Material processing method, device, equipment, storage medium and program

Abstract

The embodiment of the invention provides a material processing method and device, equipment, a storage medium and a program. The method comprises the following steps: acquiring demand driving data of a target material and market environment data corresponding to the target material; based on a deep learning algorithm, determining a sales prediction result of the target material in a set time period according to the demand driving data, and determining market demand distribution of the target material in the set time period according to the market environment data; determining a production strategy according to the sales prediction result and the market demand distribution, and performing production operation on the target material based on the production strategy; and based on the actual market information in the production operation process, the production strategy is adjusted according to a reinforcement learning algorithm and a multi-objective optimization algorithm, and production operation is performed on the target material based on the adjusted production strategy. According to the scheme, market demand changes and emergencies can be dynamically handled, and efficient processing of the material data is realized.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and particularly to a method, device, equipment, storage medium and program for material processing. Background Art

[0002] Material Requirements Planning (MRP) is a modern management tool based on production planning, inventory management and supply chain collaboration. Its core goal is to optimize procurement, production and inventory processes by accurately calculating material requirements, so as to minimize costs and ensure production continuity. Especially in the meat processing industry, the application of MRP systems has unique complexities and challenges. Specifically, meat processing mainly uses fresh raw materials (such as live livestock, poultry, frozen meat, etc.), and its perishability, short shelf life, large seasonal supply fluctuations and other characteristics require extremely high timeliness and flexibility of material planning. At the same time, there are various product types (such as cut meat, deep-processed products, cooked food, etc.), and the collaborative management of multiple types of materials such as raw materials, auxiliary materials (seasonings, additives), and packaging materials (vacuum bags, labels) needs to be covered.

[0003] The MRP systems used in the prior art mainly rely on historical data and fixed rules, and cannot dynamically respond to market demand changes and emergencies. Moreover, a large number of manual operations and inputs are required in the actual application process, with frequent manual intervention and low processing efficiency of material data. Summary of the Invention

[0004] Multiple aspects of this application provide a method, device, equipment, storage medium and program for material processing, which can dynamically respond to market demand changes and emergencies, and the entire process does not require manual participation, and can achieve efficient processing of material data.

[0005] In the first aspect, an embodiment of this application provides a method for material processing, including:

[0006] Obtain the demand-driven data of the target material and the market environment data corresponding to the target material. The demand-driven data is used to reflect various parameters affecting the demand of the target material, and the market environment data is used to reflect the market change situation corresponding to the target material;

[0007] Based on the deep learning algorithm, determine the sales volume prediction result of the target material within the set time period according to the demand-driven data, and determine the market demand distribution of the target material within the set time period according to the market environment data;

[0008] Determine the production strategy according to the sales volume prediction result and the market demand distribution, and perform production operations on the target material based on the production strategy;

[0009] Based on the actual market information in the production operation process, adjust the production strategy according to the reinforcement learning algorithm and the multi-objective optimization algorithm, so as to perform production operations on the target material based on the adjusted production strategy.

[0010] In a second aspect, an embodiment of the present application provides a material processing device, including:

[0011] An acquisition module, configured to acquire demand-driven data of a target material and market environment data corresponding to the target material, where the demand-driven data is used to reflect various parameters affecting the demand of the target material, and the market environment data is used to reflect the market change situation corresponding to the target material;

[0012] A prediction module, configured to determine a sales volume prediction result of the target material within a set time period according to the demand-driven data based on a deep learning algorithm, and determine the market demand distribution of the target material within the set time period according to the market environment data;

[0013] A production module, configured to determine a production strategy according to the sales volume prediction result and the market demand distribution, and perform production operations on the target material based on the production strategy;

[0014] An adjustment module, configured to adjust the production strategy according to the reinforcement learning algorithm and the multi-objective optimization algorithm based on the actual market information in the production operation process, so as to perform production operations on the target material based on the adjusted production strategy.

[0015] In a third aspect, an embodiment of the present application provides an electronic device, including: a memory, a processor, and a communication interface; wherein, an executable code is stored on the memory, and when the executable code is executed by the processor, the processor is caused to execute the material processing method as described in the first aspect.

[0016] In a fourth aspect, an embodiment of the present application provides a non-transitory machine-readable storage medium, on which an executable code is stored, and when the executable code is executed by a processor of an electronic device, the processor is caused to execute the material processing method as described in the first aspect.

[0017] In a fifth aspect, an embodiment of the present application provides a computer program product, including: a computer program, and when the computer program is executed by a processor of an electronic device, the processor is caused to execute the material processing method as described in the first aspect.

[0018] In the embodiments of the present application, by using deep learning algorithms, the sales volume prediction result of the target material within a set time period can be determined according to demand-driven data, and the market demand distribution of the target material within the set time period can be determined according to market environment data. There is no need to rely on historical data and fixed rules anymore, and it can dynamically respond to market demand changes and emergencies, thereby better completing the production operation of the target material. During this process, based on the actual market information in the production operation process, the production strategy is adjusted according to the reinforcement learning algorithm and the multi-objective optimization algorithm, and the production operation of the target material is carried out based on the adjusted production strategy. It can not only adjust the production strategy in real time according to the actual market information, but also consider multiple objectives simultaneously during the production process to give the optimal production strategy. The entire process does not require manual intervention, realizing the efficient processing of material data. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation to the present application. In the drawings:

[0020] Figure 1 is a flowchart of a material processing method provided by an exemplary embodiment of the present application;

[0021] Figure 2 is a schematic structural diagram of a material requirements planning system provided by an exemplary embodiment of the present application;

[0022] Figure 3 is a schematic structural diagram of a material processing device provided by an exemplary embodiment of the present application;

[0023] Figure 4 is a schematic structural diagram of an electronic device provided by an exemplary embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0024] To make the objectives, technical solutions, and advantages of the present application clearer, the technical solutions of the present application will be clearly and completely described below in conjunction with the specific embodiments of the present application and the corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0025] Material Requirements Planning (MRP) is a modern management tool based on production planning, inventory management, and supply chain collaboration. Its core objective is to optimize procurement, production, and inventory processes by accurately calculating material requirements, minimizing costs, and ensuring production continuity. Especially in the meat processing industry, the application of MRP systems has unique complexities and challenges. Specifically, meat processing mainly uses fresh raw materials (such as live livestock, poultry, frozen meat, etc.), and its perishable nature, short shelf life, and large seasonal supply fluctuations pose extremely high requirements for the timeliness and flexibility of material planning. At the same time, there is a wide variety of products (such as cut meat, deep-processed products, cooked food, etc.), and the collaborative management of multiple types of materials, including raw materials, auxiliary materials (seasonings, additives), and packaging materials (vacuum bags, labels), needs to be covered.

[0026] The MRP systems used in the prior art mainly rely on historical data and fixed rules, unable to dynamically respond to market demand changes and emergencies. Moreover, in the actual application process, a large number of manual operations and inputs are required, with frequent human intervention, and the decision-making process is easily affected subjectively, unable to achieve efficient and accurate processing of material data, thus affecting normal production operations. In view of this, the embodiments of this application provide a material processing method.

[0027] Figure 1 The flowchart of a material processing method provided by the embodiments of this application is as Figure 1 shown, and this method includes:

[0028] Step 101, obtain the demand-driven data of the target material and the market environment data corresponding to the target material. The demand-driven data is used to reflect various parameters affecting the demand of the target material, and the market environment data is used to reflect the market change situation corresponding to the target material.

[0029] Step 102, based on the deep learning algorithm, determine the sales volume prediction result of the target material within the set time period according to the demand-driven data, and determine the market demand distribution of the target material within the set time period according to the market environment data.

[0030] Step 103, determine the production strategy according to the sales volume prediction result and the market demand distribution, and perform production operations on the target material based on the production strategy.

[0031] Step 104, based on the actual market information during the production operation process, adjust the production strategy according to the reinforcement learning algorithm and the multi-objective optimization algorithm, and perform production operations on the target material based on the adjusted production strategy.

[0032] In practical applications, it is possible to first obtain the demand-driven data of the target materials (such as products that can be sold in the market, like pork, beef, sausages, corn, etc.), as well as the market environment data corresponding to the target materials. Among them, the demand-driven data is used to reflect various parameters affecting the demand for the target materials. For example, historical sales data, market reports, weather, holiday information, promotional activities, etc. The market environment data is used to reflect the market changes corresponding to the target materials. For example, market research data, consumer purchase preferences, macroeconomic indicators, etc.

[0033] After that, based on the deep learning algorithm, determine the sales volume prediction result of the target material within the set time period according to the demand-driven data, and determine the market demand distribution of the target material within the set time period according to the market environment data. The specific process is as follows: Input the demand-driven data into the trained sales volume prediction model so that the sales volume prediction model outputs the sales volume prediction result of the target material within the set time period; Input the market environment data into the trained market demand prediction model so that the market demand prediction model outputs the market demand prediction result of the target material within the set time period. Among them, the sales volume prediction model is a Long Short-Term Memory (LSTM) model, and the market demand prediction model is a Variational Autoencoder (VAE).

