in demand data through different parameter settings and variable selections, ensuring the comprehensiveness and accuracy of the final demand forecast results. The grid search method plays a key role in optimizing model parameters and selecting the best model combination, significantly improving the accuracy of demand forecasting and the robustness of the model.

[0028] Furthermore, the prediction results of each model are weighted averaged to obtain a comprehensive prediction result, and the Monte Carlo simulation method is used to sample and calculate the comprehensive prediction result multiple times to obtain a dynamic demand prediction result including a point prediction and a prediction interval, including: the root mean square of the historical prediction error of each time-varying autoregressive model during the parameter estimation and update process, and the weight of each time-varying autoregressive model is calculated according to the inverse of the root mean square of each time-varying autoregressive model; using the calculated weight, the prediction results of the corresponding time-varying autoregressive models are weighted averaged to obtain a comprehensive prediction result, and the error term of the comprehensive prediction result is tested for normality to obtain error distribution parameters; using the Monte Carlo method, samples are repeatedly drawn from the error distribution parameters, and the drawn samples are added to the comprehensive prediction result each time to obtain multiple simulated prediction values; the average value of the simulated prediction values ​​is calculated as the point prediction, and the quantiles at a specific confidence level are calculated as the upper and lower bounds of the prediction interval to obtain a dynamic demand prediction result including a point prediction and a prediction interval.

[0029] Specifically, the prediction results of each time-varying autoregressive (TVAR) model are weighted averaged to obtain a comprehensive prediction result, and the Monte Carlo simulation method is used to sample and calculate the comprehensive prediction result multiple times to generate a dynamic demand prediction result including point prediction and prediction interval. In the specific implementation process, the root mean square error (RMSE) of the historical prediction error generated by each IVAR model during parameter estimation and update is first calculated. The root mean square error formula is as follows: ; in, For the Models in time The actual demand value, is the predicted value of the model, is the total number of observation points. By calculating the RMSE of each model, we can obtain the prediction error of each model on the historical data. Then, we can calculate the weight of each model according to the inverse of its RMSE. The weight formula is as follows: ; This weight distribution method ensures that the model with smaller prediction error occupies a larger weight in the comprehensive prediction, thereby improving the accuracy and reliability of the comprehensive prediction. Subsequently, the prediction results of each TVAR model are weighted averaged using the calculated weights to obtain the comprehensive prediction result. The comprehensive prediction formula is as follows: ; in, For time The comprehensive prediction value of is the total number of models. After completing the weighted average, the error term of the comprehensive prediction result is tested for normality to determine the parameters of the error distribution. This step usually uses the Shapiro-Wilk test or the Kolmogorov-Smirnow test. The test results are used to determine whether the error term obeys the normal distribution and estimate its mean and variance. These parameters are used in subsequent Monte Carlo simulations to ensure that the generation of error terms during the simulation process conforms to the statistical characteristics of the actual data. After completing the weighted average and determining the error distribution parameters, the Monte Carlo method is used to perform multiple sampling and calculations on the comprehensive prediction results. Specifically, random samples are drawn multiple times from the determined error distribution, and the drawn error samples are added to the comprehensive prediction results each time to generate multiple simulated prediction values. The mathematical expression of the Monte Carlo simulation is: ; in, For the The predicted value of the simulation, is the first number drawn from the error distribution error samples, , is the number of simulations. Through a large number of simulated forecast values, the probability distribution of demand can be constructed, and then the point forecast value and forecast interval can be calculated. Finally, the average of all simulated forecast values ​​is calculated as the point forecast value, which represents the most likely estimate of future demand. At the same time, according to a specific confidence level (such as 95%), the corresponding quantiles of the simulated forecast values ​​are calculated to determine the upper and lower bounds of the forecast interval. For example, the 2.5th percentile and the 97.5th percentile are calculated as the forecast interval at a 95% confidence level.

