A multi-category supply chain dynamic scheduling and optimization method based on S2B2C

Abstract

The application discloses a kind of based on S2B2C multi-class supply chain dynamic scheduling and optimization method, it is related to supply chain management and intelligent decision-making technical field, solve the problem that data cooperation is poor in existing S2B2C mode, multi-class scheduling lacks differentiation and dynamic response capability is insufficient, multi-source data acquisition and multi-class portrait construction, access S, B, C three end data, set differentiated weight and priority, collaborative demand prediction, and combine time series and machine learning algorithm, introduce B end plan factor correction prediction deviation, construct multi-objective optimization model, balance cost and service level, set multi-class mixed constraint, solve by using genetic algorithm combined with reinforcement learning mechanism, generate purchase, allocation and logistics instruction, real-time monitoring and abnormal response, based on Internet of Things data triggers dynamic rescheduling, performance evaluation and closed-loop optimization, the application realizes supply chain whole link cooperation and fine scheduling, improves supply-demand matching precision, decision efficiency and supply chain risk resistance ability.

Description

Technical Field

[0001] This invention relates to the field of supply chain management technology, and in particular to a method for dynamic scheduling and optimization of multi-category supply chains based on S2B2C. Background Technology

[0002] With the rapid development of e-commerce and new retail, S2B2C (Supplier to Business to Consumer) has emerged as a new business model. This model empowers distributors / retailers (B) through a supply chain platform (S) to jointly serve end consumers (C), aiming to improve circulation efficiency by integrating supply chain resources. However, in practical applications, existing supply chain scheduling methods still suffer from many technical bottlenecks, making it difficult to meet the increasingly complex management needs under this model. The S2B2C model suffers from insufficient data collaboration and forecasting accuracy. In existing supply chain management technologies, data between suppliers, distributors, and consumers is often fragmented, forming "information silos." Upstream suppliers typically produce and stock inventory based solely on historical orders from B-end customers, lacking the ability to perceive real-time demand fluctuations from C-end customers (such as browsing history and search activity); simultaneously, sales plans from B-end customers (such as pre-promotion information) are not effectively transmitted upstream. This information asymmetry makes it difficult for traditional forecasting models to capture true market demand, easily triggering the "bullwhip effect," resulting in inventory backlogs or stockouts. Furthermore, there is a lack of refined scheduling mechanisms tailored to the differentiated characteristics of multiple product categories. Existing technologies typically employ standardized scheduling strategies, neglecting the significant differences in attributes among different product categories. In practice, fresh produce has extremely high timeliness requirements, bulk commodities are cost-sensitive, and special categories such as chemicals have strict restrictions on mixed loading. Traditional scheduling models struggle to simultaneously handle restrictions on mixed loading of multiple categories, differentiated timeliness requirements, and the balancing of multiple objectives, resulting in high transportation costs and a high risk of product damage or violations. There is a contradiction between the static nature of scheduling schemes and the uncertainties of the supply chain. Traditional scheduling systems are mostly based on static plan generation, lacking real-time monitoring and dynamic response capabilities. When unexpected situations occur during supply chain execution (such as logistics delays, demand surges, or abnormal environmental parameters), existing technologies cannot promptly detect and trigger rescheduling mechanisms, often leading to decreased fulfillment rates and an inability to automatically generate effective contingency plans. Therefore, there is an urgent need for a supply chain scheduling optimization method that can break down data barriers between the S, B, and C ends, support differentiated scheduling of multiple product categories, and possess real-time dynamic response and intelligent solution capabilities. Summary of the Invention