[0034] After obtaining the sales volume prediction result and the market demand distribution, the production strategy can be determined, and production operations can be carried out on the target materials based on the production strategy. During this process, actual market information can be obtained in real time, such as the sudden shutdown of the product supplier on which the production of the target material depends, or the holiday approaching, and the demand for the target material in the market surges. Based on the actual market information during the production operation, adjust the production strategy according to the reinforcement learning algorithm and the multi-objective optimization algorithm, and carry out production operations on the target materials based on the adjusted production strategy. Among them, the reinforcement learning algorithm includes the Deep Q-Network algorithm and the Policy Gradient algorithm. The multi-objective optimization algorithm includes the Genetic algorithm and the Ant Colony algorithm.

[0035] Based on the above, the material processing method provided by the embodiments of the present application can, by using deep learning algorithms, determine the sales volume prediction result of the target material within a set time period according to demand-driven data, and determine the market demand distribution of the target material within the set time period according to market environment data. Without relying on historical data and fixed rules, it can dynamically respond to market demand changes and emergencies, and thus better complete the production operation of the target material. In this process, based on the actual market information in the production operation process, the production strategy is adjusted according to the reinforcement learning algorithm and the multi-objective optimization algorithm, and the production operation of the target material is carried out based on the adjusted production strategy. It can not only adjust the production strategy in real time according to the actual market information, but also consider multiple objectives simultaneously during the production process to give the optimal production strategy. The whole process does not require manual intervention, realizing the efficient and accurate processing of material data and ensuring the smooth progress of the production operation of the target material.

[0036] In the embodiments of the present application, after carrying out the production operation of the target material based on the adjusted production strategy, the method further includes: adjusting the parameters of the deep learning algorithm and the reinforcement learning algorithm according to the actual sales volume and the actual market demand distribution of the target material.

[0037] In practical applications, there is likely to be an error between the sales volume prediction result and the market demand distribution of the target material obtained by the deep learning algorithm and the actual sales volume and the actual market demand distribution of the target material. In this case, in order to facilitate the efficient and accurate generation of subsequent operations, the parameters of the deep learning algorithm and the reinforcement learning algorithm can be adjusted.

[0038] For the sake of easy understanding, for example, assume that the sales volume prediction result and the market demand of the target material are 1200 tons, while the actual sales volume is 350 tons (lower than the prediction), and the remaining inventory is 400 tons (200 tons of which are approaching expiration). Then at this time, the weights of holidays can be reduced and the influence of economic indicators can be increased through the LSTM model, and the penalty coefficient for inventory backlog can be increased through DQN to update the production strategy: reduce production to 15 tons per day and start the "buy one get one free" promotion for clearance.

[0039] The following specifically introduces the specific application processes of the above algorithms in the embodiments of the present application

[0040] Example:

[0041] I. Deep learning algorithm:

[0042] Taking the sales volume prediction model as LSTM as an example, LSTM is a classic deep learning model for processing time series data, which can capture long-term dependencies and is suitable for analyzing sales historical data and seasonal trends.

[0043] In practical applications, by inputting demand-driven data into the trained LSTM, the LSTM can output the sales prediction results of the target material within the set time period. For the sake of easy understanding, specific examples are given from the following aspects:

[0044] 1. Input data: Integration of multi-dimensional time series data

[0045] The input data of the LSTM needs to contain multi-dimensional time series information related to sales. The following takes the actual scenario of a meat processing enterprise as an example:

[0046] Suppose the demand-driven data includes: historical sales data, market reports and competition dynamics, weather data, holidays and promotional activities. Specifically:

[0047] The historical sales data is: the monthly sales volume of beef sausages of an enterprise in the past 3 years (for example, 10 tons were sold in January 2021, and 25 tons were sold in December 2023). The role of this data is: to capture the long-term trend (such as annual growth rate) and periodic fluctuations (such as a sharp increase in sales at the end of the year) of the beef sausage sales volume.

[0048] Market reports and competition dynamics: Industry reports show that the demand for healthy foods increased by 20% in 2023, and the launch of low-fat beef sausages by competitors led to a 5% decline in the sales volume of this enterprise. The role of this data is: to correct the prediction model through external market information and avoid the limitations of relying solely on historical data.

[0049] Weather data: High summer temperatures (such as continuous temperatures above 30°C) led to a 30% increase in the sales volume of barbecue meat products, while the demand for hot pot meat products increased in winter. The role of this data is: to correlate weather with category demand and predict seasonal fluctuations.

[0050] Holidays and promotional activities: The sales volume of beef sausages increased by 50% in the month before the Spring Festival compared with normal days, and the online sales volume increased by 80% during the Double Eleven promotion. The role of this data is: to quantify the impact of holidays and promotions on sales volume and generate accurate peak predictions.

[0051] Specifically in implementation, for example, assume that the input data is a 5-dimensional time series (historical sales data, market reports and competition dynamics, weather data, holidays and promotional activities), and the data form at each time step (such as monthly) is as follows: [Sales volume = 15 tons, Temperature = 28°C, Holiday = 1 (Spring Festival), Promotion = 0.8 (discount activity), Competitor = 0 (no new products)]. The LSTM captures complex dependencies through multi-dimensional features.

[0052] 2. Network architecture: Hierarchical functions and information flow

[0053] Taking the LSTM network architecture of an enterprise as an example to illustrate the functions of each layer:

[0054] Input layer: Receives the standardized 5-dimensional time series data (such as monthly data for the past 24 months). Example: The input shape is (24, 5), representing 24 time steps, with 5 features at each time step.

[0055] LSTM cell layer: Configures an LSTM layer with 128 hidden units to memorize long-term dependencies through the gating mechanism (input gate, forget gate, output gate). Among them, the forget gate is used to: automatically ignore invalid features (such as extremely low sales volume caused by equipment failures in a certain month). The input gate is used to: focus on learning the relationship between holiday promotions and sudden increases in sales volume. The output gate is used to: generate the hidden state of the current time step by integrating historical information.

[0056] Fully connected layer: Maps the 128-dimensional features output by the LSTM to a 1-dimensional predicted value (sales volume for the next month). Example: The weight matrix of the fully connected layer combines features as predicted sales volume = 0.3 × historical trend + 0.5 × promotion intensity + 0.2 × weather impact.

[0057] Output layer: Outputs the predicted sales volume for the next N months (such as predicted values for the next 3 months: 18 tons, 20 tons, 22 tons).

[0058] 3. Training process: Model optimization and error correction

[0059] Training data example: Uses historical data from 2019 - 2022 as the training set and data from 2023 as the validation set.

[0060] Loss function (mean squared error, MSE): Assume the model predicts the sales volume in January 2023 to be 20 tons, and the actual sales volume is 18 tons. Then the single-point error is (20 - 18) 2 = 4. The training objective is to minimize the average error over all time steps.

[0061] Backpropagation and weight adjustment: Assume the LSTM model underestimates the impact of high temperatures in the summer of 2023 on sales volume (predicts 25 tons, actual 30 tons).

[0062] Adjustment process:

[0063] 1) Calculate the error gradient: The weight of the high-temperature feature needs to be increased.

[0064] 2) Update the weight parameters related to temperature in the LSTM cell through an optimizer (such as Adam).

[0065] 3) Iteratively train until the model can accurately reflect the non-linear relationship between temperature and sales volume (such as when the temperature > 25°C, the sales volume increases by 2% for every 1°C increase).

[0066] To facilitate an intuitive understanding of the advantages of using the LSTM model in this application, a comparison with the traditional MRP system is presented below:

[0067] Traditional MRP system: Based on linear regression, the predicted sales volume for December 2023 was 20 tons. However, due to Christmas promotions, the actual sales volume reached 30 tons, resulting in a shortage of 10 tons in inventory.

[0068] LSTM model: Accurately predicted the impact of promotions, output a predicted value of 28 tons. The enterprise procured raw materials in advance, meeting the demand and reducing losses.

[0069] Generally speaking, through the integration of multi-dimensional time-series data, the memory ability of the gating mechanism, and dynamic weight adjustment, LSTM has achieved in the material requirements planning of the meat processing industry: "Precise prediction: Quantify the impact of non-linear factors such as holidays, promotions, and weather", "Dynamic adaptation: Real-time correction of the model to cope with market fluctuations", "Cost optimization: Reduce inventory shortages / backlogs and reduce raw material waste".

[0070] Taking the market demand prediction model VAE as an example, VAE is a generative model that can be used to extract features from complex market data and then predict future market demand.

[0071] In practical applications, by inputting market environment data into the trained VAE, VAE can output the market demand prediction results of the target material within a set time period. For the sake of easy understanding, specific examples are given below in the following aspects:

[0072] 1. Input data: Integration of multi-source market data

[0073] The input of VAE needs to cover structured data (such as market research data, consumer purchase preferences, and macroeconomic indicators) and unstructured data. The following takes a certain meat processing enterprise as an example to illustrate:

[0074] Market research data: The consumer survey report shows that 60% of consumers in a certain region tend to buy "low-fat beef products", and the annual growth rate of the demand for organically certified meat products is 15%. The role of this data is to quantify consumer preferences and capture the demand trends in the segmented market.

[0075] Consumer purchase preferences: E-commerce platform data shows that the sales volume of beef balls increases by 40% in winter, while the sales volume of instant beef jerky increases by 30% in summer. The role of this data is to analyze the seasonal characteristics of products and consumer behavior patterns.

[0076] Macroeconomic indicators: According to the data, the meat consumer price index (CPI) increased by 8% year-on-year in 2023, and the GDP growth rate slowed down to 4.5%. The role of this data is to correlate the economic environment with the overall market demand fluctuations.