[0030] 103. Based on the pre-trained deep Q network model, the inventory replenishment decision is evaluated in combination with the dynamic demand forecast results and the integrated data set to obtain the inventory optimization strategy; In one embodiment of the present invention, the inventory replenishment decision is evaluated according to the pre-trained deep Q network model in combination with the dynamic demand forecast results and the integrated data set to obtain the inventory optimization strategy, including: encoding the current inventory level, supplier reliability index and inventory holding cost in the dynamic demand forecast results and the integrated data set into a state vector; inputting the state vector into the pre-trained deep Q network model to obtain Q values ​​of various possible replenishment decisions, and selecting the optimal replenishment decision according to the Q value and a preset greedy strategy; simulating the execution of the selected replenishment decision, calculating the expected inventory cost and service level, and fine-tuning the replenishment decision according to the simulation results to obtain the inventory optimization strategy.

[0031] Specifically, first, the dynamic demand forecast results are encoded with the current inventory level, supplier reliability index and inventory holding cost in the integrated data set to form a state-oriented model. Indicates at time Demand forecasts, including point forecasts and forecast intervals; current inventory levels Indicates the actual quantity of each product in stock; supplier reliability indicator Including on-time delivery rate of each supplier, product quality score, etc.; inventory holding cost Involving storage costs, capital occupation costs, etc. Integrate this information into a high-dimensional vector , which fully reflects the key status of current inventory management. Next, the state vector Input the pre-trained deep Q network model to obtain the Q value of various possible replenishment decisions. The deep Q network model can extract potential features from complex state vectors and predict the Q value of each possible action through a multi-layer neural network structure. Assume that there is possible replenishment decisions, each decision corresponds to a Q value ,in Indicates The model calculates the Q value of all decisions through forward propagation, and the formula is expressed as: ; in, represents the function of the deep Q network, is the parameter of the network. Then, the optimal replenishment decision is selected according to the preset greedy strategy. The Golden Heart strategy usually adopts S-3 greedy-3 greedy-3 greedy-3 cold ... The probability of selecting the decision with the highest Q value is The probability of randomly selecting other decisions to balance exploration and utilization. The optimal replenishment decision selected satisfy: ; In this way, the optimal replenishment decision can be reasonably selected based on the current state vector to ensure that the dynamic adjustment of inventory levels matches market demand. After that, simulation execution is performed to calculate the expected inventory cost and service level. Adjust inventory levels and evaluate the impact of this decision on inventory holding costs and service levels. Inventory Holding Costs The calculation includes storage costs, capital occupation costs, etc., and the service level It is measured by indicators such as out-of-stock rate and order fulfillment rate. Specifically, assuming that the replenishment decision The inventory level is adjusted to ,but ; in, and Represents the replenishment decision Incremental impact on inventory costs and service levels. By simulating the village bank, the economic benefits and service effects of each replenishment decision can be recalculated to ensure the scientificity and rationality of the decision. According to the simulation results, the replenishment decision can be fine-tuned to optimize the inventory strategy. If the simulation results show that the replenishment decision leads to excessively high inventory costs or insufficient service levels, the replenishment quantity is adjusted or an alternative supplier is selected to re-evaluate the new replenishment decision. The adjustment process may involve feedback optimization of the deep Q network model. By recording the difference between the actual execution results and the predicted results, the model parameters are further optimized to improve the model's prediction accuracy and decision-making ability. Finally, after simulated execution and decision fine-tuning, the optimized inventory management strategy is obtained. This strategy ensures a high service level while maintaining a low inventory cost, and comprehensively improves the efficiency and reliability of inventory management.