[0003] A method for dynamic scheduling and optimization of multi-category supply chains based on S2B2C includes the following steps; S1. Multi-source data collection and multi-category profile construction: Collect multi-source data, including C-end data, B-end data, and S-end data, and construct multi-category profiles, setting differentiated weight coefficients and scheduling priorities for different categories (e.g., fresh produce prioritizes timeliness, and bulk commodities prioritize cost). S2. Collaborative Demand Forecasting Based on S2B2C: Utilizing the ARIMA time series model combined with machine learning algorithms, the demand for each product category by consumers (C-end) in different regions and time periods is predicted. The sales plan factor of the business end is introduced to correct the prediction deviation of the C-end (e.g., the pre-heating of business end promotional activities). Dynamic setting of safety stock: Based on the prediction error and the service level agreement (SLA), the safety stock threshold of each warehousing node is dynamically calculated. For the characteristics of multiple product categories, the ABC classification management strategy is implemented, and a more sensitive inventory early warning mechanism is set for high-value and high-turnover product categories. S3. Dynamic scheduling model construction: Construct a multi-objective optimization model to balance cost, timeliness and service quality, set the objective function, and conduct constraint analysis. Consider the synergistic effects of different categories in transportation and warehousing (such as carpooling rate and shared storage), and transform the multi-category scheduling problem into a vehicle routing problem with time windows (VRPTW) or a multi-level inventory control problem. S4. Intelligent solution and scheduling scheme generation: The genetic algorithm (GA) is adopted and a reinforcement learning mechanism is introduced. The agent is trained using historical data, enabling it to make scheduling decisions quickly in a real-time environment and generate procurement scheduling, inventory transfer and logistics dispatch instructions. S5. Real-time monitoring and anomaly response mechanism: To address uncertainties in the supply chain and achieve dynamic characteristics, it monitors logistics trajectories and warehouse temperature and humidity (for special product categories) through IoT devices, monitors sudden surges or drops in demand (such as the "hot-selling" phenomenon), and dynamically reschedules: It sets trigger mechanisms (such as delays exceeding X hours, inventory falling below the warning line), and when an anomaly occurs, it triggers local or global recalculation and automatically generates contingency plans (such as emergency replenishment, switching logistics providers, and canceling low-priority orders). S6. Performance Evaluation and Closed-Loop Optimization: Form a PDCA closed loop to continuously optimize the scheduling strategy and evaluate the scheduling performance in multiple dimensions. Evaluation indicators include order fulfillment on-time rate, inventory turnover days, logistics cost ratio, B-end satisfaction, and C-end positive review rate. Feed the actual execution data of this scheduling back to the data pool of step S1. Use comparative analysis to correct the parameters of the demand forecasting model, optimize the weight configuration of the scheduling algorithm, and improve the decision accuracy of the next cycle.

[0004] Furthermore, a dynamic scheduling and optimization method for multi-category supply chains based on S2B2C is proposed. In step S1, data from C-end, B-end, and S-end are accessed, and multi-category profiles are constructed based on product attributes. This specifically includes: C-end data: Collect consumer browsing history, historical orders, search keywords, and evaluation feedback to identify real-time demand fluctuations; B2B data: Collecting information on distributors' / retailers' inventory levels, sales rates, financial status, and fulfillment capabilities; S-side data: Data collected from upstream suppliers' production capacity, raw material inventory, logistics partners' transportation capacity, and warehouse node status. Multi-category profile construction: Classify products according to their attributes (such as shelf life, volume, value density, turnover rate) (such as high-frequency standardized products, long-tail non-standardized products, perishable fresh products, etc.), and set differentiated weight coefficients and scheduling priorities for different categories (for example, fresh products are given priority in terms of timeliness, and bulk commodities are given priority in terms of cost).

