Multi-target inventory optimization method based on dynamic demand prediction and intelligent decision-making system

Abstract

The invention discloses a multi-target inventory optimization method based on dynamic demand prediction and an intelligent decision-making system, and the method comprises the steps: firstly collecting a cache library inventory state of a cache library system, including considering production order data and AGV transportation capability data, and carrying out the standardization processing of the cache library inventory state; inputting the data into an LSTM neural network model of a dynamic demand prediction module to predict raw and auxiliary material demands in a future time period; the replenishment amount and the replenishment period are solved based on an improved multi-objective optimization algorithm NSGA-II, and the optimization objectives are the lowest total cost, the lowest stockout rate and the lowest inventory fluctuation; issuing a replenishment instruction to the cache library system through the production management system, and scheduling the AGV to execute a transportation task; and the inventory state is monitored in real time in the production process, strategy adjustment or model retraining is triggered according to the fluctuation coefficient, and finally a zero-stockout, low-fluctuation and optimal-cost replenishment strategy is output, so that the method is suitable for a high-dynamic manufacturing environment.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent manufacturing and supply chain optimization, and is applicable to real-time material scheduling and warehouse management in manufacturing industries such as prepregs and composite materials. In particular, it relates to a multi-objective inventory optimization method and intelligent decision-making system based on dynamic demand forecasting. Background Art

[0002] As a key intermediate material for high-performance composite materials, prepregs are experiencing a continuous growth in demand for applications in aviation, aerospace, shipbuilding, automotive, and other fields. The production process of prepreg products involves the automatic logistics and transportation of raw and auxiliary materials and finished products, the threading and unthreading of roll-type raw and auxiliary materials, automatic loading and unloading production, automatic feeding for inspection, automatic packaging, automatic transportation of waste products, and information traceability. The prepreg production process requires high levels of equipment coordination, involving the multi-link coordination of more than ten kinds of raw and auxiliary materials such as carbon fiber, release paper, PE film, and adhesive film. It is also constrained by dynamic factors such as equipment loading and unloading rhythm, AGV transportation capacity, and storage capacity, making it difficult for traditional inventory management models to adapt to complex production needs.

[0003] In traditional prepreg production, fixed thresholds or order quantity (EOQ) models are used for production and replenishment, which relies on manual experience and cannot adapt to dynamic demand and multi-objective conflicts. Due to unreasonable safety stock settings, it is easy to cause frequent stock-outs or inventory backlogs. When emergency orders are inserted in production or equipment fails, traditional methods cannot quickly generate dynamic strategies that take into account both robustness and economy. The consumption rate of raw and auxiliary materials in prepreg production is significantly affected by fluctuations in production orders, while traditional static models cannot capture timing dependency characteristics and have inventory response lags. They cannot perform dynamic real-time responses to prepreg production processes with high-frequency and multi-variable demand characteristics. In addition, prepreg production needs to balance multi-objective conflicts such as total cost, equipment downtime risks caused by stock-outs, and inventory fluctuations caused by storage capacity limitations. Traditional single-objective optimization algorithms or manual rules are difficult to quantify complex constraints such as equipment loading and unloading coordination, replenishment rate, and storage capacity limitations. Summary of the Invention

[0004] Purpose of the invention: The present invention aims to provide a multi-objective inventory optimization method and intelligent decision-making system based on dynamic demand forecasting. The present invention links LSTM prediction with NSGA-II multi-objective optimization, and adopts a dynamic constraint processing strategy based on equipment loading and unloading rhythm matching to solve the complex production demand problem of prepreg mentioned in the above background technology.

[0005] Technical solution: A multi-objective inventory optimization method based on dynamic demand forecasting, including:

[0006] S1. Collect cache inventory status, production order data, and AGV transportation capacity data, and standardize the obtained data, including generating a standardized data set D norm ;

[0007] S2. Build a dynamic demand forecasting model based on the LSTM neural network to predict the demand for raw materials and auxiliary materials in the future period; the dynamic demand forecasting model includes the following objective function:

[0008] Objective function 1 is to minimize the total cost of replenishment cost, inventory holding cost and out-of-stock penalty cost, which can be expressed mathematically as follows:

[0009]

[0010] Objective function 2 is to minimize the risk of out-of-stock, and its mathematical expression is:

[0011]

[0012] Objective function 3 is to minimize inventory volatility, and its mathematical expression is:

[0013]

[0014] Where C order,i represents the replenishment cost of material i, which includes AGV scheduling, equipment energy consumption, and manual inspection costs. i represents the forecast demand rate of material i, Q i represents the replenishment quantity of material i, C hold,i represents the inventory holding cost of material i, T i represents the replenishment cycle of material i, C short,i represents the out-of-stock penalty cost of material i, I i0 represents the initial inventory of material i; σ(I i (t)) is the standard deviation of the inventory level of the i-th material; u(I i (t)) is the mean inventory level of the i-th material; I i (t) is the inventory level function of the i-th type of material at time t; n is the total number of materials. The smaller the Ti / Pi ratio is, the closer the replenishment cycle is to the equipment loading and unloading capacity, and the lower the inventory fluctuation risk is.

[0015] S3. Solve the dynamic demand forecasting model described in step S2 based on the improved NSGA-II algorithm. The optimization objectives are to minimize total cost, minimize out-of-stock rate, and minimize inventory fluctuation. Output the Pareto optimal solution set and replenishment quantity Q. i and replenishment cycle T i Replenishment strategy;

[0016] The following constraints are considered during the solution process:

[0017] Replenishment and cycle non-negative constraints:

[0018] Q i≥0,T i ≥0,I i0 ≥0(i=1,2,…,n)

[0019] Cache library location capacity limit:

[0020] 0≤I i (t)+Q i ≤S i

[0021] Where, I i (t) represents the inventory of material i at time t, S i Indicates the upper limit of the storage capacity of material i;

[0022] Equipment loading and unloading rhythm matching:

[0023]

[0024] T i represents the replenishment cycle of material i; P i represents the loading and unloading cycle of the production equipment associated with material i; k i is the loading and unloading efficiency coefficient of material i, k ii ≥1;

[0025] Inventory dynamic constraints:

[0026] Q i ≥D i ·T i -I i0

[0027] AGV transport capacity limitations:

[0028]

[0029] t k represents the time taken for the kth AGV transport task; T AGV,max Indicates the maximum continuous working time of AGV;

[0030] S4. Based on the solution of step S3, the production management system sends a replenishment instruction to the cache system and schedules the AGV to perform the transportation task;

[0031] S5. Inventory fluctuation warning and dynamic adjustment: Real-time monitoring of inventory status, triggering strategy adjustment or retraining of dynamic demand forecasting models based on the fluctuation coefficient, including adjusting the objective function weight based on the inventory fluctuation coefficient. The fluctuation coefficient calculation formula is:

[0032]

[0033] I i represents the daily inventory of material i, u(Ii ) represents the average inventory level of material i, σ(I i ) represents the standard deviation of the inventory level of material i.