[0077] Unstructured data: Social media comments (such as "The price of beef has recently increased, so I switched to buying chicken"), industry news (such as an outbreak of foot-and-mouth disease in a certain country leading to import restrictions). The role of this data is to extract sentiment and event features through natural language processing (NLP).

[0078] In specific implementation, for example, assume that the input data forms a multi-dimensional feature vector after preprocessing: [Demand for low-fat products = 0.6, Growth rate of organic products = 0.15, Sales volume of winter beef balls = +40%, Meat CPI = 8%, Social media sentiment = -0.3 (negative)]. The VAE compresses these complex data into low-dimensional latent features for generating demand forecasts.

[0079] 2. Network architecture: Encoder, latent space, and decoder

[0080] Taking the VAE architecture of an enterprise as an example to illustrate the functions of each component:

[0081] Encoder:

[0082] Maps the input multi-dimensional data (such as 10-dimensional features) to the latent space (such as 3-dimensional).

[0083] Example: The encoder outputs the latent variable z = [z1 = 1.2, z2 = -0.5, z3 = 0.8], where z1 represents the intensity of consumers' health awareness, z2 represents the economic environment impact factor, and z3 represents the seasonal fluctuation index.

[0084] Latent Space:

[0085] Represents the uncertainty of the latent variable through a probability distribution (such as a Gaussian distribution).

[0086] Example: The latent variable z follows the distribution N(μ,σ), where μ = [1.2, -0.5, 0.8] (mean) and σ = [0.1, 0.3, 0.2] (standard deviation). The VAE generates diverse demand scenarios through sampling.

[0087] Decoder:

[0088] Reconstructs the market demand forecast from the latent variable z.

[0089] Example: The decoder outputs the probability distribution of beef demand for the next 3 months. For example:

[0090] Month 1: N(18 tons, 2 tons)

[0091] Month 2: N(20 tons, 3 tons)

[0092] Month 3: N(22 tons, 4 tons)

[0093] 3. Training process: maximizing likelihood and minimizing reconstruction error

[0094] Training data example: Use market data from 2019-2022 to train the VAE model and verify it with data from 2023.

[0095] Loss function:

[0096] Reconstruction loss: Mean square error (MSE), which measures the gap between the decoder prediction value and the actual demand.

[0097] KL divergence: constrains the distribution of latent variables to be close to the standard normal distribution to prevent overfitting.

[0098] Backpropagation and optimization:

[0099] Scenario: Assume that the VAE model underestimates the drop in demand caused by the economic recession in 2023 (predicting 25 tons, actual 20 tons).

[0100] Adjustment process:

[0101] 1) Calculate the reconstruction loss gradient and adjust the decoder weights to enhance the influence of the economic factor.

[0102] 2) Maintaining generative diversity by penalizing excessive deviation of latent variables through KL divergence.

[0103] 3) After iterative training, the VAE model is able to generate a demand distribution that includes "economic downturn risk" (such as the predicted value is adjusted to N(20 tons, 5 tons)).

[0104] 4. Output: Probability distribution guides decision making

[0105] Forecast example: The VAE model outputs the probability distribution of beef demand in the next three months as follows:

[0106] Month 1: Mean 18 tons, standard deviation 2 tons (68% probability of falling between 16 and 20 tons).

[0107] Month 2: Mean 20 tons, standard deviation 3 tons (95% probability of falling between 14-26 tons).

[0108] Month 3: Mean 22 tons, standard deviation 4 tons (99.7% probability of falling between 10-34 tons).

[0109] Enterprise application scenarios:

[0110] Sales strategy optimization:

[0111] If the demand cap in the third month is 34 tons, the company can sign a flexible supply agreement with the distributor in advance.

[0112] For extreme demands with low probabilities (such as 10 tons), develop promotion plans (such as discount clearance).

[0113] Material Requirements Planning:

[0114] Safety stock: Stock up according to the mean μ + 2σ (stock up 30 tons in the third month).

[0115] Production schedule: Adjust the production line priority according to the probability distribution (such as prioritizing production in months with high-probability demands).

[0116] To facilitate an intuitive understanding of the advantages of using the VAE model in this application, the following takes "a certain enterprise using VAE to predict the demand in the fourth quarter of 2023" as an example to compare with traditional methods:

[0117] Traditional method: The linear regression prediction is 25 tons, but actually only 18 tons were sold due to an economic recession, resulting in a backlog of 7 tons in inventory.

[0118] VAE model prediction: Generate a demand distribution N(20 tons, 5 tons), and the enterprise formulates a plan according to the probability quantile:

[0119] Optimistic scenario (25 tons): Reserve 10% production capacity flexibility.

[0120] Pessimistic scenario (15 tons): Reduce the purchase volume and give priority to consuming the existing inventory.

[0121] The actual demand is 19 tons, and the enterprise avoids backlogs through dynamic adjustment.

[0122] Generally speaking, VAE improves the robustness of demand forecasting in the meat processing industry through the following core advantages: "Complex data modeling: Integrate structured and unstructured data to capture non-linear relationships (such as economic indicators and consumer sentiment)", "Uncertainty quantification: Generate probability distributions to support risk-sensitive decision-making (such as flexible inventory stocking, multi-scenario plans)", "Feature decoupling: Separate key influencing factors in latent variables (such as health awareness, seasonality) to facilitate targeted optimization strategies".

[0123] It can be understood that in deep learning algorithms, LSTM is good at long-term dependencies in time series (such as holiday cycles), while VAE is good at dealing with high-dimensional, multi-source heterogeneous data and quantifying prediction uncertainties. When used in combination, the dual goals of "accurate prediction" and "risk control" can be achieved, promoting the transformation of supply chain management from "experience-driven" to "data-intelligence-driven".

[0124] Based on the above, the embodiments of the present application use deep learning algorithms (such as LSTM and VAE) for sales and market demand forecasting, which can handle complex time series and multi-dimensional data and dynamically capture market trends. The system can not only learn the long-term dependencies in historical data but also adapt to seasonal fluctuations, holiday effects, and sudden market events, thus providing more accurate demand forecasts.

[0125] II. Reinforcement Learning Algorithm

[0126] As an implementation, the reinforcement learning algorithm is the Deep Q-Network (DQN for short). The Deep Q-Network is a classic reinforcement learning algorithm that can handle complex and high-dimensional state spaces and is suitable for optimizing inventory and production decisions.

[0127] The following will illustrate the application of the Deep Q-Network in inventory and production optimization in the meat processing industry from several aspects:

[0128] 1. State Space: A dynamic representation of the system state, which needs to comprehensively reflect the real-time operation state of the enterprise, including the following key variables (the combination of these variables determines the overall situation of current production and inventory):

[0129] Inventory Level: The current inventory quantity (e.g., the inventory of beef sausages is 50 tons).

[0130] Production Progress: The status of the production line (e.g., the current production capacity is 5 tons per day, and the remaining production order is 30 tons).

[0131] Market Demand Forecast: The order volume for the next week (e.g., the predicted demand is 8 tons per day).

[0132] Personnel Availability: The attendance rate of employees (e.g., the current shift is fully staffed with no absences).

[0133] Supplier Status: The supply situation of raw materials (e.g., the beef supplier's delivery is delayed by 2 days).

[0134] Shelf Life Pressure: The quantity of products approaching the shelf life in inventory (e.g., 10 tons of beef sausages need to be sold within 7 days).

[0135] Example Scenario: On a certain day, the system state is:

Inventory = 50 tons, Production Progress = 30 tons, Demand Forecast = 8 tons / day, Personnel = 100%, Supplier = Delayed 2 days, Shelf Life Pressure = 10 tons

[0136] 2. Action Space: It includes decisions such as adjusting the production plan (e.g., accelerating or slowing down the production rhythm), adjusting the inventory level (e.g., increasing or decreasing the procurement batches), and personnel scheduling (e.g., adjusting shifts or reallocating tasks). For ease of understanding, the following examples are provided:

[0137] Adjust production rhythm: Accelerate production (+20% capacity), maintain the current speed, decelerate production (-20% capacity).

[0138] Adjust procurement batches: Increase the procurement quantity (e.g., purchase 1000 kg more beef), decrease the procurement quantity, maintain the original plan.

[0139] Personnel scheduling: Increase shifts (e.g., night shift), decrease shifts, reallocate tasks.

[0140] Promotion for clearance: Launch a promotion campaign for products approaching the expiration date (e.g., a 10% price cut).

[0141] Example actions: In a scenario where there is overstock and supplier delay, possible actions are:

[0142] Action A: Decelerate production (-20% capacity), decrease the procurement quantity.

[0143] Action B: Maintain the production speed, launch a promotion for clearance.

[0144] Action C: Accelerate production (stock up to meet future demand), increase the procurement batches.

[0145] 3. Reward function: The reward function defines the payoff for each decision, with the goal of maximizing long-term payoff. Specific rewards may include reducing inventory holding costs, minimizing material waste, improving production efficiency, meeting market demand, etc. Among the above rewards, the reward items can include:

[0146] Inventory holding cost: Inventory quantity × unit storage cost (e.g., the cost per ton per day of inventory is 100 yuan).

[0147] Material waste cost: Quantity of expired products × unit loss (e.g., the loss per ton of expired beef sausage is 5000 yuan).

[0148] Production efficiency reward: Degree of matching between actual output and planned output (e.g., a 50-yuan reward for every 1% increase in the matching degree).

[0149] Market demand satisfaction: Actual delivery quantity ÷ demand quantity × reward coefficient (e.g., a 1000-yuan reward for fully meeting the demand).