[0005] Furthermore, a dynamic scheduling and optimization method for multi-category supply chains based on S2B2C is proposed. In step S2, time series models combined with machine learning algorithms are used to predict the demand for each product category by consumers (C-end) in different regions and time periods. A sales plan factor from the business side (B-end) is introduced to correct prediction biases, and a safety stock threshold is dynamically set. The specific steps are as follows: S21. Data Preprocessing and Feature Engineering: Clean the collected historical sales data from the C-end: remove outliers, smooth the data, extract time features (week, month, quarter, holidays), spatial features (regional climate, population density) and product features, structure the sales plan data uploaded from the B-end, and extract promotion types (discounts, flash sales, live streaming), discount intensity, preheating duration and expected target audience tags. S22. Constructing a Hybrid Forecasting Model: A dual-stream fusion architecture is adopted for basic demand forecasting. For each product category, the ARIMA model is used to capture the linear trend and periodic fluctuations (such as seasonal fluctuations) of historical sales, outputting a basic forecast value Yb. Using the XGBoost algorithm, external features (weather, competitor prices, macroeconomic indicators) and C-end behavioral features (browse conversion rate, search popularity) are input to predict the demand fluctuation correction value Ye. Through a weighted fusion strategy, the natural C-end demand forecast value is obtained. ,in α This represents the proportion of the time series model in the final prediction result; S23. Introducing deviation correction for B-end sales plan factors: Constructing a promotion elasticity coefficient model to... To address the issue of forecasts lagging behind planned activities, adjustments are made: Planning factor coefficients are calculated, and key factors are extracted based on the sales plan uploaded by the B-end: discount rate, marketing channel weight, and historical conversion rate of similar activities. An additive or multiplicative model is used to adjust the demand. During the pre-sale period, the final collaborative demand forecast is calculated using a formula. ,in βConfidence weights for B-end plans (dynamically adjusted based on historical B-end plan execution rates). For incremental demand; S24. Forecast Error Analysis and Dynamic Safety Stock Setting: Based on the corrected forecast results, calculate the safety stock for each storage node: Calculate the standard deviation of the forecast error: Use a sliding window to statistically analyze the deviation between recent forecast values ​​and actual sales, and calculate the standard deviation. σ e ; Determine the service level coefficient: Based on the product profile and SLA agreement constructed in S1, query the Z-value of the normal distribution (e.g., for fresh produce, which requires high timeliness, set Z=2.33 to correspond to a 99% service level; for bulk commodities, set Z=1.65 to correspond to a 95% service level). Dynamic threshold calculation: combining replenishment lead time L and its variance. Substitute into the dynamic safety stock formula: ,in This represents the average daily demand. S25. ABC Classification Inventory Warning Configuration: Based on the classification results in S1, configure differentiated inventory warning strategies: Category A: Establish a high-frequency inventory counting mechanism; when inventory falls below S... A red alert is immediately triggered when the value reaches 1.2 times the normal value, and the S3 scheduling model is activated to prioritize processing. Category B / C: Appropriately relax the warning threshold to S 0.8 times, to reduce inventory holding costs and avoid capital accumulation.

[0006] Furthermore, a method for dynamic scheduling and optimization of multi-category supply chains based on S2B2C is characterized by, In step S3, a multi-objective optimization model is constructed, and the objective function and constraints are set, as follows; Setting the objective function: Balancing conflicting objectives through weighted scoring; cost control dimensions: Total costs include: Procurement costs: the principal amount paid to purchase goods from suppliers; Warehousing costs: rent and management fees incurred in storing goods in warehouses; Transportation costs: fuel, tolls, and driver costs for vehicle delivery; and Stockout penalty costs: compensation losses caused by stockouts. Service levels include: satisfaction rate and on-time rate; The system automatically assigns weights based on different product categories (such as fresh produce which emphasizes timeliness, and bulk commodities which emphasize cost), and calculates a comprehensive score for optimization. Constraints include hard constraints and soft constraints: Set hard constraints: Supplier supply limit: The order quantity cannot exceed the supplier's current production capacity or inventory. Warehouse capacity limit: The total volume of goods entering the warehouse cannot exceed the remaining storage space in the warehouse; Vehicle carrying capacity limits: The total weight of the loaded goods shall not exceed the truck's rated load capacity, and the total volume shall not exceed the truck's cargo compartment volume; Multi-category mixed loading prohibition: enforce product mutual exclusion rules, for example: if the system detects that there are chemicals and food in the same vehicle, it will directly determine that the plan is illegal and they must be transported in separate vehicles; Set soft constraints: delivery time window for C-end customers, designated logistics provider for B-end customers; Transforming the above logic into a computer-processable structure: the delivery task is transformed into a vehicle routing problem with time windows, namely: multiple vehicles, multiple delivery points, multiple time windows, and mixed loading restrictions, to find the shortest path and the lowest cost.