[0034] Furthermore, the LSTM neural network structure of the dynamic demand forecasting model includes:

[0035] Bidirectional two-layer LSTM unit, each LSTM unit's gating mechanism is a forget gate, the input gate uses a sigmoid activation function, and the candidate state uses a tanh activation function; the input layer dimension is consistent with the number of features, and the time step window is 24 hours;

[0036] The output of the LSTM layer is followed by a fully connected layer with ReLU as the activation function to extract high-order temporal features. A Dropout layer is added after the fully connected layer to prevent overfitting, and L2 regularization is used on the weights to optimize the generalization ability of the model.

[0037] The output layer uses the linear activation function Linear to directly predict material demand in the next 6 hours;

[0038] The loss function of the training strategy is the weighted mean square error W-MSE, the weight decays over time, and the Adam optimizer dynamically adjusts the learning rate.

[0039] Furthermore, the dynamic demand forecasting model is trained using cache inventory status data, which includes real-time inventory information, historical raw material consumption data, and insufficient inventory alarm information, for training the dynamic demand forecasting model.

[0040] The training process uses the root mean square error (RMSE) or the mean absolute percentage error (MAPE) as the model accuracy judgment indicator. When RMSE>δ×RMSEbase or MAPE is greater than the set threshold, the model retraining is triggered. δ is the adjustment coefficient, and its value is greater than or equal to 1. RMSEbase is the benchmark error in the model verification stage; the final output is the forecast value of raw material and auxiliary material demand from t to t+6 hours in the future.

[0041] The present invention takes into account that the storage costs incurred during the inventory period are positively correlated with the storage time and occupied space, and mathematically expresses them as follows:

[0042] C hold,i =α i ·V i ·T hold

[0043] α i V is the daily storage cost per unit volume; i is the volume of a single piece of material; T hold is the average storage time;

[0044] The direct economic loss caused by out-of-stock includes downtime loss and emergency restocking premium, so the out-of-stock penalty cost is mathematically expressed as:

[0045] C short,i =β i ·S i +γ i ·b(S i >0)

[0046] β i is the downtime loss per unit of stock-out; γ i b(S i >0) is an indicator function, which takes 1 when out of stock and 0 otherwise;

[0047] The standard deviation of the inventory level of material i σ(I i (t)) is defined as follows:

[0048]

[0049] u i is the average inventory of the i-th category material.

[0050] In this method, the improved NSGA-II algorithm introduces adaptive crossover probability:

[0051]

[0052] Among them, P max is the initial maximum crossover probability, P min is the minimum crossover probability, Gen is the current evolutionary generation, Gen max is the maximum evolutionary generation.

[0053] The present invention also provides an intelligent decision-making system based on multi-objective inventory management, which implements the above method to perform dynamic inventory demand forecasting. The system includes:

[0054] Inventory dynamic monitoring module, used to obtain material inventory information in real time, including material consumption and equipment energy consumption;

[0055] Dynamic demand forecasting module, used to execute the dynamic demand forecasting model and output the forecast value of raw material and auxiliary material demand in a certain time period in the future;

[0056] The multi-objective optimization module obtains the forecast demand data transmitted by the dynamic demand forecast module and the AGV status data of the AGV scheduling system. Based on the NSGA-II algorithm, it selects new populations through non-dominated sorting stratification, genetic evolution, and elite strategy selection, and outputs the Pareto optimal solution set and replenishment quantity Q. i and replenishment cycle T i Replenishment strategy;

[0057] The replenishment strategy execution module receives the replenishment strategy generated by the multi-objective optimization module and transmits it to the production management system to convert it into production replenishment instructions, including material type, quantity, target storage location and latest arrival time;

[0058] The feedback and adaptive module does not trigger replenishment if the inventory is sufficient. Otherwise, it generates a replenishment instruction according to the replenishment strategy in step S3 of the above method and transmits it to the production management system to execute the replenishment operation.

[0059] Compared with the prior art, the advantages and beneficial effects of the present invention are:

[0060] (1) The neural network LSTM time series modeling is combined with the improved NSGA-II multi-objective optimization algorithm. The LSTM is used to capture the time series characteristics such as equipment operation status, historical orders, inventory fluctuations, etc. in real time. The above multi-modal data are fused to form a dynamic demand forecast for multi-specification materials. The dynamic demand forecast results are input into the improved NSGA-II optimization algorithm, which significantly improves the input data accuracy of the multi-objective optimization. Compared with the traditional ARIMA model and the single genetic algorithm, the prediction accuracy of the deep fusion model of the dynamic demand forecast model based on deep learning and the improved multi-objective optimization algorithm is significantly improved, and the replenishment response time is significantly shortened;

[0061] (2) Compared with the traditional static inventory model and single-objective optimization algorithm, the use of LSTM and the improved NSGA-II multi-objective algorithm can achieve the coordinated optimization of total cost, out-of-stock rate, and inventory fluctuation, improve the robustness of the system, and dynamically adjust the safety stock level based on the LSTM prediction results to achieve adaptive optimization of the safety stock, which can effectively solve the problem of inventory redundancy or out-of-stock caused by prediction lag;

[0062] (3) By building an intelligent inventory optimization system covering data perception, dynamic demand forecasting, multi-objective optimization, replenishment execution and adaptive feedback closed loop, the replenishment strategy can be dynamically adjusted to achieve accurate matching of demand fluctuations and replenishment strategies, solving the problems of dynamic demand response lag, multi-objective conflict quantification difficulties, and low efficiency of equipment and logistics coordination in prepreg production scenarios. In particular, it can effectively respond to emergencies such as emergency task insertions and equipment failures that occur during the prepreg production process, thereby reducing the company's production costs, improving the resilience of the supply chain, and laying a good foundation for the company's digital transformation. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 This is an overall flow chart of the method of the present invention;

[0064] Figure 2 Optimization flow chart for the LSTM and improved NSGA-II composite optimization algorithm;

[0065] Figure 3 This is the flowchart of AGV task execution in the multi-objective inventory optimization method based on dynamic demand forecasting;

[0066] Figure 4 It is the application framework of the method of the present invention;

[0067] Figure 5 It is a flowchart of the film production process. DETAILED DESCRIPTION

[0068] The specific technical solutions of the present invention are further described below with reference to the accompanying drawings to help those skilled in the art further understand the present invention without limiting the rights thereof.