[0032] Furthermore, the step of inputting the state vector into a pre-trained deep Q network model to obtain Q values ​​of various possible replenishment decisions, and selecting the optimal replenishment decision according to the Q values ​​and a preset greedy strategy includes: The state vector is input into a pre-trained deep Q network model, and the input state vector is normalized to obtain a standardized state representation; the standardized state representation is input into the input layer of the deep Q network, and forward propagation calculation is performed through a multi-layer convolutional neural network and a fully connected layer to obtain the Q value of each possible replenishment decision; the Q values ​​are sorted, and according to the preset ε-greedy strategy, the replenishment decision with the highest Q value is selected with a probability of 1-ε, and a replenishment decision is randomly selected with a probability of ε to obtain candidate replenishment decisions; action noise is applied to the candidate replenishment decisions, and the optimal replenishment decision is obtained by adding Gaussian noise with a mean of zero and a standard deviation that decays over time.

[0033] Specifically, the encoded state vector is first input into the pre-trained deep Q network model, and the input state vector is normalized to obtain a standardized state representation. It includes dynamic demand forecast results, current inventory levels, supplier reliability indicators, and inventory holding costs. Through normalization, the data of each dimension is standardized to a unified scale range to ensure that data from different dimensions are processed under the same standard. The Z-score normalization method is used, that is, the mean of each dimension is subtracted. and divided by the standard deviation , the formula is as follows: ; in, is the normalized state vector, and Represent the mean and standard deviation of each dimension respectively. The purpose of standardization is to eliminate the influence of dimensional differences and numerical ranges in the data, so that the deep Q network model can be more efficient and stable when processing input data. The standardized state vector It is passed to the input layer of the deep Q network model as the input data of the model. Through this processing, it is ensured that the model can accurately capture the relationship between different features during training and prediction, and improve the prediction ability and generalization performance of the model. The normalized state vector not only improves the training efficiency of the model, but also reduces the problem of gradient disappearance or explosion, ensuring that the deep neural network can smoothly perform forward propagation and back propagation calculations in the multi-layer structure, thereby generating accurate Q value predictions. Subsequently, the normalized state vector It is input into the input layer of the deep Q network, and forward propagates through multiple layers of convolutional neural networks (CNN) and fully connected layers to generate the Q value of each possible replenishment decision. The deep Q network model consists of multiple convolutional layers and fully connected layers. The convolutional layer is responsible for extracting spatial features and patterns in the input state vector. It detects and extracts features by applying multiple filters to enhance the model's ability to identify key features in the input data. Each convolutional layer is usually followed by a nonlinear activation function, such as ReLU (Rectified Linear Unit), to introduce nonlinear characteristics and improve the model's expressiveness. The extracted features are processed by pooling layers for dimensionality reduction, reducing the amount of calculation and the number of parameters while retaining important feature information. Then, the features are integrated and mapped at a high level through fully connected layers, converting the local features extracted by the convolutional layer into global feature representations. Finally, the output layer of the deep Q network maps the integrated features to the Q value of each possible replenishment decision through a linear activation function, which is expressed as 200 January 20 ; in, Indicates in status Take the replenishment decision The expected return value, represents the function of the deep Q network, is the set of network parameters. Through the combination of multi-layer convolution and fully connected layers, the deep Q network can effectively capture the complex features and nonlinear relationships in the input state vector and generate accurate Q value predictions. This process not only improves the model's ability to understand state features, but also enhances the model's decision-making ability in complex inventory management environments, ensuring that the generated Q value can truly reflect the potential benefits of each replenishment decision, thereby providing a solid foundation for subsequent decision-making choices.