[0007] Furthermore, a dynamic scheduling and optimization method for multi-category supply chains based on S2B2C is proposed. In step S4, a genetic algorithm combined with reinforcement learning is used to solve the model and generate procurement scheduling, inventory transfer, and logistics dispatch instructions. The specific steps are as follows: S41. Chromosome Encoding and Intelligent Initialization: This transforms complex business scheduling schemes into computer-processable gene sequences using a multi-dimensional encoding strategy. It employs segmented chromosome encoding, where one chromosome represents a complete scheduling scheme, containing three segments. Procurement gene segment: The numerical sequence represents the order quantity placed with each supplier; Inventory gene segment: The numerical sequence represents the transfer flow between warehouses; Logistics gene segment: The numerical sequence represents the vehicle route and delivery order; Reinforcement learning-assisted initialization: Unlike traditional random generation of initial solutions, this method utilizes a pre-trained reinforcement learning agent to quickly generate a set of better initial solutions as a seed population based on current real-time data (such as inventory shortages or sudden congestion), thus shortening the algorithm's search time. S42. Fitness Function Construction and Evaluation: Calculate the merits of each chromosome (scheduling plan) and substitute it into the objective function: decode the chromosome back into the business plan, calculate the total cost (procurement + warehousing + transportation + penalty) and service level defined in step S3, and calculate the fitness: fitness value = service level score - normalized total cost. The higher the score, the better the plan. S43. Genetic Operations and Iterative Evolution: Simulating the biological evolution process to find the global optimum. Selection operation: Keep the solutions with high fitness (survival of the fittest), and eliminate the solutions with poor fitness; Cross-operation: Interchange parts of the genes of two excellent solutions (e.g., combine the procurement plan of solution A with the logistics route of solution B) to generate new solutions and explore better combinations; Mutation operation: Randomly change some details of the solution (e.g., randomly change a supplier or change the route of a car) to prevent the algorithm from getting stuck in a local optimum dead end; S44. Local Search Guided by Reinforcement Learning Mechanism: During the iteration process of the genetic algorithm, a reinforcement learning agent is introduced to intervene. The agent perceives the current environmental state in real time (such as a route violating the time window constraint or a commodity having extremely low inventory). Based on the trained policy network, the agent makes targeted corrections to the individuals in the genetic algorithm. For example, if the agent finds that a vehicle is overloaded, it directly issues an action command to adjust the loading order instead of waiting for a long period of random mutation to solve the problem. When the algorithm is close to the optimal solution but the oscillation does not converge, the reinforcement learning agent directly locks the key parameters based on historical experience and quickly outputs the final result. S45. Termination Decision and Instruction Decoding Output: Determine if the maximum number of iterations and fitness improvement threshold have been reached; when the optimal individual with the highest fitness is output, perform a decoding operation on the optimal individual and map it to a specific business execution instruction: Procurement scheduling instructions: Determine the order quantity and order time for each supplier; Inventory transfer instructions: Determine the transfer routes and quantities of goods between various storage nodes; Logistics dispatch instructions: determine vehicle allocation, delivery route sequence, and carrier selection.