[0069] The method described in the present invention achieves precise execution of demand fluctuations and replenishment strategies by constructing an intelligent inventory optimization system covering data perception, dynamic demand forecasting, multi-objective optimization, replenishment execution and dynamic adjustment. It solves the problems existing in prepreg production scenarios, such as delayed dynamic demand response, difficulty in quantifying multi-objective conflicts, and low efficiency of equipment and logistics collaboration. It can effectively solve the problem of inventory redundancy or downtime caused by shortages due to forecast lags, reduce enterprise production costs, and improve supply chain resilience.

[0070] Furthermore, the implementation of the present invention can be combined with the production management system as a basis, and the necessary parts include: first, the data acquisition and preprocessing module collects the cache inventory status of the cache system, the production order data of the production management system and the AGV transportation capacity data of the AGV scheduling system, and standardizes the above data; the above standardized data set, the time series data of the historical consumption of raw and auxiliary materials, and the inventory fluctuation warning information are input into the LSTM neural network model of the dynamic demand prediction module, and the raw and auxiliary material demand in the future period is predicted by the LSTM neural network; the replenishment quantity Q is solved based on the improved multi-objective optimization algorithm NSGA-II i and replenishment cycle T i The optimization goals are to minimize total cost, lowest out-of-stock rate and smallest inventory fluctuation; the production management system sends replenishment instructions to the cache system, and dispatches AGV to perform transportation tasks; and monitors inventory status in real time during the production process, triggers policy adjustments or model retraining based on the fluctuation coefficient, and finally outputs a replenishment strategy with zero out-of-stock, low fluctuation and optimal cost, which is suitable for highly dynamic manufacturing environments.

[0071] Specifically, combined Figure 1 , the implementation steps of the present invention are as follows:

[0072] S1. Data collection and preprocessing: Collect cache inventory status, production order data, and AGV transportation capacity data, and standardize the above data.

[0073] In combination with the system, the acquisition of cache inventory status refers to the data acquisition and preprocessing module acquiring real-time inventory information, raw material and auxiliary material historical consumption data, and insufficient inventory alarm information from the cache system;

[0074] The acquisition of production order data refers to the data acquisition and preprocessing module obtaining production order data (product number, quantity, delivery date), production plan, equipment parameters (temperature, pressure, speed, beat, working status), and historical production data information from the production management system;

[0075] The collection of AGV transport capacity data refers to the data collection and preprocessing module obtaining AGV transport capacity constraints, optimization strategies, and transport records (time consumption, path) from the AGV scheduling system;

[0076] S2. Build a dynamic demand forecasting model: Build a dynamic demand forecasting model based on the LSTM neural network to predict the demand for raw materials and auxiliary materials in a certain period of time in the future;

[0077] The dynamic demand forecasting model inputs the above-mentioned standardized data set, time series data of historical consumption of raw and auxiliary materials, and inventory fluctuation warning information into the LSTM neural network model of the dynamic demand forecasting module, trains the model until the prediction accuracy meets the set requirements, and uses the root mean square error (RMSE) or the mean absolute percentage error (MAPE) as the model accuracy judgment indicator. When RMSE>δ×RMSEbase or MAPE is greater than the set threshold, the model retraining is triggered. Here, δ is the adjustment coefficient, its value is greater than or equal to 1, and RMSEbase is the benchmark error in the model verification stage; the final output is the forecast value of raw and auxiliary material demand from t to t+6 hours in the future;

[0078] The LSTM neural network structure is a bidirectional two-layer LSTM unit. The gating mechanism of each LSTM unit is a forget gate, the input gate uses a sigmoid activation function, and the candidate state uses a tanh activation function. The input layer dimension is consistent with the number of features, and the time step window is 24 hours.

[0079] The output of the LSTM layer is followed by a fully connected layer with a ReLU activation function to extract high-order temporal features. A Dropout layer is added after the fully connected layer to prevent overfitting, and L2 regularization is used on the weights to optimize the generalization ability of the model.

[0080] The output layer uses a linear activation function to directly predict material demand for the next 6 hours. The loss function of the training strategy is the weighted mean square error (W-MSE), the weight decays over time, and the Adam optimizer dynamically adjusts the learning rate.

[0081] S3. Construction and solution of multi-objective optimization model

[0082] The mathematical model is solved based on the improved NSGA-II algorithm. The model is a multi-objective optimization model. The optimization objectives are to minimize the total cost, the lowest out-of-stock rate and the smallest inventory fluctuation. The Pareto optimal solution set and the replenishment quantity Q are output. i and replenishment cycle T i Replenishment strategy;

[0083] S4. Replenishment Strategy Execution

[0084] The production management system sends replenishment instructions to the cache system, and dispatches AGV to perform transportation tasks;

[0085] S5. Inventory fluctuation warning and dynamic adjustment

[0086] Monitor inventory status in real time and trigger strategy adjustments or model retraining based on fluctuation coefficients.

[0087] Based on the above technical solution, the present invention provides a dynamic demand forecasting system, whose framework includes: data acquisition and preprocessing module, dynamic demand forecasting module, multi-objective optimization module, replenishment strategy execution module, inventory monitoring module, feedback and adaptation module;

[0088] Data Collection and Preprocessing Module: This module executes step S1 of the aforementioned method and is responsible for the real-time collection, cleaning, and standardization of multi-source data, providing high-quality input for subsequent prediction and optimization. By integrating interfaces with the production management system, cache system, and AGV scheduling system, it collects equipment operating parameters, cache inventory status, and AGV transport capacity data. The module features a built-in multi-protocol adaptation engine that supports industrial protocols such as OPC-UA and MQTT, enabling seamless access to heterogeneous device data. It also uses dynamic sliding window technology to adaptively adjust data processing granularity, ensuring that data characteristics are synchronized with the production rhythm.

[0089] Dynamic Demand Forecasting Module: Executes step S2 of the above method and constructs a demand forecasting model based on a long short-term memory (LSTM) network. This model inputs historical consumption data, equipment status, and order priority labels. A two-layer LSTM unit captures the nonlinear relationship between temporal dependencies and equipment operating modes, and outputs forecasts of future material demand over multiple time periods. A weighted mean square error (W-MSE) is introduced as a loss function, and the forecast results are transmitted to the multi-objective optimization module via a protocol buffer. A dynamic retraining mechanism is also designed. When the continuous forecast error exceeds a threshold or a sudden change in the equipment operating mode is detected, incremental training is automatically triggered to update the model parameters, ensuring that the forecast capability is continuously optimized as the production environment changes.