[0150] Example calculation: Assume that on a certain day, Action A (decelerate production) is taken, resulting in the following:

[0151] 1) Inventory decreases by 5 tons → Inventory cost reduction: 5 tons × 100 yuan = 500 yuan.

[0152] 2) Due to decelerated production, the daily output is only 4 tons (the original plan was 5 tons) → Production efficiency penalty: (5 - 4) / 5 × 100% × 50 yuan = -100 yuan.

[0153] 3) Shortage of raw materials caused by supplier delays, possible out - of - stock in the future → Potential penalty estimate: - 200 yuan. Total reward: 500 - 100 - 200 = + 200 yuan.

[0154] 4. Q - value update: Optimization of neural network and Bellman equation

[0155] Training process:

[0156] Experience replay: Store historical state - action - reward - new state tuples (such as (s, a, r, s')).

[0157] Q - value prediction: Use the current neural network to estimate the Q - values of all actions in state s.

[0158] Target Q - value calculation: According to the Bellman equation, the target Q - value is r + γ×max(Q(s')) (γ is the discount factor, such as 0.9).

[0159] Loss function: Minimize the mean square error between the predicted Q - value and the target Q - value.

[0160] Parameter update: Adjust the neural network weights through backpropagation.

[0161] Example training steps:

[0162] Current state s: Inventory = 50 tons, Supplier = Delayed, Demand = 8 tons / day.

[0163] Action a: Slow down production (Q predicted value = 200).

[0164] Reward r: + 200 yuan.

[0165] New state s': Inventory = 45 tons, Supplier = Delayed, Demand = 8 tons / day.

[0166] Target Q - value: 200 + 0.9*max(Q(s')). Assume max(Q(s')) = 180, then the target value is 200 + 162 = 362.

[0167] Loss calculation: (362 - 200) 2 = 26244, update network parameters through gradient descent.

[0168] The following takes a meat processing enterprise facing the following problems as an example for illustration:

[0169] Problem 1. Demand fluctuation: The demand surges during holidays (e.g., the demand for beef sausage increases from 5 tons per day to 15 tons before the Mid - Autumn Festival).

[0170] Problem 2. Supply chain risk: The key supplier suspends production due to the epidemic, and the raw material supply decreases by 30%.

[0171] The DQN decision-making process is as follows:

[0172] 1) State perception: Inventory = 20 tons, production progress = 8 tons per day, demand forecast = 15 tons / day, supplier = production cut by 30%.

[0173] 2) Action selection:

[0174] Accelerate production (+20% capacity to 9.6 tons per day).

[0175] Urgent procurement of raw materials from alternative suppliers (cost increases by 10%).

[0176] 3) Reward calculation:

[0177] Meet festival demand → Reward +1500 yuan.

[0178] Increase in raw material cost → Penalty -500 yuan.

[0179] Reduction in the risk of inventory depletion → Reward +300 yuan.

[0180] Total reward: 1500 - 500 + 300 = +1300 yuan.

[0181] Q-value update: Through iterative training, the model learns that "increasing production capacity in advance + multi-source procurement" is the optimal strategy to cope with sudden demand surges and supply shortages.

[0182] Generally speaking, DQN realizes the dynamic optimization of the meat processing industry through the following mechanisms: "Comprehensive state perception: integrating multi-dimensional data such as inventory, production, and supply chain", "Flexible action decision-making: selecting the optimal adjustment strategy with the best cost-benefit in a complex environment", "Fine reward design: balancing short-term costs and long-term stability", and "Continuous learning and optimization: adapting to the dynamic changes of the market and supply chain through Q-value update".

[0183] As another implementation method, the reinforcement learning algorithm is the policy gradient algorithm. In some continuous decision-making scenarios, such as production rhythm adjustment and supply chain management, the policy gradient method can better solve the problem of continuous action space.

[0184] The following will illustrate the policy gradient algorithm in inventory and production optimization of the meat processing industry from several aspects:

[0185] 1. Policy representation: Modeling of continuous action space

[0186] Definition: The policy gradient algorithm directly outputs the probability distribution of actions through a deep neural network and is applicable to continuous adjustment scenarios (such as fine-tuning of production speed).

[0187] Network architecture:

[0188] Input layer: State variables (inventory level, demand forecast, supply chain status, etc.).

[0189] Hidden layer: A fully connected layer (e.g., with 128 neurons) extracts features.

[0190] Output layer: Gaussian distribution parameters: The mean (μ) and standard deviation (σ) of the output action. For example, the action of adjusting the production rhythm is a continuous value (such as the percentage increase or decrease in daily production capacity).

[0191] Example scenario:

[0192] State s: [Inventory = 30 tons, Demand forecast = 10 tons / day, Supplier reliability = 0.8 (80% on-time delivery)]

[0193] Output of the policy network:

[0194] Mean of production adjustment μ = +5% (increase production capacity by 5%).

[0195] Standard deviation σ = 2% (allowing random exploration within the range of +3% to +7%).

[0196] Action execution: Sample from the distribution N(5%, 2%), and the actual action is +6%, that is, the daily production capacity is increased from 10 tons to 10.6 tons.

[0197] 2. Policy update: By performing gradient ascent on the cumulative reward during the decision-making process, the policy gradient algorithm adjusts the weights of the neural network to select a better policy in future states.

[0198] Training process:

[0199] Trajectory sampling: Execute the policy in the environment to generate a state-action-reward sequence.

[0200] Cumulative reward calculation: Discount and accumulate the rewards for each trajectory (e.g., G_t = Σγ k r_{t+k}).

[0201] Gradient calculation: Calculate the gradient direction that maximizes the cumulative reward through the policy gradient theorem.

[0202] Parameter update: Adjust the weights of the neural network along the gradient direction.

[0203] Example training process:

[0204] Scenario: The enterprise status and actions within a certain week are as follows:

[0205]

[0206]

[0207] Cumulative Reward (γ = 0.9): G1 = 500 + 0.9×800 + 0.9 2 ×1200 = 500 + 720 + 97² = 2192 yuan.

[0208] Gradient Update: The weights of the policy network are adjusted in the direction of increasing high-reward actions (such as +8% - +10% production capacity adjustment), reducing the probability of inefficient actions (such as overproduction leading to inventory backlog).

[0209] 3. Advantages: Policy gradients perform well in handling continuous actions and high-dimensional state spaces, and can dynamically adjust production plans, material procurement, and inventory scheduling to adapt to the complex requirements of the meat processing industry.

[0210] For ease of understanding, the following provides two specific examples:

[0211] Example 1: Fine-tuning of production rhythm

[0212] Problem: Traditional solutions can only select discrete actions (such as "increase production by 10%" or "decrease production by 10%") and cannot achieve fine-tuning.

[0213] Policy gradient solution:

[0214] Output continuous actions (such as "increase production by 6.3%") to accurately match demand fluctuations.

[0215] Effect: Through continuous adjustment, an enterprise increased its production capacity utilization rate from 75% to 92%, reducing production capacity waste.

[0216] Example 2: Multivariable collaborative optimization

[0217] Problem: Meat processing requires simultaneous adjustment of production, procurement, and promotion, with a high-dimensional action space.

[0218] Policy gradient solution:

[0219] Joint action output:

[0220] Production adjustment: +5%

[0221] Procurement batches: +8%

[0222] Promotion intensity: 12% price cut

[0223] Effect: During the low-demand period, through a combination strategy of slightly increasing production + promotion, the inventory turnover rate increased by 30%.

[0224] The following provides a specific scenario implementation example to illustrate this embodiment:

[0225] Scenario: Coping with sudden supply chain disruptions

[0226] It should be noted that in the original text, "972" was wrongly written as "97²" in line , and this has been corrected in the translation.Background: Due to the discontinuation of supply by the beef supplier, an enterprise needs to dynamically adjust its production and procurement strategies.

[0227] Status s:

Inventory = 15 tons, Demand = 8 tons per day, Supplier status = Discontinued supply, Cost of alternative supplier = +20%

[0228] Output of the policy network:

[0229] Production adjustment: -10% (reduce the production capacity of beef sausages).

[0230] Procurement action: Switch to an alternative supplier, with a procurement volume increase of +15%.

[0231] Promotion action: Reduce the price of in-stock products by 8% to accelerate shipments.

[0232] Reward calculation:

[0233] Avoiding raw material depletion by reducing production: +300 yuan.

[0234] Increase in procurement cost: -200 yuan.

[0235] Clearing inventory through promotion: +150 yuan.

[0236] Total reward: 300 - 200 + 150 = +250 yuan

[0237] Long-term effect: Through policy gradient learning, in similar supply chain disruption events, the system automatically selects a combined strategy of "production reduction + multi-source procurement + promotion", reducing the out-of-stock rate by 45% and cost overruns by 30%.

[0238] Generally speaking, the policy gradient algorithm improves the intelligent level of supply chain management in the meat processing industry through the following mechanisms: "Continuous action space: Production rhythm, procurement volume, etc. support fine-tuning (such as increasing production by 6.3% instead of a fixed 10%)", "Multi-variable collaborative decision-making: Synchronously optimize production, procurement, and promotion strategies to avoid local optima", "High-dimensional state processing ability: Integrate dozens of dimensions of states such as inventory, demand, and supply chain to generate a global optimal strategy", and "Dynamic adaptability: Quickly adjust strategies in the face of sudden fluctuations (such as the epidemic, weather), reducing manual intervention".