[0008] Furthermore, a dynamic scheduling and optimization method for multi-category supply chains based on S2B2C is proposed. Step S5 involves monitoring the supply chain execution status and setting triggering and exception response mechanisms, specifically including: Real-time collection of logistics trajectory data and warehouse environment parameters via IoT devices: Logistics trajectory monitoring: Using GPS positioning modules and RFID tags, the real-time location, speed and stationary status of transport vehicles are collected at a preset frequency (e.g. every 30 seconds). The system compares the real-time trajectory with the planned route generated in step S4 and calculates the estimated time of arrival (ETA). Warehouse environment monitoring: For special categories built in S1: fresh produce, pharmaceuticals, and chemicals, temperature and humidity sensors and gas concentration sensors are deployed in the warehouse nodes and transport compartments. The sensors upload environmental parameters in real time, and the system determines whether they are within the preset compliance range (e.g., the temperature in the fresh produce compartment needs to be maintained at 0-4℃). Once the environmental parameters exceed the limit, an alarm is triggered immediately. Real-time monitoring of demand fluctuations, identifying sudden surges or drops in demand. Fluctuation identification: Calculate the deviation ratio between the actual demand in the current period and the predicted demand in step S2. If the deviation ratio exceeds the preset abnormal threshold (e.g., actual demand suddenly increases to 200% of the predicted value, i.e., the "hot item" phenomenon; or suddenly drops to 10%), the system automatically determines it as an abnormal demand and marks the status as a "high priority event". Dynamic triggering mechanism and rescheduling judgment: The system has a preset abnormal trigger threshold library and monitors in real time whether key indicators have reached the threshold. Time-sensitive trigger: When logistics tracking shows that the vehicle delay time exceeds a preset threshold (e.g., exceeding the planned arrival time by 2 hours), it is judged as a serious delay; Inventory trigger: When the real-time inventory level of a certain warehouse node falls below the dynamic safety inventory warning line (calculated in step S2), it is determined to be a stockout risk; Environmental trigger: When the environmental parameters (such as temperature) of a special category continue to be abnormal for more than a preset time, it is judged as a risk of cargo damage. Once any of the above trigger conditions are met, the system will suspend the regular scheduling logic and start the dynamic rescheduling engine. Emergency plans are automatically generated and executed, including emergency replenishment, switching logistics providers, and canceling low-priority orders.

[0009] The beneficial effects of this invention are as follows: By introducing a "B-end sales plan factor" to correct C-end demand forecasts, the problem of forecast lag caused by relying solely on historical data is effectively solved. This collaborative forecasting mechanism not only reduces inventory fluctuations caused by the "bullwhip effect" but also enables upstream suppliers to respond to market changes in advance, significantly improving the overall supply and demand matching degree and response speed of the supply chain. A full-link real-time monitoring and anomaly response mechanism has been established. Utilizing IoT technology, logistics trajectories and warehousing environments are perceived in real time, and a trigger mechanism is set to automatically initiate dynamic rescheduling. This mechanism changes the passive situation of the traditional "static execution" scheduling scheme, enabling real-time response to sudden situations such as demand surges and logistics delays, automatically generating emergency replenishment or capacity switching contingency plans, effectively ensuring order fulfillment rates and improving the supply chain's resilience. A solution strategy combining "genetic algorithm and reinforcement learning" is adopted. The introduction of the reinforcement learning mechanism effectively solves the problems of slow convergence speed and easy getting trapped in local optima when traditional genetic algorithms solve large-scale VRPTW (vehicle routing problem with time windows). This method can quickly generate high-quality procurement, inventory, and logistics scheduling instructions, meeting the stringent requirements of real-time decision-making in actual business scenarios. By constructing multi-category profiles and setting differentiated weighting coefficients and scheduling priorities for different categories (such as prioritizing timeliness for fresh produce and cost for bulk commodities), constraints such as restrictions on mixed packaging of multiple categories were introduced. This refined scheduling strategy not only ensures the service quality of high-value, perishable goods but also effectively controls the overall cost of long-tail goods, significantly reducing spoilage rates and the risk of violations. Attached Figure Description