[0090] Multi-Objective Optimization Module: Executes step S3 of the aforementioned method and uses the improved NSGA-II algorithm to solve the replenishment policy. The objective function includes the total cost of replenishment and inventory holding, the cost of out-of-stock penalties, and the standard deviation of inventory fluctuation. The algorithm uses real-number encoding to represent the replenishment quantity and replenishment cycle, uses simulated binary crossover and polynomial mutation operations to generate a child population, and combines non-dominated sorting and congestion calculation to select high-quality solutions. The optimization results are output as a multidimensional objective space solution set, supporting decision makers in selecting the final strategy based on business preferences.

[0091] The replenishment strategy execution module executes step S4 of the above method, receives the replenishment strategy generated by the multi-objective optimization module, and converts it into production task replenishment instructions through the production management system, including key information such as material type, quantity, target storage location, and latest arrival time. The production management system sends the replenishment instructions to the cache system, which generates inbound and outbound tasks based on the replenishment instructions. The AGVs execute the transportation tasks and distribute them to idle AGVs through a load balancing strategy, ensuring that transportation response times meet production cycle requirements. The module supports digital signature verification of instructions and interception of abnormal instructions to ensure operational security.

[0092] Inventory monitoring module, feedback and adaptive module: Execute step S5 of the above method, the inventory monitoring module tracks the inventory status of the cache in real time, dynamically sets the safety stock threshold and monitors its compliance. When it is detected that the inventory fluctuation exceeds the preset threshold, or the inventory deviation continues to expand, an alarm signal is triggered and the warning information is pushed to the feedback and adaptive module. Among them, the feedback and adaptive module is used to build a closed-loop feedback mechanism to enhance the robustness of the system. The comprehensive cost reduction index within the cycle, the number of out-of-stock events, and the improvement index of inventory turnover rate can be counted, and the multi-objective optimization weights can be dynamically adjusted according to the inventory fluctuation coefficient. In order to solve the performance degradation problem of the prediction model, a triggered retraining logic is designed to trigger the dynamic demand prediction module to retrain the LSTM model, and the training results are stored in the knowledge base for optimizing the initial population generation strategy for the next cycle.

[0093] Example 1

[0094] Combine Figure 1 and Figure 2 In this embodiment 1, the dynamic demand forecasting and multi-objective inventory optimization application in the film production process is taken as an example. There are 6 film machines in total (the corresponding loading points are represented as M1-M6). The film production process is as follows: the raw and auxiliary materials are shipped from the cache. Except for the resin that is directly transported to the film machine, the other rolls are first transferred to the shaft-threading position by the AGV. After the rolls and thin paper tubes are threaded, they are transferred to the loading position of the film machine. The truss manipulator grabs the rolls or thin paper tubes for loading and then starts film production; Among them, the AGV operation: after a single full charge, it works for 6 hours and is responsible for the synchronous transportation of wide and narrow raw and auxiliary materials; the cache is limited by the spatial layout, and its storage position upper limit is I maxThe demand for raw materials and auxiliary materials in the film production process is strongly correlated with the film machine specifications (wide / narrow), requiring differentiated forecasting and replenishment. Specifically, combining existing film production processes and existing production management systems, this invention provides the following application scenarios.

[0095] The production management system issues production orders to the intelligent decision-making system via an API. The intelligent decision-making system synchronizes production task information with the workshop cache system and also provides feedback to the production management system regarding order completion, AGV equipment status, and material requirements. Materials such as release paper, PE film, and fine and coarse paper tubes are conveyed via conveyor lines into the workshop cache. Before entering the cache, the materials are bound to the pallet barcode by scanning it. The intelligent decision-making system sends the scanned pallet barcode to the production management system. After retrieving the material and pallet information associated with the barcode, the production management system transmits the detailed information back to the cache system. After the intelligent decision-making system receives the incoming materials and puts them on the shelves, it synchronizes the incoming information with the production management system.

[0096] After the production site calls for materials, raw and auxiliary materials such as release paper, PE film, thin paper tubes, and thick paper tubes are shipped from the buffer warehouse. An AGV first transports the rolls or tubes shipped from the buffer warehouse to the threading machine. The production management system then issues the threading task to the threading machine, completing the threading process. After receiving the threading completion signal, the intelligent decision-making system synchronizes the signal with the AGV scheduling system, dispatching the AGV to the threading machine exit to retrieve the material. Based on the material information, the AGV is then delivered to the loading position of the corresponding buffer rack for the film. Once the AGV is finished, a completion signal is sent back to the intelligent decision-making system.

[0097] The intelligent decision-making system issues a task to the film machine truss to load a new roll of material. After the truss is safely confirmed, when the truss scans the code to load the material, it calls the production management system interface to perform a material legitimacy check and transmits the verification result back to the intelligent decision-making system. If the verification passes, the production management system automatically generates a work order material delivery record. If the verification finds an error, the material is returned to the cache on site. The film machine truss takes the material from the cache rack and places it into the film machine chuck. After closing the chuck, the film machine truss sends a completion signal to the intelligent decision-making system, which replies to the production management system with a successful material request. Among them, the AGV operation: After a single full charge, it can work for 6 hours and is responsible for the simultaneous transportation of wide-width and narrow-width raw and auxiliary materials; the cache is limited by the spatial layout, and its storage capacity limit is Imax; the demand for raw and auxiliary materials in the film production process is strongly related to the specifications of the film machine (wide / narrow), requiring differentiated prediction and replenishment.

[0098] At present, a fixed threshold or order quantity (EOQ) model is used for production replenishment in prepreg production, that is, when a certain material is lower than the safety stock, such as when it is lower than the delivery quantity of one full pallet, the cache library initiates a replenishment request and sends the corresponding replenishment information to the production management system to apply for replenishment. The safety stock setting of this replenishment strategy relies on manual experience, cannot capture the timing dependency characteristics, has a lag in inventory response, cannot perform dynamic real-time response to the prepreg production process with high frequency and multivariate demand characteristics, and is prone to frequent out-of-stock. If the safety stock is set too large, it is easy to cause inventory backlog problems, resulting in inventory waste, thereby reducing production efficiency and resulting in higher production costs. In order to solve the above-mentioned technical problems, the present invention has developed an intelligent decision-making system that can realize multi-objective inventory optimization based on dynamic demand forecasting, which is used for real-time dynamic material scheduling in production processes such as prepregs and composite materials.