[0239] Based on the above, through the use of the reinforcement learning algorithm, the embodiments of the present application can autonomously learn and optimize decisions in a real-time changing environment. The system continuously collects real-time data such as the market, inventory, and production resources, dynamically adjusts the production plan and procurement strategy, realizes intelligent real-time response and adaptive adjustment, and ensures the ability to quickly respond to changes in market demand and supply chain fluctuations. The introduction of reinforcement learning enables the system to optimize decisions through continuous trial and error, improve the response speed, and reduce manual intervention.

[0240] III. Multi-objective optimization algorithm

[0241] As an implementation, the multi-objective optimization algorithm is the Genetic Algorithm (GA for short). The genetic algorithm is an optimization method based on natural selection and genetic mechanisms, which can handle multi-objective problems. The following will illustrate the application of the genetic algorithm in inventory and production optimization in the meat processing industry from several aspects:

[0242] 1. Chromosome encoding: Each solution is encoded as a chromosome, where genes represent different decision variables in the material requirements plan (such as inventory level, production rhythm, procurement batch, etc.). For the sake of easy understanding, for example:

[0243] Scenario: A meat processing enterprise needs to optimize the following decision variables:

[0244] Inventory level (unit: ton);

[0245] Production rhythm (daily output, unit: ton);

[0246] Procurement batch (quantity per procurement batch, unit: ton);

[0247] Promotion frequency (number of promotions per week).

[0248] Chromosome encoding example: Encode the above variables into a gene sequence, and each gene corresponds to the value of a variable. For example:

[0249] Gene 1 (inventory level): The value range is 50 - 100 tons → encoded as binary 0110010 (decimal 50 tons).

[0250] Gene 2 (production rhythm): The value range is 5 - 15 tons / day → encoded as 1001 (decimal 9 tons / day).

[0251] Gene 3 (procurement batch): 10 - 50 tons per batch → encoded as 001110 (decimal 28 tons).

[0252] Gene 4 (promotion frequency): 0 - 3 times per week → encoded as 11 (decimal 3 times).

[0253] Complete chromosome: 0110010 1001 001110 11, this complete chromosome represents: inventory 50 tons, daily output 9 tons, procurement batch 28 tons, and 3 promotions per week.

[0254] 2. Fitness function: The fitness function of multi-objective optimization includes multiple measurement indicators, such as inventory holding cost, shelf-life constraint, production efficiency, market demand satisfaction, etc. The algorithm evaluates the quality of each chromosome by weighted summation of these fitnesses.

[0255] First, define the objective function:

[0256] Inventory cost: Quantity in stock × Unit storage cost (e.g., 50 tons × 100 yuan / ton = 5,000 yuan).

[0257] Shelf-life loss: Quantity of products approaching expiration × Unit loss (e.g., 10 tons × 5,000 yuan = 50,000 yuan).

[0258] Production efficiency: Actual output / Maximum production capacity (e.g., 9 tons / 12 tons = 75% → Reward coefficient 75% × 1,000 yuan = 750 yuan).

[0259] Market demand satisfaction: Actual delivery quantity / Demand quantity (e.g., Delivery rate 90% → Reward 900 yuan).

[0260] Secondly, perform fitness calculation: Convert multiple objectives into a single fitness value through weighted summation: Fitness = w1 × (1 / Inventory cost) + w2 × (1 / Shelf-life loss) + w3 × Production efficiency + w4 × Market demand satisfaction.

[0261] Assume weights w1 = 0.4, w2 = 0.3, w3 = 0.2, w4 = 0.1, then: Fitness = 0.4 × (1 / 5,000) + 0.3 × (1 / 50,000) + 0.2 × 0.75 + 0.1 × 0.9 ≈ 0.00008 + 0.000006 + 0.15 + 0.09 = 0.240086

[0262] Objective: Maximize the fitness value.

[0263] 3. Genetic operations: Through genetic operations, the genetic algorithm selects, crosses, and mutates the solutions of each generation to find the global optimal solution. In each generation, the optimal solution gradually evolves until the optimal material requirements plan that balances multiple objectives is found. For ease of understanding, for example:

[0264] Initial population: Randomly generate 100 chromosomes, each representing a material plan. <L

[0265] Selection: Rank by fitness and retain the top 20% of high-quality individuals (e.g., the plan with 60 tons in stock and a daily output of 10 tons).

[0266] Crossover: Exchange some genes of two parental chromosomes to generate new individuals.

[0267] Parent 1: 0110010 (Inventory 50 tons)|1001 (Production 9 tons)

[0268] Parent 2: 0111100 (Inventory 60 tons)|1010 (Production 10 tons)

[0269] Offspring: 0111100 (Inventory 60 tons) | 1001 (Production 9 tons)

[0270] Mutation: Randomly modify a certain bit of a certain gene (e.g., the inventory gene changes from 0110010 to 0110000 → 48 tons).

[0271] Iteration result: After 50 generations of evolution, the optimal solution is: Inventory = 70 tons (balancing the shortage risk), Daily production = 12 tons (full-load production), Purchase batch = 40 tons (economic order quantity), Promotion = 2 times / week (balancing inventory clearance and profit).

[0272] At this time, the fitness value has increased to 0.35, significantly better than the initial population.

[0273] As another implementation method, the multi-objective optimization algorithm is the Ant Colony Optimization (ACO) algorithm. The ant colony algorithm is an optimization algorithm based on swarm intelligence that can find the global optimal solution under complex constraints. The application process of the ant colony algorithm is illustrated as follows in several aspects:

[0274] 1. Ant colony behavior simulation: By simulating the process of ants finding paths between food sources and nests, and continuously iterating, the optimal material scheduling plan is found. Each "ant" represents a possible material requirement plan, which is illustrated as follows:

[0275] Scenario: A certain enterprise needs to optimize the supply chain path from raw material procurement to production and delivery, with the goal of minimizing inventory costs and maximizing production efficiency.

[0276] Path definition:

[0277] Nodes: Suppliers (S1, S2), Production workshops (P1, P2), Warehouses (W1, W2), Customers (C1, C2).

[0278] Path: The path selected by the ant (i.e., the solution) represents a scheduling plan. For example:

[0279] S1 → P1 → W1 → C1 (Supplier 1 supplies goods → Workshop 1 produces → Warehouse 1 stores → Deliver to Customer 1).

[0280] Example of ant behavior:

[0281] Ant A selects the path: S1 → P1 → W1 → C1, Total cost = Transportation cost 2000 yuan + Inventory cost 3000 yuan = 5000 yuan.

[0282] Ant B selects the path: S2 → P2 → W2 → C2, Total cost = Transportation cost 2500 yuan + Inventory cost 2500 yuan = 5000 yuan.

[0283] Ant C selects the path: S1→P2→W1→C2, total cost = transportation cost of 1800 yuan + inventory cost of 3500 yuan = 5300 yuan.

[0284] 2. Pheromone update: Each ant leaves pheromone on the path according to the quality of the solution it finds. The quality of the path depends on the evaluation result of the multi-objective function. The higher the pheromone concentration, the better the path.

[0285] Pheromone rule: Ants release pheromone on the path, and the pheromone concentration is proportional to the fitness of the path (such as the reciprocal of the total cost).

[0286] Pheromone update formula: τ_{ij}(t + 1) = (1 - ρ)τ_{ij}(t) + Δτ

[0287] where ρ = 0.1 is the evaporation rate, and Δτ = Q / C (Q is a constant, and C is the total cost of the path).

[0288] Example calculation:

[0289] The path cost of Ant A = 5000 yuan → Δτ = 1000 / 5000 = 0.2, and the pheromone increases by 0.2.

[0290] The path cost of Ant B = 5000 yuan → also increases by 0.2.

[0291] The path cost of Ant C = 5300 yuan → Δτ = 1000 / 5300 ≈ 0.19, and the pheromone increases less.

[0292] In summary: The pheromone concentrations of the paths S1→P1→W1→C1 and S2→P2→W2→C2 are higher, and subsequent ants are more likely to choose these paths.

[0293] 3. Global optimization: Through the accumulation and gradual strengthening of pheromone, a global optimal solution that balances multiple objectives (such as inventory cost, production efficiency, market response speed, etc.) can be found.

[0294] Multi-objective fitness function:

[0295] Fitness = 0.5×(1 / inventory cost) + 0.3×production efficiency + 0.2×on-time delivery rate

[0296] Iterative optimization process:

[0297] Initial exploration: Ants randomly select paths to generate diverse solutions.

[0298] Pheromone accumulation: The pheromone concentrations of low-cost and high-efficiency paths gradually increase.

[0299] Convergence result: After 100 iterations, the optimal path is S2→P2→W2→C2, and its advantages are as follows:

[0300] Inventory cost: 2,500 yuan (the warehousing rate of W2 is lower).

[0301] Production efficiency: The capacity utilization rate of workshop P2 reaches 95%.

[0302] On-time delivery rate: 100% (warehouse W2 is close to customer C2).

[0303] Actual effect: The inventory cost is reduced by 20%, the production efficiency is increased by 15%, and the on-time delivery rate of orders is increased from 80% to 98%.