[0010] Figure 1 A flowchart of a multi-category supply chain dynamic scheduling and optimization method based on S2B2C; Detailed Implementation A method for dynamic scheduling and optimization of multi-category supply chains based on S2B2C, such as Figure 1 As shown, it includes the following steps; S1. Multi-source data collection and multi-category profile construction: Collect multi-source data, including C-end data, B-end data, and S-end data, and construct multi-category profiles, setting differentiated weight coefficients and scheduling priorities for different categories (e.g., fresh produce prioritizes timeliness, and bulk commodities prioritize cost). S2. Collaborative Demand Forecasting Based on S2B2C: Utilizing the ARIMA time series model combined with machine learning algorithms, the demand for each product category by consumers (C-end) in different regions and time periods is predicted. The sales plan factor of the business end is introduced to correct the prediction deviation of the C-end (e.g., the pre-heating of business end promotional activities). Dynamic setting of safety stock: Based on the prediction error and the service level agreement (SLA), the safety stock threshold of each warehousing node is dynamically calculated. For the characteristics of multiple product categories, the ABC classification management strategy is implemented, and a more sensitive inventory early warning mechanism is set for high-value and high-turnover product categories. S3. Dynamic scheduling model construction: Construct a multi-objective optimization model to balance cost, timeliness and service quality, set the objective function, and conduct constraint analysis. Consider the synergistic effects of different categories in transportation and warehousing (such as carpooling rate and shared storage), and transform the multi-category scheduling problem into a vehicle routing problem with time windows (VRPTW) or a multi-level inventory control problem. S4. Intelligent solution and scheduling scheme generation: The genetic algorithm (GA) is adopted and a reinforcement learning mechanism is introduced. The agent is trained using historical data, enabling it to make scheduling decisions quickly in a real-time environment and generate procurement scheduling, inventory transfer and logistics dispatch instructions. S5. Real-time monitoring and anomaly response mechanism: To address uncertainties in the supply chain and achieve dynamic characteristics, it monitors logistics trajectories and warehouse temperature and humidity (for special product categories) through IoT devices, monitors sudden surges or drops in demand (such as the "hot-selling" phenomenon), and dynamically reschedules: It sets trigger mechanisms (such as delays exceeding X hours, inventory falling below the warning line), and when an anomaly occurs, it triggers local or global recalculation and automatically generates contingency plans (such as emergency replenishment, switching logistics providers, and canceling low-priority orders). S6. Performance Evaluation and Closed-Loop Optimization: Form a PDCA closed loop to continuously optimize the scheduling strategy and evaluate the scheduling performance in multiple dimensions. Evaluation indicators include order fulfillment on-time rate, inventory turnover days, logistics cost ratio, B-end satisfaction, and C-end positive review rate. Feed the actual execution data of this scheduling back to the data pool of step S1. Use comparative analysis to correct the parameters of the demand forecasting model, optimize the weight configuration of the scheduling algorithm, and improve the decision accuracy of the next cycle.

[0011] Furthermore, a dynamic scheduling and optimization method for multi-category supply chains based on S2B2C is proposed. In step S1, data from C-end, B-end, and S-end are accessed, and multi-category profiles are constructed based on product attributes. This specifically includes: C-end data: Collect consumer browsing history, historical orders, search keywords, and evaluation feedback to identify real-time demand fluctuations; B2B data: Collecting information on distributors' / retailers' inventory levels, sales rates, financial status, and fulfillment capabilities; S-side data: Data collected from upstream suppliers' production capacity, raw material inventory, logistics partners' transportation capacity, and warehouse node status. Multi-category profile construction: Classify products according to their attributes (such as shelf life, volume, value density, turnover rate) (such as high-frequency standardized products, long-tail non-standardized products, perishable fresh products, etc.), and set differentiated weight coefficients and scheduling priorities for different categories (for example, fresh products are given priority in terms of timeliness, and bulk commodities are given priority in terms of cost).