[0099] Combine Figure 3-5 During the production process, the intelligent decision-making system obtains the inventory status of the cache, the production order data issued by the production management system, the AGV transportation capacity data, the data of the processing and loading and unloading of rolls by equipment such as the shaft threading machine and the film machine in real time. The LSTM neural network model constructed by the dynamic demand prediction module obtains the predicted demand data, and solves the optimal replenishment strategy (replenishment quantity and replenishment cycle) through the improved multi-objective optimization algorithm. The replenishment instructions are sent to the cache system through the production management system, and the task information is synchronized to the AGV scheduling system to schedule the AGV to perform the transportation task. The intelligent decision-making system monitors the inventory status in real time during the production process, triggers strategy adjustment or model retraining according to the fluctuation coefficient, and finally outputs a replenishment strategy with zero out-of-stock, low fluctuation and optimal cost, which is suitable for highly dynamic manufacturing environments.

[0100] The above-mentioned intelligent decision-making system is the multi-objective inventory management decision-making system provided by the present invention, which includes a data acquisition and preprocessing module, a dynamic demand forecasting module, a multi-objective optimization module, a replenishment strategy execution module, an inventory monitoring module, a feedback and adaptation module, and integrates and interacts with the production management system, the cache system, and the AGV scheduling system for information data to achieve precise execution of demand fluctuations and replenishment strategies.

[0101] Further integration Figure 1-5 , the specific implementation steps of the present invention are as follows:

[0102] Step 1. Collect production order data, AGV transport capacity data, and cache inventory status, and standardize the above data;

[0103] The standardization process refers to cleaning and normalizing inventory data, equipment operation data, order data, and AGV transportation capacity data to generate a standardized data set D norm ;

[0104] Among them, the normalization method adopts Min-Max normalization, and the formula is:

[0105]

[0106] The acquisition of production order data refers to the data acquisition and preprocessing module obtaining production order data (product number, quantity, delivery date), production plan, equipment parameters (temperature, pressure, speed, beat, working status), and historical production data information from the production management system;

[0107] Product specifications: wide-width film (such as order number xx-001, width 1000mm, daily output 10,000 meters, delivery time 24 hours), narrow-width film (order number xx-305, width 650mm, daily output 4,000 meters, delivery time 12 hours).

[0108] Equipment parameters: The temperature of the wide-width film machine is set to 25℃±2℃, the pressure is 0.2-0.4MPa, the compounding speed is 25m / min, and the loading and unloading rhythm is to complete the material roll replacement every 1.6 hours; the loading and unloading rhythm of the narrow-width film machine is once every 0.85 hours.

[0109] Historical production data: Statistical analysis of the average daily consumption of release paper rolls, PE film rolls, and thin paper rolls for wide-format films over the past 30 days.

[0110] Normalization is performed based on order priority, with emergency orders marked as 1 and regular orders marked as 0. The order is then normalized to the [0, 1] interval using Min-Max normalization.

[0111] The collection of AGV transport capacity data refers to the data collection and preprocessing module obtaining AGV transport capacity constraints, optimization strategies, and transport records (time consumption, path) from the AGV scheduling system;

[0112] From the cache to the threading machine: The standard route takes 2 minutes, which increases to 3 minutes during congestion.

[0113] From the shaft threading machine to the film machine loading position: The standard route takes 4.5 minutes, transporting one roll of release paper or PE film and two rolls of thin paper tubes at a time.

[0114] Transport capacity constraints: The maximum continuous working time of a single AGV is 6 hours, and it can complete a maximum of 15 transport tasks (wide width) or 18 (narrow width) per hour.

[0115] Transportation time: Normalized by route segment. For example, the time from the cache to the spindle threading machine is mapped to a scale of 0-1 (baseline value 2 minutes = 0.2, congestion 3 minutes = 0.3, maximum threshold 10 minutes = 1.0). The time from the spindle threading machine to the laminating machine is mapped to a scale of 0-1 (baseline value 4.5 minutes = 0.45, increasing to 6 minutes in congestion = 0.6).

[0116] Task load: Divide the AGV task queue length (e.g., the current number of queued tasks is 6) by the maximum load capacity (20) and normalize it to 0.3.

[0117] The acquisition of cache inventory status refers to the data acquisition and preprocessing module acquiring real-time inventory information, raw material consumption history data, and insufficient inventory alarm information from the cache system;

[0118] Release paper: The current inventory is 30 rolls (upper limit Imax = 300 rolls), the historical consumption rate is 3 rolls / hour, and when the inventory is less than or equal to 7 rolls, a replenishment instruction is triggered to the intelligent decision-making system.

[0119] PE film: The current inventory is 30 rolls (upper limit Imax = 300 rolls), the historical consumption rate is 12 rolls / hour, and when the inventory is less than or equal to 14 rolls, a replenishment instruction is triggered to the intelligent decision-making system.

[0120] Thin paper rolls: Scan the pallet code when entering the warehouse and bind the material information (specification 1200mm, quantity 50 rolls). The real-time inventory is 30 rolls.

[0121] Inventory normalization: Release paper inventory is mapped from 0 to 300 rolls to 0-1 (30 rolls is mapped to 0.1).

[0122] Alarm signal coding: insufficient inventory alarm is marked as 1, normal is 0, and a binary sequence is generated according to the time window.

[0123] Using a 10-minute time window, we time-aligned the equipment, logistics, and other data to generate a 24-hour × 144-time-step sequence. An example of a standardized dataset for film production (normalized) is shown in the following table:

[0124]

[0125] Step 2. Build a dynamic demand forecasting model and use the LSTM neural network to predict the demand for raw materials and auxiliary materials in the future period;

[0126] (1) The dynamic demand forecasting model is constructed by building an LSTM neural network model through a dynamic demand forecasting module:

[0127] The LSTM neural network structure is a bidirectional two-layer LSTM unit. The gating mechanism of each LSTM unit is a forget gate, the input gate uses a sigmoid activation function, and the candidate state uses a tanh activation function. The input layer dimension is consistent with the number of features, and the time step window is 24 hours.

[0128] The bidirectional LSTM structure has two layers in the forward and backward directions, with 64 neurons in the hidden layer. The outputs are integrated by splicing to preserve the bidirectional temporal characteristics.