[0304] It should be noted that among the above two implementation methods, the genetic algorithm is suitable for long-term plan optimization (such as quarterly production scheduling), and can flexibly integrate multiple variables such as inventory, production, and promotion through gene coding (a typical example under this algorithm is as follows: a certain enterprise reduces the expired loss from 15% to 5% through GA, and the production efficiency is increased by 25%). The ant colony algorithm is suitable for dynamic scheduling problems (such as daily logistics path planning). It can adjust the strategy in real time through pheromone guidance to cope with supply chain fluctuations (a typical example under this algorithm is as follows: a cold chain logistics enterprise uses ACO to optimize the distribution path, the transportation cost is reduced by 18%, and the on-time delivery rate is increased to 95%).

[0305] Based on the above, the embodiments of the present application introduce multi-objective optimization algorithms (such as genetic algorithms, ant colony algorithms), consider multiple objectives such as inventory cost, production efficiency, and shelf life at the same time, and find the optimal material requirements plan and production scheduling strategy. These algorithms can achieve a balance between different objectives, ensuring not only reducing inventory costs, but also ensuring on-time production and delivery, reducing waste, and optimizing the overall resource allocation of the enterprise.

[0306] Furthermore, the present application can integrate the above algorithms into a unified material requirements planning system. The system generates dynamic material requirements plans and production scheduling strategies by real-time monitoring and analyzing sales forecasts, market demands, inventory status, production progress, and supply chain status, and combining deep learning, reinforcement learning, and multi-objective optimization algorithms.

[0307] As Figure 2 shown, the material requirements planning system includes: a data input layer 21, a prediction layer 22, a decision layer 23, and a feedback layer 24.

[0308] Among them, the data input layer 21 is used to: collect data related to meat processing in real time, preprocess it, and input it into the deep learning and reinforcement learning models.

[0309] For the sake of easy understanding, the following is an example to introduce the collection and preprocessing process of the above data related to meat processing:

[0310] 1. Sales data:

[0311] Source: ERP system, e-commerce platform, POS terminal.

[0312] Example: In a certain month, the sales volume of beef sausage is 500 tons, and the online channel accounts for 40%.

[0313] Preprocessing: Clean outliers (such as return data), standardize the format (such as unified to tons / month).

[0314] 2. Market demand data:

[0315] Source: Market research reports, social media sentiment, macroeconomic indicators.

[0316] Example: The number of mentions of "low-fat beef products" on social media has increased by 20%.

[0317] Preprocessing: Extract keywords through NLP, quantify the sentiment index (such as positive sentiment +0.8).

[0318] 3. Inventory status:

[0319] Source: Warehouse Management System (WMS), RFID tags.

[0320] Example: The current inventory is 200 tons, of which 50 tons are approaching the expiration date.

[0321] Preprocessing: Calculate the inventory turnover rate and the shelf-life pressure index.

[0322] 4. Production progress:

[0323] Source: MES system, production line sensors.

[0324] Example: The current daily production capacity is 10 tons, and the remaining order quantity is 100 tons.

[0325] Preprocessing: Calculate the production completion rate and the remaining construction period.

[0326] 5. Supply chain status:

[0327] Source: Supplier ERP, logistics tracking system.

[0328] Example: A certain supplier has reduced production by 30% due to the epidemic, and the delivery is delayed by 2 days.

[0329] Preprocessing: Quantify the supply chain reliability (such as delay probability = 0.2).

[0330] The output of the data input layer 21 is a standardized data matrix, for example: [Sales volume = 500 tons, Market demand index = 0.8, Inventory = 200 tons, Production progress = 10 tons / day, Supply chain reliability = 0.8].

[0331] The prediction layer 22 is used to: use a deep learning model for sales and market demand prediction, and the prediction results are directly used for production and inventory management.

[0332] The following is an example to introduce the input and output of the deep learning model:

[0333] LSTM model:

[0334] Input: Sales volume in the past 24 months, holiday markers, promotional activities.

[0335] Output: Sales volume prediction for the next 3 months (such as [600 tons, 650 tons, 700 tons]).

[0336] VAE model:

[0337] Input: Market research data, consumer preferences, macroeconomic indicators.

[0338] Output: Probability distribution of future demand (such as N(650 tons, 50 tons)).

[0339] Example scenario:

[0340] Input: [Historical sales volume = 500 tons, Holiday = 1 (Spring Festival), Promotional intensity = 0.8, Economic indicator = 0.9]

[0341] Output:

[0342] LSTM model: Sales volume for the next 3 months [600 tons, 650 tons, 700 tons].

[0343] VAE model: Sales volume for the next 1 month N(600 tons, 50 tons).

[0344] The decision-making layer 23 is used to: combine the reinforcement learning model and the multi-objective optimization model to generate the optimal solutions for production scheduling, material requirements, and procurement plans.

[0345] The following is an example to introduce the application process of the reinforcement learning model and the multi-objective optimization model:

[0346] Reinforcement learning (DQN):

[0347] State: Current inventory, production progress, demand prediction, supply chain status.

[0348] Action: Adjust production rhythm, procurement batches, promotional strategies.

[0349] Rewards: Inventory cost, production efficiency, market demand satisfaction.

[0350] Output: Optimal action strategy (such as increasing production by 10% and increasing the number of procurement batches by 15%).

[0351] Multi-objective optimization (GA / ACO):

[0352] Objective: Minimize inventory cost, shelf-life loss, and maximize production efficiency.

[0353] Output: Pareto optimal solution (such as inventory = 700 tons, daily production = 12 tons, procurement = 800 tons).

[0354] Example scenario:

[0355] Status:

Inventory = 200 tons, production progress = 10 tons / day, demand forecast = 650 tons, supply chain reliability = 0.8

[0356] Decision:

[0357] DQN: Increase production by 10% → Increase daily production to 11 tons.

[0358] GA: Optimize the procurement batch to 800 tons and the inventory level to 700 tons.

[0359] Result: Inventory cost reduced by 15%, production efficiency increased by 20%, and order fulfillment rate increased to 95%.

[0360] The feedback layer 24 is used to: Continuously adjust the prediction model and decision-making strategy through a real-time feedback mechanism to ensure that the system can adapt to the dynamic changes of the environment.

[0361] The following is a specific example of the application process of the feedback layer 24 step by step:

[0362] S1. Data monitoring: Real-time collection of the latest sales, inventory, and production data.

[0363] S2. Error calculation: Compare the predicted value with the actual value (such as predicted sales volume of 600 tons and actual of 550 tons)

[0364] S3. Model update:

[0365] Adjust the weights of LSTM and reduce the weights affected by holidays.

[0366] Update the DQN reward function and increase the penalty coefficient for inventory cost.

[0367] S4. Strategy optimization: Regenerate the production and procurement plan (such as reducing production by 5% and reducing the number of procurement batches by 10%).

[0368] Combined with the above material requirements planning system, the following provides two specific scenario examples to illustrate the solution of this application:

[0369] Scenario Example 1: Coping with the sudden increase in demand during holidays and inventory backlogs

[0370] Background: A meat processing enterprise mainly produces beef sausages and ready-to-eat meat products, and the shelf life of the products is 3 months. The enterprise faces the following challenges:

[0371] Fluctuation in demand during holidays: The sales volume surges by 200% before the Spring Festival, but the demand drops sharply after the festival, resulting in inventory backlogs.

[0372] Difficulties in shelf life management: The backlogged inventory is prone to approaching the expiration date, causing waste.

[0373] Rigid production planning: The traditional MRP system cannot flexibly adjust production capacity, resulting in insufficient production capacity during peak seasons and overproduction capacity during off-seasons.

[0374] Solution: Achieve dynamic optimization through the material requirements planning system:

[0375] Data input layer 21:

[0376] Collect real-time data, including: historical sales data (sales volume around the Spring Festival in the past 3 years), current inventory status (inventory of 500 tons, of which 100 tons will expire within 1 month), supply chain status (supplier A's delivery is delayed by 3 days), and market demand signals (social media shows that the search volume for "Spring Festival gift boxes" has increased by 150%).

[0377] Prediction layer 22:

[0378] LSTM demand prediction:

[0379] Input: Sales volume, promotion activities, and weather data during the Spring Festival in the past 3 years.

[0380] Output: The predicted sales volume 1 month before the Spring Festival is 1200 tons (a 200% increase compared to normal days).

[0381] VAE probability prediction:

[0382] Output: The demand distribution after the Spring Festival is N(400 tons, 100 tons), with a high probability of slow-moving risk.

[0383] Decision-making layer 23:

[0384] Reinforcement learning (DQN) dynamic adjustment:

[0385] Status: Inventory = 500 tons, demand prediction = 1200 tons, production progress = 20 tons / day, supplier delay = 3 days.

[0386] Actions:

[0387] Increase production to 30 tons per day (a 50% increase in production capacity).

[0388] Urgent procurement of raw materials from alternative suppliers (a 10% increase in cost).

[0389] Launch a pre-Spring Festival promotion (a 10% increase in sales volume through the full reduction activity).

[0390] Reward calculation:

[0391] Reward for meeting demand: +5000 yuan.

[0392] Penalty for inventory backlog: -2000 yuan (calculated based on the expiration risk).

[0393] Multi-objective optimization (GA):

[0394] Chromosome encoding: [Inventory = 600 tons, production rhythm = 30 tons per day, promotion frequency = 3 times per week].

[0395] Fitness function: Balance inventory cost, production efficiency, and shelf-life loss.

[0396] Output: The Pareto optimal solution is inventory = 700 tons, production = 28 tons per day, and promotion = 4 times per week.

[0397] Feedback layer 24:

[0398] Data monitoring after the festival: The actual sales volume is 350 tons (lower than the forecast), and the remaining inventory is 400 tons (200 tons of which are approaching expiration).