[0012] Furthermore, a dynamic scheduling and optimization method for multi-category supply chains based on S2B2C is proposed. In step S2, time series models combined with machine learning algorithms are used to predict the demand for each product category by consumers (C-end) in different regions and time periods. A sales plan factor from the business side (B-end) is introduced to correct prediction biases, and a safety stock threshold is dynamically set. The specific steps are as follows: S21. Data Preprocessing and Feature Engineering: Clean the collected historical sales data from the C-end: remove outliers, smooth the data, extract time features (week, month, quarter, holidays), spatial features (regional climate, population density) and product features, structure the sales plan data uploaded from the B-end, and extract promotion types (discounts, flash sales, live streaming), discount intensity, preheating duration and expected target audience tags. S22. Constructing a Hybrid Forecasting Model: A dual-stream fusion architecture is adopted for basic demand forecasting. For each product category, the ARIMA model is used to capture the linear trend and periodic fluctuations (such as seasonal fluctuations) of historical sales, outputting a basic forecast value Yb. Using the XGBoost algorithm, external features (weather, competitor prices, macroeconomic indicators) and C-end behavioral features (browse conversion rate, search popularity) are input to predict the demand fluctuation correction value Ye. Through a weighted fusion strategy, the natural C-end demand forecast value is obtained. ,in α This represents the proportion of the time series model in the final prediction result; S23. Introducing deviation correction for B-end sales plan factors: Constructing a promotion elasticity coefficient model to... To address the issue of forecasts lagging behind planned activities, adjustments are made: Planning factor coefficients are calculated, and key factors are extracted based on the sales plan uploaded by the B-end: discount rate, marketing channel weight, and historical conversion rate of similar activities. An additive or multiplicative model is used to adjust the demand. During the pre-sale period, the final collaborative demand forecast is calculated using a formula. ,in β Confidence weights for B-end plans (dynamically adjusted based on historical B-end plan execution rates). For incremental demand; S24. Forecast Error Analysis and Dynamic Safety Stock Setting: Based on the corrected forecast results, calculate the safety stock for each storage node: Calculate the standard deviation of the forecast error: Use a sliding window to statistically analyze the deviation between recent forecast values ​​and actual sales, and calculate the standard deviation. σ e ; Determine the service level coefficient: Based on the product profile and SLA agreement constructed in S1, query the Z-value of the normal distribution (e.g., for fresh produce, which requires high timeliness, set Z=2.33 to correspond to a 99% service level; for bulk commodities, set Z=1.65 to correspond to a 95% service level). Dynamic threshold calculation: combining replenishment lead time L and its variance. Substitute into the dynamic safety stock formula: ,in This represents the average daily demand. S25. ABC Classification Inventory Warning Configuration: Based on the classification results in S1, configure differentiated inventory warning strategies: Category A: Establish a high-frequency inventory counting mechanism; when inventory falls below S... A red alert is immediately triggered when the value reaches 1.2 times the normal value, and the S3 scheduling model is activated to prioritize processing. Category B / C: Appropriately relax the warning threshold to S 0.8 times, to reduce inventory holding costs and avoid capital accumulation.

[0013] Furthermore, a method for dynamic scheduling and optimization of multi-category supply chains based on S2B2C is characterized by, In step S3, a multi-objective optimization model is constructed, and the objective function and constraints are set, as follows; Setting the objective function: Balancing conflicting objectives through weighted scoring; cost control dimensions: Total costs include: Procurement costs: the principal amount paid to purchase goods from suppliers; Warehousing costs: rent and management fees incurred in storing goods in warehouses; Transportation costs: fuel, tolls, and driver costs for vehicle delivery; and Stockout penalty costs: compensation losses caused by stockouts. Service levels include: satisfaction rate and on-time rate; The system automatically assigns weights based on different product categories (such as fresh produce which emphasizes timeliness, and bulk commodities which emphasize cost), and calculates a comprehensive score for optimization. Constraints include hard constraints and soft constraints: Set hard constraints: Supplier supply limit: The order quantity cannot exceed the supplier's current production capacity or inventory. Warehouse capacity limit: The total volume of goods entering the warehouse cannot exceed the remaining storage space in the warehouse; Vehicle carrying capacity limits: The total weight of the loaded goods shall not exceed the truck's rated load capacity, and the total volume shall not exceed the truck's cargo compartment volume; Multi-category mixed loading prohibition: enforce product mutual exclusion rules, for example: if the system detects that there are chemicals and food in the same vehicle, it will directly determine that the plan is illegal and they must be transported in separate vehicles; Set soft constraints: delivery time window for C-end customers, designated logistics provider for B-end customers; Transforming the above logic into a computer-processable structure: the delivery task is transformed into a vehicle routing problem with time windows, namely: multiple vehicles, multiple delivery points, multiple time windows, and mixed loading restrictions, to find the shortest path and the lowest cost.