[0129] The output of the LSTM layer is followed by a fully connected layer with a ReLU activation function to extract high-order temporal features. A Dropout layer is added after the fully connected layer to prevent overfitting, and L2 regularization is used on the weights to optimize the generalization ability of the model.

[0130] The output layer uses a linear activation function to directly predict material demand for the next 6 hours. The loss function of the training strategy is the weighted mean square error (W-MSE), with weights decaying over time.

[0131] Optimally, if the RMSE of the baseline model is still much larger than δ×RMSEbase after multiple rounds of optimization, the number of hidden layers can be increased to 3, while strengthening regularization and improving the Dropout probability; δ is the adjustment coefficient, and its value is greater than or equal to 1. RMSEbase is the benchmark error in the model verification stage.

[0132] (2) Model training and verification

[0133] The standardized dataset, time series data on historical raw material consumption, and inventory fluctuation warning information were input into the LSTM neural network model of the dynamic demand forecasting module. The Adam optimizer was used to dynamically adjust the learning rate and the weighted mean square error (W-MSE) was used as the loss function. The Adam input was a 24-hour time series window, and the output was the demand forecast for the next 6 hours. The training cycle was 400 times.

[0134] The standard data is divided into: 70% as training set, 15% as validation set, and 15% as test set;

[0135] The root mean square error (RMSE) or mean absolute percentage error (MAPE) is used as the model accuracy metric. The RMSE and MAPE of the predicted value and the true value are calculated on the test set to verify the model effectiveness. In particular, when RMSE>δ×RMSEbase or MAPE is greater than the set threshold, the model retraining is triggered.

[0136] Step 3. Construction and solution of multi-objective optimization model.

[0137] The mathematical model is solved based on the improved NSGA-II algorithm. The optimization objectives are to minimize the total cost, the lowest out-of-stock rate and the smallest inventory fluctuation. The Pareto optimal solution set and the replenishment quantity Q are output. i and replenishment cycle T i Replenishment strategy;

[0138] The multi-objective optimization model constructs a mathematical model of multi-objective optimization through a multi-objective optimization module. The constructed objective function 1 is to minimize the total cost of replenishment cost, inventory holding cost and out-of-stock penalty cost:

[0139]

[0140] Among them, the fixed costs generated by a single replenishment operation cover AGV scheduling, equipment operation, and manual inspection;

[0141] C order,i =C AGV,i +C 设备,i +C 人工,i

[0142] Here, C AGV,i is the AGV scheduling cost; C 设备,i Energy consumption cost for operating equipment such as axle threading machines, hoists, and trusses; C 人工,i Cost of manual inspection and confirmation;

[0143] Furthermore, the storage costs incurred by materials during inventory are positively correlated with storage time and occupied space;

[0144] C hold,i =α i ·V i ·T hold

[0145] Here, α i V is the daily storage cost per unit volume; i is the volume of a single piece of material; T hold is the average storage time;

[0146] Out-of-stock results in direct economic losses, including downtime losses and emergency restocking premiums, so the out-of-stock penalty costs are:

[0147] C short,i =β i ·S i +γ i ·b(S i >0)

[0148] β i is the downtime loss per unit of stock-out; γ i b(S i >0) is the indicator function (1 when out of stock, 0 otherwise);

[0149] Here, i: the total number of material types, C order,i : Fixed cost of replenishing the i-th type of material once; D i : The predicted demand rate of the i-th type of material, output by the LSTM model; Q i : Replenishment quantity of the i-th category material; C hold,i : Unit inventory holding cost of the i-th type of material; T i : Replenishment cycle of the i-th category material; C short,i: Out-of-stock penalty cost of the i-th material; Ii0: Initial inventory of the i-th material; D i ·T i :Total demand during the replenishment cycle; max(0,D i ·T i -(I i0 +Q i )): out-of-stock quantity;

[0150] The constructed objective function 2 is to minimize the risk of out-of-stock:

[0151]

[0152] The constructed objective function 3 is to minimize the inventory volatility:

[0153]

[0154] Here, σ(I i (t) is the standard deviation of the inventory level of the i-th material; u i is the mean of the inventory level of the i-th material; I i (t) is the inventory level function of the i-th type of material at time t; T i is the replenishment cycle of the i-th material; n is the total number of materials. The smaller the Ti / Pi ratio is, the closer the replenishment cycle is to the equipment loading and unloading capacity, and the lower the inventory fluctuation risk is.

[0155] in,

[0156]

[0157] Here, u i It can actually be expressed as the average inventory of the i-th category material;

[0158] The mathematical model, its prerequisites (constraints) include:

[0159] (1) Replenishment and cycle non-negative constraints:

[0160] Q i ≥0,T i ≥0,I i0 ≥0(i=1,2,…,n)

[0161] (2) Cache storage capacity limit:

[0162] 0≤I i (t)+Q i ≤S i

[0163] Where, I i (t) represents the inventory of material i at time t, S iIndicates the upper limit of the storage capacity of material i;

[0164] (3) Equipment loading and unloading rhythm matching:

[0165]

[0166] T i represents the replenishment cycle of material i; P i represents the loading and unloading cycle of the production equipment associated with material i; k i is the loading and unloading efficiency coefficient of material i, k ii ≥1;

[0167] This constraint limits the replenishment cycle T i The lower limit of , directly affects the minimum total cost minf1 of the objective function and the minimum inventory volatility minf3:

[0168] Replenishment cycle T i An increase in will reduce the replenishment frequency but increase the inventory holding cost.

[0169] Equipment loading and unloading rhythm matching constraint mandatory T i Not less than the equipment loading and unloading capacity, to avoid loading and unloading conflicts caused by frequent replenishment (such as AGV transportation congestion), thereby indirectly controlling costs.

[0170] T i / P i The smaller the ratio, the closer the replenishment cycle is to the equipment loading and unloading capacity, and the lower the risk of inventory fluctuations.

[0171] The constraint condition is to limit T i ≥P i / k i , ensure that T i / P i ≥1 / k i , thereby avoiding drastic inventory fluctuations caused by too short a replenishment cycle (insufficient loading and unloading capacity).

[0172] The significance of this constraint: It quantifies the actual capacity of the equipment into a mathematical model, breaking through traditional static constraints (such as fixed replenishment cycles). By binding the replenishment cycle to loading and unloading capacity, it avoids equipment-inventory disconnects and ensures that the replenishment strategy neither exceeds the equipment's processing limits (leading to task backlogs) nor is it overly conservative (leading to inventory redundancy). In the optimization of three objectives: cost, stockouts, and fluctuations, the equipment loading and unloading cycle constraint allows for a precise match between replenishment strategies and production capacity, making it suitable for highly dynamic manufacturing environments (such as prepreg production).