[0399] Model adjustment:

[0400] The LSTM model reduces the weight of holidays and increases the impact of economic indicators.

[0401] The DQN increases the inventory backlog penalty coefficient.

[0402] Policy update:

[0403] Reduce production to 15 tons per day and launch a "buy one get one free" promotion to clear the inventory.

[0404] Effect:

[0405] Reduction in inventory backlog: The expiration loss is reduced from 15% to 5%, saving 500,000 yuan in cost.

[0406] Improvement in demand satisfaction rate: The on-time delivery rate of Spring Festival orders is increased from 70% to 95%.

[0407] Optimization of production capacity utilization: The waste of production capacity in the off-season is reduced by 30%.

[0408] Example 2: Coping with supply chain disruptions and raw material shortages

[0409] Background: A company relies on a single beef supplier. Due to an outbreak, the supplier halted production and faced the following challenges:

[0410] Shortage of raw materials: inventory can only sustain 7 days of production, with a gap of 200 tons.

[0411] Risk of production stagnation: Failure to purchase alternative raw materials in a timely manner will result in order default.

[0412] Cost surge: Alternative supplier quotes have increased by 20%, and there is a need to balance procurement costs with delivery pressure.

[0413] Solution: Real-time optimization of supply chain and production strategies through the system:

[0414] Data input layer 21:

[0415] Real-time data collection includes: supply chain status (the original supplier has stopped production, and the alternative supplier B’s quotation has increased by 20%), 100 tons of raw materials remaining in inventory (15 tons consumed daily), and customer order priority (VIP customer orders must be met first).

[0416] Prediction layer 22:

[0417] VAE market demand forecast:

[0418] Output: The demand distribution in the next two weeks is N(150 tons, 30 tons).

[0419] LSTM supply chain risk prediction:

[0420] Input: Supplier’s historical delivery data and epidemic updates.

[0421] Output: The probability of the original supplier resuming work is 10%, and the probability of delay of alternative supplier B is 5%.

[0422] Decision-making layer 23:

[0423] Reinforcement Learning (Policy Gradient) Dynamic Scheduling:

[0424] Status: Raw material inventory = 100 tons, demand = 150 tons, supplier B cost = +20%.

[0425] action:

[0426] Purchase 80 tons from supplier B (enough for 5 days of production).

[0427] Adjust production priorities: give priority to VIP customer orders (accounting for 50% of total demand).

[0428] Start internal raw material allocation (allocate 20 tons from other factories).

[0429] Reward function:

[0430] Order delivery reward: +3000 yuan (full delivery of VIP orders).

[0431] Purchasing cost penalty: -1000 yuan.

[0432] Ant Colony Optimization (ACO) path optimization:

[0433] Path definition: Supplier B → production workshop → customer.

[0434] Pheromone update: Select the path with the lowest cost and the fastest delivery.

[0435] Output the optimal path: Urgently airlift 10 tons of raw materials, and transport the remaining 70 tons by land.

[0436] Feedback layer 24:

[0437] Real-time monitoring: The delivery of Supplier B is delayed by 1 day, and the delivery progress of VIP orders lags behind.

[0438] Strategy adjustment:

[0439] Increase the weight of "delivery on-time rate" in the policy gradient model.

[0440] ACO re-plans the path and adds 5 tons of raw materials to be airlifted.

[0441] Effect:

[0442] Guarantee production continuity: 100% of VIP orders are delivered on time, avoiding a liquidated damages loss of 1 million yuan.

[0443] Cost control: Through the mixed transportation strategy, the purchasing cost only increases by 12% (20% lower than expected).

[0444] Improve supply chain resilience: Establish a multi-supplier backup mechanism to reduce the out-of-stock risk by 40%.

[0445] In summary, the above two scenario embodiments demonstrate the practical application value of the material requirements planning system in complex scenarios:

[0446] 1. Dynamic adaptability: Respond to demand fluctuations and supply chain disruptions through real-time data feedback and model adjustment.

[0447] 2. Multi-objective optimization: Balance conflicting objectives such as cost, efficiency, and delivery on-time rate to achieve global optimality.

[0448] 3. Industry customization: Provide precise solutions for the particularities of meat processing (such as shelf life, seasonality).

[0449] Based on the above, the advantages of the embodiments of this application compared with the traditional solutions are briefly described below:

[0450] 1. High-precision market demand forecasting

[0451] By introducing deep learning algorithms (such as LSTM), this application can extract long-term dependence features from historical sales data and, at the same time, analyze by combining multi-dimensional market factors (such as holiday effects, market promotions, and competitor movements) to generate dynamic demand forecasts. The specific advantages are as follows:

[0452] Capturing complex non-linear relationships: Compared with traditional linear models such as time series analysis and regression analysis, deep learning can automatically learn complex non-linear relationships, which is particularly suitable for scenarios with large demand fluctuations in the meat processing industry.

[0453] More adaptable prediction: It can automatically adjust the prediction model as market data is updated, and respond in real time to seasonal, sudden market events, etc.

[0454] Greatly improved prediction accuracy: It reduces the problems of material shortages or overstocking caused by market demand fluctuations, and helps enterprises maintain agility and efficient operation in a highly competitive market environment.

[0455] Enhancing market response speed: The system can give early warnings of market demand changes, enabling production and procurement to be adjusted more promptly and avoiding supply-demand mismatches.

[0456] 2. Intelligent production scheduling and inventory management

[0457] By adopting reinforcement learning algorithms (such as DQN and policy gradient algorithms), the system of this application can perform dynamic optimization based on real-time production, inventory, procurement, and sales data. Reinforcement learning can autonomously learn the fluctuations and feedback in the supply chain and generate optimal material demand plans and production scheduling strategies. The specific advantages are as follows:

[0458] Real-time adaptive adjustment: The system achieves self-adaptability through reinforcement learning and can quickly adjust procurement, production plans, and inventory strategies in the face of supply chain fluctuations and market demand changes without manual intervention.

[0459] Reducing production lags and inventory overstocking: Through real-time monitoring and dynamic adjustment, the system can effectively prevent problems of excessive inventory or insufficient supply. Especially in fresh foods such as meat products, shelf life is a key issue.

[0460] Reducing inventory costs: Intelligently optimizing inventory levels, reducing unnecessary inventory overstocking, reducing waste caused by expiration or damage, and significantly reducing the operating costs of enterprises.

[0461] Improving production efficiency: The system optimizes production scheduling according to current demand and production capacity, making production plans more accurate, reducing production capacity waste, and improving resource utilization.

[0462] 3. Global Optimal Decision-making for Multi-objective Optimization

[0463] The present invention applies multi-objective optimization algorithms (such as genetic algorithms and particle swarm algorithms) to different links of the supply chain to balance multiple conflicting objectives (such as inventory cost, production efficiency, delivery time, and shelf life) in order to find the optimal solution. Multi-objective optimization can comprehensively consider multiple key factors in enterprise operations to achieve global optimality. The specific advantages are as follows:

[0464] Multi-dimensional Objective Balance: The system can find a balance among multiple conflicting objectives (such as cost, efficiency, and shelf life), unlike traditional systems that can only focus on a single objective (such as inventory minimization).

[0465] Solving Complex Decision-making Problems: For highly complex multi-variable problems such as production scheduling, inventory management, and procurement planning, the multi-objective optimization algorithm of the present invention can achieve the optimal material requirements planning by searching for the global optimal solution.

[0466] Maximizing Resource Utilization: By optimizing the allocation of enterprise resources, reducing redundant production and ineffective inventory, the overall efficiency of the supply chain is improved.

[0467] Reducing Production and Procurement Errors: The system ensures a high degree of matching between material requirements and actual production capacity and market demand according to the multi-objective optimization decision, avoiding unnecessary production and procurement losses.

[0468] 4. Intelligent Supply Chain Management throughout the Whole Process

[0469] This application constructs an integrated supply chain management system to automate and intelligently manage the whole process of sales forecasting, procurement planning, inventory management, and production scheduling. Through real-time data transmission and the integration of intelligent algorithms, the system can globally optimize all links of the supply chain. The specific advantages are as follows:

[0470] Whole-process Data Integration: Compared with traditional modular systems, the present invention realizes real-time data sharing and linkage among different links in the supply chain, avoiding the problem of information silos.

[0471] High Degree of Automation: The system can automatically generate and adjust the material requirements plan according to real-time data, reducing manual operations and interventions, and improving the response speed and decision-making accuracy.

[0472] Enhancing Operational Efficiency: From demand forecasting to procurement and production, the system forms a closed-loop automated decision-making process, reducing the interference of human factors on decision-making and ensuring the efficient operation of production and inventory management.

[0473] Coping with complex supply chain environments: In the face of supply chain fluctuations, raw material shortages, or sudden changes in market demand, the system can quickly make adjustments to reduce losses caused by information lags or improper decisions.

[0474] 5. Customized optimization for the meat processing industry

[0475] This application is specifically tailored to the characteristics of the meat processing industry, integrating shelf-life management, batch production management, and seasonal demand fluctuation analysis to ensure that the system meets the special requirements of the meat processing industry for shelf-life, inventory, and production batches. The specific advantages are as follows:

[0476] Precise shelf-life management: The system combines shelf-life and inventory turnover time to ensure that inventory management can minimize product expiration or waste.