[0034] After the deep Q network generates the Q values ​​of each replenishment decision, these Q values ​​are sorted and sorted according to the preset decision making. – The greedy strategy is a decision-making method that balances exploration and utilization. Its core idea is to choose the current optimal decision most of the time, while retaining a small probability to explore other possible decisions to prevent falling into the local optimum. In the specific implementation process, the Q values ​​of all replenishment decisions are first sorted, and the decision with the highest Q value is identified. Then, according to the preset ε value (such as 0.1), The probability of choosing ,by The probability of randomly selecting other replenishment decisions The selection process generates a set of candidate replenishment decisions , which includes the optimal decision and random decision. This strategy introduces a certain amount of randomness to ensure that the model still has the opportunity to explore potential better decisions while relying mainly on the known optimal decision, thereby improving the overall decision performance and the adaptability of the model. In the specific implementation, assuming that there is The Q value of each decision is first calculated, and then the Q value is compared with a randomly generated value. The value determines whether to choose the decision with the highest Q value or randomly select another decision. The formula is as follows: ; This process ensures the diversity and flexibility of decision-making options, enabling the inventory management system to make more intelligent and efficient replenishment decisions in a dynamic and uncertain environment, improving the overall responsiveness and operational efficiency of the supply chain. After selecting candidate replenishment decisions, action noise is applied to these decisions. By adding Gaussian noise with a mean of zero and a standard deviation that decays over time, the final replenishment decision is further optimized. In the specific implementation process, Gaussian noise is first generated. ,in Over time Gradually decay to reduce the impact of randomness on the decision. Then, the noise is added to the candidate replenishment decision On the other hand, we get the final optimal replenishment decision: ; The introduction of this action noise makes the replenishment decision have a certain degree of randomness and exploration while maintaining the main decision logic, which can cope with emergencies and demand fluctuations that may occur in the inventory management process. In addition, as time goes by, the standard deviation of the noise Gradually reduce, to ensure that the model gradually reduces randomness during the learning process, and improve the stability and accuracy of decision-making. In this way, the optimal replenishment decision Not only is it based on the accurate prediction of the deep Q network, but it also combines the flexible adjustment of action noise to ensure that the inventory management system can make efficient and reliable replenishment decisions in a dynamic and complex supply chain environment. This process builds an efficient and innovative inventory optimization mechanism through precise state vector processing, the powerful prediction ability of the deep learning model, and intelligent decision-making strategies, which significantly improves the responsiveness and operational efficiency of the supply chain.

[0035] 104. Perform multi-objective Pareto optimization on the target enterprise's inventory management based on the inventory optimization strategy, and obtain and implement the intelligent inventory management plan.

[0036] In one embodiment of the present invention, performing multi-objective Pareto optimization on the inventory management of the target enterprise according to the inventory optimization strategy to obtain and execute an intelligent inventory management plan includes: extracting key decision parameters of the inventory optimization strategy, wherein the key decision parameters include replenishment point, replenishment quantity and safety inventory level; using the decision parameters to construct a multi-objective optimization model that includes minimizing inventory cost, maximizing service level and minimizing supply chain risk; applying an improved multi-objective particle swarm algorithm to the multi-objective optimization model to generate a series of non-dominated solutions to form a Pareto front; performing fuzzy C-means clustering on the solutions on the Pareto front to obtain multiple inventory management plans, and using the TOPSIS method combined with the evaluation results of the inventory optimization strategy to comprehensively evaluate and rank the multiple inventory management plans to determine an intelligent inventory management plan; converting the intelligent inventory management plan into specific inventory control parameters and operating instructions, and distributing them to relevant departments for execution through the enterprise resource planning system of the target enterprise.

[0037] Specifically, the multi-objective Pareto optimization of the target enterprise's inventory management is carried out according to the inventory optimization strategy, and finally the intelligent inventory management plan is generated and implemented. First, the key decision parameters are extracted from the inventory optimization strategy, including the replenishment point, replenishment quantity and safety stock level. The replenishment point determines when the inventory needs to be replenished to prevent out-of-stock situations caused by a surge in demand or supply chain delays. The replenishment quantity determines the specific quantity of each replenishment to ensure that the inventory can meet the current demand without increasing the holding cost due to excessive backlog. The safety stock level is used to cope with the uncertainty of demand and the fluctuation of the supply chain, ensuring that a certain inventory level can still be maintained in emergencies to ensure that the service level is not affected. These key decision parameters directly affect the efficiency and cost control of inventory management. Therefore, in the extraction process, it is necessary to combine historical sales data, supplier delivery capabilities, market demand fluctuations and other information to ensure the accuracy and practicality of the parameters.