[0173] (4) Dynamic inventory constraints:

[0174] Q i ≥D i ·Ti -I i0

[0175] D i : The predicted demand rate of the i-th material, output by LSTM;

[0176] (5) AGV transport capacity limitations:

[0177]

[0178] t k represents the time taken for the kth AGV transport task; T AGV,max Indicates the maximum continuous working time of AGV;

[0179] The improved NSGA-II algorithm introduces adaptive crossover probability:

[0180]

[0181] Among them, P max is the initial maximum crossover probability, P min is the minimum crossover probability, Gen is the current evolutionary generation, Gen max is the maximum evolutionary generation.

[0182] The improved NSGA-II algorithm is used to solve the mathematical model. The algorithm optimization process is as follows: Figure 2 As shown, the specific process is:

[0183] (1) Initialize the population.

[0184] Generate the initial parent population according to the variable range of the material replenishment quantity Qi and the replenishment cycle Ti. The chromosome code represents the replenishment quantity and replenishment cycle of each material. The population size is N and the maximum generation number is Gen_max.

[0185] The chromosome encoding is [Q1,T1,Q2,T2,...,Q n ,T n ], corresponding to the replenishment quantity and cycle of n types of materials;

[0186] Taking film as an example, the chromosome encodes 8 variables, corresponding to the replenishment quantity and replenishment cycle of 4 types of materials: chromosome = [Q release paper, T release paper, QPE film, TPE film, Q thin paper tube, T thin paper tube, Q resin, T resin]

[0187] (2) Genetic manipulation.

[0188] High-quality individuals are selected from the parent population using a tournament selection strategy. The selection strategy prioritizes individuals with higher non-dominated sorting levels. If the levels are the same, the one with greater crowding is selected. Simulated binary crossover (SBX) is used to generate offspring individuals, and a polynomial mutation strategy is used to perturb individual gene positions.

[0189] (3) Non-dominated sorting and elite selection

[0190] Merge the parent and child populations (size 2N) and divide the hierarchy (F1, F2, ... F n ); Calculate the crowding degree of individuals at the same level, and select the top N individuals according to the level priority and crowding degree to form the new parent generation;

[0191] (4) Dynamic feedback mechanism

[0192] Adjust the out-of-stock calculation in the objective function based on the updated demand forecast data Di, and dynamically modify the transportation cost weight based on the AGV status data;

[0193] (5) Termination conditions and decision output

[0194] When the number of iterations reaches the maximum evolutionary generation Gen_max, the three-dimensional Pareto optimal solution set is output, and the solution that satisfies the shortage quantity S is selected from the Pareto solution set. i = 0 and the replenishment strategy with the lowest total cost, output the replenishment quantity Q i and replenishment cycle T i ;

[0195] The Pareto solution set screening rule is to give priority to the solution where all material shortages Si=0; if there are multiple zero shortage solutions, the solution with the lowest total cost and the smallest inventory fluctuation coefficient is selected;

[0196] The final output is a replenishment strategy with zero out-of-stock, low volatility and optimal cost;

[0197] NSGA-II multi-objective optimization results (Gen_max=200)

[0198]

[0199] Step 4. Execute the replenishment strategy. The production management system sends replenishment instructions to the cache system, and the AGV is dispatched to perform the transportation task.

[0200] The replenishment strategy execution module receives the replenishment strategy of replenishment cycle Ti and replenishment quantity Qi generated by the multi-objective optimization module and transmits it to the production management system. The production management system analyzes the optimization results and converts them into executable production replenishment instructions, including key information such as material type, quantity, target storage location and latest arrival time. The task priority is sorted by the urgency of equipment loading and unloading;

[0201] The production management system sends replenishment instructions to the cache system. The cache system dynamically assigns storage location priorities based on equipment location, uses the proximity principle and emergency task queue-jumping mechanism to schedule AGVs to perform transportation tasks, and allocates tasks to idle AGVs through a load balancing strategy to ensure that the transportation response time meets the production rhythm requirements.

[0202] Acquiring real-time equipment data refers to the production management system acquiring it through the OPC-UA protocol; integrating the MQTT protocol to receive AGV transport logs, including path coordinates, time consumption, congestion coefficient and other data;

[0203] Step 5. Inventory fluctuation warning and dynamic adjustment. Monitor inventory status in real time and trigger strategy adjustments or model retraining based on fluctuation coefficients.

[0204] The inventory fluctuation warning and dynamic adjustment mechanism refers to real-time monitoring of inventory status through the inventory monitoring module and dynamic adjustment through the feedback and adaptive module. If the inventory is sufficient, replenishment is not triggered. Otherwise, replenishment instructions are generated according to the replenishment strategy in step 3 and transmitted to the production management system to execute the replenishment operation.

[0205] The dynamic adjustment mechanism includes adjusting the objective function weight according to the inventory fluctuation coefficient. The fluctuation coefficient calculation formula is:

[0206]

[0207] I i represents the daily inventory of material i, u(I i ) represents the average inventory level of material i, σ(I i ) represents the standard deviation of the inventory level of material i.

[0208] The trigger condition for model retraining is that the prediction error MAPE>α for three consecutive times, α is the set percentage error; or RMSE>δ×RMSEbase, δ is the adjustment coefficient, its value is greater than or equal to 1, and RMSEbase is the benchmark error in the model verification stage.

[0209] The above embodiments illustrate and describe the basic principles and main features of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention, and such changes and modifications fall within the scope of the invention as claimed.