[0477] Strong adaptability: The system can adjust production and procurement plans according to factors such as the shelf-life of different products and seasonal demands to ensure the efficient operation of the supply chain.

[0478] Reduce losses: In response to the shelf-life requirements of meat products, the system reduces product waste caused by expiration through refined inventory and production management.

[0479] Improve production efficiency and market response ability: For seasonal demands and special batch management, the optimization algorithm of this application can timely adjust production and inventory strategies to ensure that product supply matches market demand.

[0480] 6. Complex data processing and intelligent prediction

[0481] This application can process a large amount of complex multi-dimensional data, such as market demand, historical sales, supply chain data, production capacity, and human resources, by integrating advanced algorithms such as deep learning, reinforcement learning, and multi-objective optimization. Deep learning algorithms can extract key features from these complex data to achieve high-precision market demand prediction. The system shows higher prediction accuracy when dealing with non-linear and multi-dimensional data, especially when coping with seasonal changes and market fluctuations.

[0482] 7. Optimize supply chain and inventory management

[0483] The reinforcement learning algorithm in this application can adaptively adjust inventory management strategies in complex supply chain environments, dynamically optimize inventory levels according to real-time market demand, supply chain conditions, and production plans, and reduce inventory costs and material waste. The multi-objective optimization algorithm also ensures that the system can maintain a reasonable inventory level while meeting shelf-life management requirements, maximizing the utilization of warehousing and production resources.

[0484] 8. Reduce manual intervention and improve automation level

[0485] By integrating artificial intelligence technology, this application significantly reduces manual intervention. The reinforcement learning algorithm autonomously generates material requirements plans and production strategies by automatically learning the changes in the production, inventory, market demand, and other environments, greatly enhancing the degree of automation. By reducing the dependence on manual operations and empirical judgments, the system can make more accurate decisions in a shorter time, improving the operational efficiency of enterprises.

[0486] 9. Integrated Optimization of the Whole Process

[0487] This application integrates the entire process of sales forecasting, inventory management, production scheduling, and procurement planning to construct a closed-loop intelligent material requirements planning system. Through the unified management of the entire supply chain, the system can share and update data in real time, achieve the optimization of the entire process from demand forecasting to production execution, ensure the coordination of all links, and improve production efficiency and market response speed.

[0488] Figure 3 The following is a schematic structural diagram of a material processing device provided by an embodiment of this application, as Figure 3 shown, the device includes: an acquisition module 31, a prediction module 32, a production module 33, and an adjustment module 34.

[0489] The acquisition module 31 is used to acquire the demand-driven data of the target material and the market environment data corresponding to the target material. The demand-driven data is used to reflect the various parameters affecting the demand of the target material, and the market environment data is used to reflect the market changes corresponding to the target material.

[0490] The prediction module 32 is used to determine the sales volume prediction result of the target material within a set time period based on the demand-driven data according to the deep learning algorithm, and determine the market demand distribution of the target material within a set time period according to the market environment data.

[0491] The production module 33 is used to determine the production strategy according to the sales volume prediction result and the market demand distribution, and perform production operations on the target material based on the production strategy.

[0492] The adjustment module 34 is used to adjust the production strategy based on the actual market information during the production operation process according to the reinforcement learning algorithm and the multi-objective optimization algorithm, so as to perform production operations on the target material based on the adjusted production strategy.

[0493] Optionally, the prediction module 32 is specifically used for: inputting the demand-driven data into the trained sales volume prediction model so that the sales volume prediction model outputs the sales volume prediction result of the target material within a set time period; inputting the market environment data into the trained market demand prediction model so that the market demand prediction model outputs the market demand prediction result of the target material within a set time period.

[0494] Optionally, the sales volume prediction model is a long short-term memory network model, and the market demand prediction model is a variational autoencoder.

[0495] Optionally, the reinforcement learning algorithm includes a deep Q-network algorithm and a policy gradient algorithm.

[0496] Optionally, the multi-objective optimization algorithm includes a genetic algorithm and an ant colony algorithm.

[0497] Optionally, the device further includes an adjustment module configured to adjust the parameters of the deep learning algorithm and the reinforcement learning algorithm according to the actual sales volume and the actual market demand distribution of the target material.

[0498] Figure 3 The device shown can execute the steps in the foregoing embodiments. For the detailed execution process and technical effects, refer to the descriptions in the foregoing embodiments and will not be elaborated herein.

[0499] In a possible design, the structure of the foregoing Figure 3 shown material processing device can be implemented as an electronic device. As Figure 4 shown, the electronic device may include: a processor 41, a memory 42, and a communication interface 43. Among them, an executable code is stored on the memory 42. When the executable code is executed by the processor 41, the processor 41 can at least implement the material processing method provided in the foregoing embodiments.

[0500] In addition, an embodiment of the present invention provides a non-transitory machine-readable storage medium. An executable code is stored on the non-transitory machine-readable storage medium. When the executable code is executed by a processor of an electronic device, the processor can at least implement the material processing method provided in the foregoing embodiments.

[0501] In addition, an embodiment of the present invention provides a computer program product. The computer program product includes a computer program. When the computer program is executed by a processor of an electronic device, the processor can at least implement the material processing method provided in the foregoing embodiments.

[0502] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0503] This application is described with reference to the flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, and the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processors of general-purpose computers, special-purpose computers, embedded processors, or other programmable data processing devices to generate a machine, such that the instructions executed by the processors of the computer or other programmable data processing devices produce means for implementing the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or in multiple blocks.

[0504] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory produce a manufactured article including instruction means that implement the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or in multiple blocks.

[0505] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or in multiple blocks.

[0506] In a typical configuration, a computing device includes one or more processors (CPUs), an input / output interface, a network interface, and memory.

[0507] The memory may include non-permanent memory in the form of computer-readable media, random access memory (RAM), and / or non-volatile memory such as read-only memory (ROM) or flash memory (flash RAM). The memory is an example of computer-readable media.

[0508] Computer-readable media includes permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory computer-readable media (transitory media), such as modulated data signals and carrier waves.

[0509] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, commodity, or apparatus that includes the element.

[0510] The above are merely embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should all be included within the scope of the claims of the present application.

Claims

1. A material handling method, characterized in that, Including: Obtain the demand-driven data of the target material and the market environment data corresponding to the target material. The demand-driven data is used to reflect various parameters affecting the demand of the target material, and the market environment data is used to reflect the market changes corresponding to the target material; Based on the deep learning algorithm, determine the sales volume prediction result of the target material within a set time period according to the demand-driven data, and determine the market demand distribution of the target material within a set time period according to the market environment data; Determine the production strategy according to the sales volume prediction result and the market demand distribution, and perform production operations on the target material based on the production strategy; Based on the actual market information in the production operation process, adjust the production strategy according to the reinforcement learning algorithm and the multi-objective optimization algorithm, so as to perform production operations on the target material based on the adjusted production strategy.

2. The method according to claim 1, characterized in that, The step of, based on the deep learning algorithm, determining the sales volume prediction result of the target material within a set time period according to the demand-driven data, and determining the market demand distribution of the target material within a set time period according to the market environment data, includes: Input the demand-driven data into the trained sales volume prediction model, so that the sales volume prediction model outputs the sales volume prediction result of the target material within a set time period; Input the market environment data into the trained market demand prediction model, so that the market demand prediction model outputs the market demand prediction result of the target material within a set time period.

3. The method according to claim 2, characterized in that, The sales volume prediction model is a long short-term memory network model, and the market demand prediction model is a variational autoencoder.

4. The method according to claim 1, wherein The reinforcement learning algorithm includes the deep Q-network algorithm and the policy gradient algorithm.

5. The method according to claim 1, wherein The multi-objective optimization algorithm includes the genetic algorithm and the ant colony algorithm.

6. The method according to any one of claims 1-5, characterized in that, After performing production operations on the target material based on the adjusted production strategy, the method further includes: Adjust the parameters of the deep learning algorithm and the reinforcement learning algorithm according to the actual sales volume and the actual market demand distribution of the target material.

7. A material handling device, characterized in that, Including: An acquisition module, configured to acquire the demand-driven data of the target material and the market environment data corresponding to the target material. The demand-driven data is used to reflect various parameters affecting the demand of the target material, and the market environment data is used to reflect the market changes corresponding to the target material; A prediction module, configured to, based on the deep learning algorithm, determine the sales volume prediction result of the target material within a set time period according to the demand-driven data, and determine the market demand distribution of the target material within a set time period according to the market environment data; A production module, configured to determine the production strategy according to the sales volume prediction result and the market demand distribution, and perform production operations on the target material based on the production strategy; An adjustment module, configured to, based on the actual market information in the production operation process, adjust the production strategy according to the reinforcement learning algorithm and the multi-objective optimization algorithm, so as to perform production operations on the target material based on the adjusted production strategy.

8. An electronic device, characterized in that, Including: A memory, a processor, and a communication interface; wherein, executable code is stored on the memory, and when the executable code is executed by the processor, the processor is caused to execute the material processing method according to any one of claims 1 to 6.

9. A non-transitory machine-readable storage medium, characterized in that, Executable code is stored on the non-transitory machine-readable storage medium, and when the executable code is executed by the processor of the electronic device, the processor is caused to execute the material processing method according to any one of claims 1 to 6.

10. A computer program product, characterized in that, Comprising: A computer program, which when executed by the processor of the electronic device, causes the processor to execute the material processing method according to any one of claims 1 to 6.

Images