[0038] Specifically, the extracted decision parameters are then used to construct a multi-objective optimization model that includes inventory cost minimization, service level maximization, and supply chain risk minimization. The model comprehensively evaluates various aspects of inventory management by comprehensively considering multiple factors such as inventory holding costs, ordering costs, out-of-stock costs, and service levels. For example, inventory holding costs include warehousing costs, capital occupation costs, etc., ordering costs involve fixed and variable costs for each replenishment, and out-of-stock costs reflect potential revenue losses and reduced customer satisfaction due to the inability to meet customer needs in a timely manner. Service levels are usually measured by out-of-stock rates or order fulfillment rates, while supply chain risks include uncertainties such as supplier delivery delays and market demand fluctuations. The multi-objective optimization model weighs these goals to find the best balance between cost, service, and risk, ensuring that the inventory management strategy is both cost-effective and flexible to respond to market changes and supply chain uncertainties.

[0039] Specifically, next, the improved multi-objective particle swarm optimization algorithm (MOPSO) is applied to the constructed multi-objective optimization model to generate a series of non-dominated solutions and form a Pareto front. The MOPSO algorithm simulates the movement of particle swarms in the search space, iteratively optimizes multiple objective functions, and searches for a set of equilibrium solutions. These solutions achieve a balance between different objectives, and the improvement of any one objective does not come at the expense of significant deterioration of other objectives. On the basis of traditional particle swarm optimization, the improved MOPSO algorithm adds a diversity maintenance mechanism and a dynamic weight adjustment strategy, which enhances the algorithm's search ability and solution diversity when dealing with complex multi-objective optimization problems. Through multiple iterations, MOPSO can effectively explore and converge to the Pareto frontier, generate a set of non-dominated solutions representing different trade-offs, and provide a variety of alternatives for subsequent decision-making choices.

[0040] Specifically, after generating the Pareto frontier, the fuzzy C-means clustering (FCM) method is used to cluster the solutions on the frontier to obtain multiple inventory management solutions. FCM achieves flexible classification of solutions by defining the membership of each solution to each cluster, ensuring that each solution can partially belong to multiple clusters, thereby improving the accuracy and flexibility of clustering. The clustering results form several different inventory management solutions, each of which represents a specific target trade-off, such as low cost and high service level or high risk and low cost. Subsequently, the TOPSIS (Technique for Order Preference by Similarity to Ideal Solution) method is used to comprehensively evaluate and rank these inventory management solutions in combination with the evaluation results of the inventory optimization strategy. The TOPSIS method evaluates the relative advantages and disadvantages of each solution by calculating the distance between it and the ideal solution and the anti-ideal solution, ensuring that the final selected solution is as close to the ideal state as possible in all objectives. Through this comprehensive evaluation, the optimal intelligent inventory management solution is determined and converted into specific inventory control parameters and operation instructions, and finally distributed to relevant departments for execution through the target enterprise's enterprise resource planning (ERP) system. The entire process uses systematic multi-objective optimization and comprehensive evaluation to ensure that the generated inventory management solution achieves the best balance between cost, service and risk. In this embodiment, by collecting and preprocessing data from the target enterprise's enterprise resource planning system, supplier management system, and point of sale system, an integrated data set is obtained, and then multiple time-varying autoregressive demand forecasting models are constructed, and the prediction results of multiple time-varying autoregressive demand forecasting models are combined in combination with an integrated learning method to obtain dynamic demand forecasting results; according to the deep Q network model, the inventory replenishment decision is evaluated in combination with the dynamic demand forecasting results and the integrated data set to obtain an inventory optimization strategy; according to the inventory optimization strategy, the inventory management of the target enterprise is multi-objective Pareto optimized to obtain and execute an intelligent inventory management plan. The present invention significantly improves the comprehensiveness and dynamism of demand forecasting by comprehensively utilizing the data of various systems within the enterprise and applying advanced time-varying prediction models and integrated learning techniques.