Claims

1. A multi-objective inventory optimization method based on dynamic demand forecasting, characterized in that: include: S1. Collect cache inventory status, production order data, and AGV transportation capacity data, and standardize the obtained data, including generating a standardized data set D norm ; S2. Build a dynamic demand forecasting model based on the LSTM neural network to predict the demand for raw materials and auxiliary materials in the future; The dynamic demand forecasting model includes the following objective function: Objective function 1: The total cost of replenishment cost, inventory holding cost and out-of-stock penalty cost is minimized, which can be expressed mathematically as: Objective function 2: Minimize out-of-stock risk. Its mathematical expression is: Objective function 3: Minimize inventory volatility. Its mathematical expression is: Where C order,i represents the replenishment cost of material i, which includes AGV scheduling, equipment energy consumption, and manual inspection costs. i represents the forecast demand rate of material i, Q i represents the replenishment quantity of material i, C hold,i represents the inventory holding cost of material i, T i represents the replenishment cycle of material i, C short,i represents the out-of-stock penalty cost of material i, I i0 represents the initial inventory of material i; σ(I i (t)) is the standard deviation of the inventory level of the i-th material; u(I i (t)) is the mean inventory level of the i-th material; I i (t) is the inventory level function of the i-th type of material at time t; n is the total number of materials. The smaller the Ti / Pi ratio is, the closer the replenishment cycle is to the equipment loading and unloading capacity, and the lower the inventory fluctuation risk is. S3. Solve the dynamic demand forecasting model described in step S2 based on the improved NSGA-II algorithm. The optimization objectives are to minimize total cost, minimize out-of-stock rate, and minimize inventory fluctuation. Output the Pareto optimal solution set to determine the replenishment quantity Q. i and replenishment cycle T i Replenishment strategy; The following constraints are considered during the solution process: Replenishment and cycle non-negative constraints: Q i ≥0,T i ≥0,I i0 ≥0(i=1,2,…,n) Cache library location capacity limit: 0≤I i (t)+Q i ≤S i Where, I i (t) represents the inventory of material i at time t, S i Indicates the upper limit of the storage capacity of material i; Equipment loading and unloading rhythm matching: T i represents the replenishment cycle of material i; P i represents the loading and unloading cycle of the production equipment associated with material i; k i is the loading and unloading efficiency coefficient of material i, k ii ≥1; Inventory dynamic constraints: Q i ≥D i ·T i -I i0 AGV transport capacity limitations: t k represents the time taken for the kth AGV transport task; T AGV,max Indicates the maximum continuous working time of AGV; S4. Based on the solution of step S3, the production management system sends a replenishment instruction to the cache system and schedules the AGV to perform the transportation task; S5. Inventory fluctuation warning and dynamic adjustment: Real-time monitoring of inventory status, triggering strategy adjustment or retraining of dynamic demand forecasting models based on the fluctuation coefficient, including adjusting the objective function weight based on the inventory fluctuation coefficient. The fluctuation coefficient calculation formula is: I i represents the daily inventory of material i, u(I i ) represents the average inventory level of material i, σ(I i ) represents the standard deviation of the inventory level of material i.

2. The multi-objective inventory optimization method based on dynamic demand forecasting according to claim 1 is characterized in that: The LSTM neural network structure of the dynamic demand forecasting model includes: Bidirectional two-layer LSTM unit, each LSTM unit's gating mechanism is a forget gate, the input gate uses a sigmoid activation function, and the candidate state uses a tanh activation function; the input layer dimension is consistent with the number of features, and the time step window is 24 hours; The output of the LSTM layer is followed by a fully connected layer with ReLU as the activation function to extract high-order temporal features. A Dropout layer is added after the fully connected layer to prevent overfitting, and L2 regularization is used on the weights to optimize the generalization ability of the model. The output layer uses the linear activation function Linear to predict material demand in the next 6 hours; The loss function of the training strategy is the weighted mean square error W-MSE, the weight decays over time, and the Adam optimizer dynamically adjusts the learning rate.

3. The multi-objective inventory optimization method based on dynamic demand forecasting according to claim 1 or 2, characterized in that: The dynamic demand forecasting model is trained using cache inventory status data, which includes real-time inventory information, historical raw material consumption data, and insufficient inventory alarm information, and is used to train the dynamic demand forecasting model. The training process uses the root mean square error (RMSE) or the mean absolute percentage error (MAPE) as the model accuracy judgment indicator. When RMSE>δ×RMSEbase or MAPE is greater than the set threshold, the model retraining is triggered. δ is the adjustment coefficient, and its value is greater than or equal to 1. RMSEbase is the benchmark error in the model verification stage; the final output is the forecast value of raw material and auxiliary material demand from t to t+6 hours in the future.

4. The multi-objective inventory optimization method based on dynamic demand forecasting according to claim 1, characterized in that: The storage costs incurred by materials during inventory are positively correlated with the storage time and space occupied, which can be expressed mathematically as: C hold,i =α i ·V i ·T hold α i V is the daily storage cost per unit volume; i is the volume of a single piece of material; T hold is the average storage time; The direct economic loss caused by out-of-stock includes downtime loss and emergency restocking premium, so the out-of-stock penalty cost is mathematically expressed as: C short,i =b i ·S i +g i ·b(S i >0) β i is the downtime loss per unit of stock-out; γ i b(S i >0) is an indicator function, which takes 1 when out of stock and 0 otherwise; The standard deviation of the inventory level of material i σ(I i (t)) is defined as follows: u i is the average inventory of the i-th category material.

5. The multi-objective inventory optimization method based on dynamic demand forecasting according to claim 1 is characterized in that: The improved NSGA-II algorithm introduces adaptive crossover probability: Among them, P max is the initial maximum crossover probability, P min is the minimum crossover probability, Gen is the current evolutionary generation, Gen max is the maximum evolutionary generation.

6. An intelligent decision-making system based on multi-objective inventory management, characterized in that: The system executes the method described in claims 1-5 to perform dynamic inventory demand forecasting and inventory management.

7. The intelligent decision-making system according to claim 6, characterized in that: The system includes: Inventory dynamic monitoring module, used to obtain material inventory information in real time, including material consumption and equipment energy consumption; Dynamic demand forecasting module, used to execute the dynamic demand forecasting model and output the forecast value of raw material and auxiliary material demand in a certain time period in the future; The multi-objective optimization module obtains the predicted demand data transmitted by the dynamic demand forecast module and the AGV status data of the AGV scheduling system. Based on the NSGA-II algorithm, it screens the new population through non-dominated sorting stratification, genetic evolution, and elite strategy selection, and outputs the Pareto optimal solution set and the replenishment strategy of replenishment quantity and replenishment cycle; The replenishment strategy execution module receives the replenishment strategy generated by the multi-objective optimization module and transmits it to the production management system to convert it into production replenishment instructions, including material type, quantity, target storage location and latest arrival time; The feedback and adaptive module does not trigger replenishment if the inventory is sufficient. Otherwise, it generates a replenishment instruction according to the replenishment strategy in step S3 and transmits it to the production management system to execute the replenishment operation.

8. The intelligent decision-making system according to claim 6 or 7, characterized in that: The system is suitable for production management of films and prepregs. The production process takes into account production order data and AGV transportation capabilities, and combines material supply to realize dynamic demand decision-making and forecasting of material inventory.

Images