[0041] The above describes the intelligent inventory management method in the embodiment of the present invention. The following describes the intelligent inventory management device in the embodiment of the present invention. Figure 2 In one embodiment of the present invention, an intelligent inventory management device includes: Data integration module 201, used to collect and pre-process data from the target enterprise's enterprise resource planning system, supplier management system, and point of sale system to obtain an integrated data set; The demand forecasting module 202 is used to construct multiple time-varying autoregressive demand forecasting models based on the integrated data set, and combine the forecast results of multiple time-varying autoregressive demand forecasting models by using an integrated learning method to obtain a dynamic demand forecasting result including a point forecast and a forecast interval; A replenishment decision module 203 is used to evaluate the inventory replenishment decision according to the pre-trained deep Q network model, combined with the dynamic demand forecast result and the integrated data set, to obtain an inventory optimization strategy; The inventory optimization module 204 is used to perform multi-objective Pareto optimization on the inventory management of the target enterprise according to the inventory optimization strategy, and obtain and execute an intelligent inventory management solution.

[0042] In an embodiment of the present invention, the intelligent inventory management device runs the above-mentioned intelligent inventory management method, and the intelligent inventory management device collects and preprocesses data from the enterprise resource planning system, supplier management system and point of sale system of the target enterprise to obtain an integrated data set, and then constructs multiple time-varying autoregressive demand forecasting models, and combines the prediction results of multiple time-varying autoregressive demand forecasting models in combination with an integrated learning method to obtain dynamic demand forecasting results; according to the deep Q network model, the inventory replenishment decision is evaluated in combination with the dynamic demand forecasting results and the integrated data set to obtain an inventory optimization strategy; according to the inventory optimization strategy, the inventory management of the target enterprise is multi-objective Pareto optimized to obtain and execute the intelligent inventory management plan. The present invention significantly improves the comprehensiveness and dynamism of demand forecasting by comprehensively utilizing the data of various systems within the enterprise and applying advanced time-varying prediction models and integrated learning technologies.

[0043] above Figure 2 The intelligent inventory management device in the embodiment of the present invention is described in detail from the perspective of modular functional entities. The intelligent inventory management device in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0044] Figure 33 is a schematic diagram of the structure of an intelligent inventory management device provided by an embodiment of the present invention. The intelligent inventory management device 300 may have relatively large differences due to different configurations or performances, and may include one or more processors (central processing units, CPU) 310 (for example, one or more processors) and a memory 320, and one or more storage media 330 (for example, one or more mass storage device terminals) storing application programs 333 or data 332. Among them, the memory 320 and the storage medium 330 can be short-term storage or permanent storage. The program stored in the storage medium 330 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations in the intelligent inventory management device 300. Furthermore, the processor 310 may be configured to communicate with the storage medium 330, and execute a series of instruction operations in the storage medium 330 on the intelligent inventory management device 300 to implement the steps of the above-mentioned intelligent inventory management method.

[0045] The intelligent inventory 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 and output interfaces 360, and / or one or more operating systems 331, such as Windows Serve, Mac OS X, Unix, Linux, FreeBSD, etc. It will be appreciated by those skilled in the art that Figure 3 The structure of the intelligent inventory management device shown does not constitute a limitation on the intelligent inventory management device provided by the present invention, and may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[0046] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the intelligent inventory management method.

[0047] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described system, device, or unit can refer to the corresponding process in the aforementioned method embodiment and will not be repeated here.

[0048] If the integrated unit is implemented in the form of 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 is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0049] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent inventory management method, characterized in that: The intelligent inventory management method comprises: Collect and preprocess data from the target enterprise's enterprise resource planning system, supplier management system, and point-of-sale system to obtain an integrated data set; According to the integrated data set, multiple time-varying autoregressive demand forecasting models are constructed, and the forecast results of multiple time-varying autoregressive demand forecasting models are combined with an integrated learning method to obtain a dynamic demand forecasting result including a point forecast and a forecast interval; According to the pre-trained deep Q network model, the inventory replenishment decision is evaluated in combination with the dynamic demand forecast results and the integrated data set to obtain an inventory optimization strategy; The inventory management of the target enterprise is subjected to multi-objective Pareto optimization according to the inventory optimization strategy, and an intelligent inventory management solution is obtained and executed.

2. The intelligent inventory management method according to claim 1, characterized in that: The method of constructing multiple time-varying autoregressive demand forecasting models based on the integrated data set and combining the forecast results of multiple time-varying autoregressive demand forecasting models with an integrated learning method to obtain a dynamic demand forecasting result including a point forecast and a forecast interval includes: Performing time series decomposition on the integrated data set to obtain trend components, seasonal components and random components; Based on the trend, seasonality and random components, multiple time-varying autoregressive models with different autoregressive orders and exogenous variable combinations are constructed; The Kalman filter algorithm is used to estimate and update the parameters of each time-varying autoregressive model to obtain the prediction results of each time-varying autoregressive model; The prediction results of each model are weighted averaged to obtain a comprehensive prediction result. The Monte Carlo simulation method is used to perform multiple sampling and calculations on the comprehensive prediction result to obtain a dynamic demand prediction result that includes point prediction and prediction interval.

3. The intelligent inventory management method according to claim 2, characterized in that: The constructing of multiple time-varying autoregressive models with different autoregressive orders and exogenous variable combinations according to the trend, seasonality and random components includes: Perform polynomial fitting on the trend component to obtain the trend function that describes the long-term trend. Perform Fourier series expansion on the seasonal component to obtain the seasonal function that describes the periodic change. Perform autocorrelation analysis on the random component to determine the maximum autoregressive order. Selecting exogenous variables from the integrated data set, and constructing a preliminary autoregressive model based on the trend function, seasonal function and autoregressive terms, wherein the exogenous variables include economic indicators, market data and supply chain characteristics; Each exogenous variable is added to the preliminary autoregressive model one by one to form multiple extended models, and each extended model is evaluated using an information criterion to select an optimal subset of exogenous variables; A grid search method is used to generate all combinations of different autoregressive orders and different exogenous variables in the exogenous variable subset, and for each combination, a corresponding time-varying autoregressive model is constructed.

4. The intelligent inventory management method according to claim 2, characterized in that: The prediction results of each model are weighted averaged to obtain a comprehensive prediction result, and the Monte Carlo simulation method is used to perform multiple sampling and calculations on the comprehensive prediction result to obtain a dynamic demand prediction result including point prediction and prediction interval, including: The root mean square of the historical forecast error of each time-varying autoregressive model during the parameter estimation and updating process is calculated, and the weight of each time-varying autoregressive model is calculated according to the inverse of the root mean square of each time-varying autoregressive model; By calculating the weights, weighted averaging is performed on the prediction results of the corresponding time-varying autoregressive models to obtain a comprehensive prediction result, and a normality test is performed on the error term of the comprehensive prediction result to obtain an error distribution parameter; Using the Monte Carlo method, samples are drawn from the error distribution parameters multiple times, and the drawn samples are added to the comprehensive prediction result each time to obtain multiple simulated prediction values; The average value of the simulated prediction values ​​is calculated as the point prediction, and the quantiles at a specific confidence level are calculated as the upper and lower bounds of the prediction interval to obtain a dynamic demand forecast result including the point prediction and the prediction interval.

5. The intelligent inventory management method according to claim 1, characterized in that: The inventory replenishment decision is evaluated based on the pre-trained deep Q network model in combination with the dynamic demand forecast results and the integrated data set to obtain the inventory optimization strategy including: encoding the dynamic demand forecast result and the current inventory level, supplier reliability index and inventory holding cost in the integrated data set into a state vector; Input the state vector into a pre-trained deep Q network model to obtain Q values ​​of various possible replenishment decisions, and select the optimal replenishment decision based on the Q values ​​and a preset greedy strategy; The selected replenishment decision is simulated and executed, the expected inventory cost and service level are calculated, and based on the simulation results, the replenishment decision is fine-tuned to obtain the inventory optimization strategy.

6. The intelligent inventory management method according to claim 5, characterized in that: The step of inputting the state vector into a pre-trained deep Q network model to obtain Q values ​​of various possible replenishment decisions, and selecting the optimal replenishment decision according to the Q values ​​and a preset greedy strategy includes: Input the state vector into the pre-trained deep Q network model, normalize the input state vector, and obtain a standardized state representation; Input the standardized state representation into the input layer of the deep Q network, perform forward propagation calculation through a multi-layer convolutional neural network and a fully connected layer, and obtain the Q value of each possible replenishment decision; The Q values ​​are sorted, and according to a preset ε-greedy strategy, a replenishment decision with the highest Q value is selected with a probability of 1-ε, and a replenishment decision is randomly selected with a probability of ε to obtain a candidate replenishment decision; Action noise is applied to the candidate replenishment decisions to obtain the optimal replenishment decision by adding Gaussian noise with zero mean and standard deviation decaying over time.

7. The intelligent inventory management method according to claim 1, characterized in that: The performing multi-objective Pareto optimization on the inventory management of the target enterprise according to the inventory optimization strategy to obtain and execute the intelligent inventory management solution comprises: According to extracting the key decision parameters of the inventory optimization strategy, wherein the key decision parameters include replenishment point, replenishment quantity and safety stock level; Using the decision parameters, a multi-objective optimization model including minimizing inventory cost, maximizing service level and minimizing supply chain risk is constructed; An improved multi-objective particle swarm algorithm is applied to the multi-objective optimization model to generate a series of non-dominated solutions and form the Pareto front; Perform fuzzy C-means clustering on the solutions on the Pareto front to obtain multiple inventory management solutions, and use the TOPSIS method combined with the evaluation results of the inventory optimization strategy to comprehensively evaluate and rank the multiple inventory management solutions to determine the intelligent inventory management solution; The intelligent inventory management solution is converted into specific inventory control parameters and operation instructions, and distributed to relevant departments for execution through the target enterprise's enterprise resource planning system.

8. An intelligent inventory management device, characterized in that: The intelligent inventory management device comprises: The data integration module is used to collect and preprocess data from the target enterprise's enterprise resource planning system, supplier management system, and point of sale system to obtain an integrated data set; A demand forecasting module is used to construct multiple time-varying autoregressive demand forecasting models based on the integrated data set, and combine the forecast results of multiple time-varying autoregressive demand forecasting models in combination with an integrated learning method to obtain a dynamic demand forecasting result including a point forecast and a forecast interval; A replenishment decision module, used to evaluate the inventory replenishment decision based on the pre-trained deep Q network model, combined with the dynamic demand forecast results and the integrated data set, to obtain an inventory optimization strategy; The inventory optimization module is used to perform multi-objective Pareto optimization on the inventory management of the target enterprise according to the inventory optimization strategy, and obtain and execute an intelligent inventory management solution.

9. An intelligent inventory management device, characterized in that: The intelligent inventory management device comprises: a memory and at least one processor, wherein instructions are stored in the memory; The at least one processor calls the instructions in the memory to enable the intelligent inventory management device to perform the steps of the intelligent inventory management method as described in any one of claims 1-7.

10. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the steps of the intelligent inventory management method as described in any one of claims 1-7 are implemented.

Images