[0014] Furthermore, a dynamic scheduling and optimization method for multi-category supply chains based on S2B2C is proposed. In step S4, a genetic algorithm combined with reinforcement learning is used to solve the model and generate procurement scheduling, inventory transfer, and logistics dispatch instructions. The specific steps are as follows: S41. Chromosome Encoding and Intelligent Initialization: This transforms complex business scheduling schemes into computer-processable gene sequences using a multi-dimensional encoding strategy. It employs segmented chromosome encoding, where one chromosome represents a complete scheduling scheme, containing three segments. Procurement gene segment: The numerical sequence represents the order quantity placed with each supplier; Inventory gene segment: The numerical sequence represents the transfer flow between warehouses; Logistics gene segment: The numerical sequence represents the vehicle route and delivery order; Reinforcement learning-assisted initialization: Unlike traditional random generation of initial solutions, this method utilizes a pre-trained reinforcement learning agent to quickly generate a set of better initial solutions as a seed population based on current real-time data (such as inventory shortages or sudden congestion), thus shortening the algorithm's search time. S42. Fitness Function Construction and Evaluation: Calculate the merits of each chromosome (scheduling plan) and substitute it into the objective function: decode the chromosome back into the business plan, calculate the total cost (procurement + warehousing + transportation + penalty) and service level defined in step S3, and calculate the fitness: fitness value = service level score - normalized total cost. The higher the score, the better the plan. S43. Genetic Operations and Iterative Evolution: Simulating the biological evolution process to find the global optimum. Selection operation: Keep the solutions with high fitness (survival of the fittest), and eliminate the solutions with poor fitness; Cross-operation: Interchange parts of the genes of two excellent solutions (e.g., combine the procurement plan of solution A with the logistics route of solution B) to generate new solutions and explore better combinations; Mutation operation: Randomly change some details of the solution (e.g., randomly change a supplier or change the route of a car) to prevent the algorithm from getting stuck in a local optimum dead end; S44. Local Search Guided by Reinforcement Learning Mechanism: During the iteration process of the genetic algorithm, a reinforcement learning agent is introduced to intervene. The agent perceives the current environmental state in real time (such as a route violating the time window constraint or a commodity having extremely low inventory). Based on the trained policy network, the agent makes targeted corrections to the individuals in the genetic algorithm. For example, if the agent finds that a vehicle is overloaded, it directly issues an action command to adjust the loading order instead of waiting for a long period of random mutation to solve the problem. When the algorithm is close to the optimal solution but the oscillation does not converge, the reinforcement learning agent directly locks the key parameters based on historical experience and quickly outputs the final result. S45. Termination Decision and Instruction Decoding Output: Determine if the maximum number of iterations and fitness improvement threshold have been reached; when the optimal individual with the highest fitness is output, perform a decoding operation on the optimal individual and map it to a specific business execution instruction: Procurement scheduling instructions: Determine the order quantity and order time for each supplier; Inventory transfer instructions: Determine the transfer routes and quantities of goods between various storage nodes; Logistics dispatch instructions: determine vehicle allocation, delivery route sequence, and carrier selection.

[0015] Furthermore, a dynamic scheduling and optimization method for multi-category supply chains based on S2B2C is proposed. Step S5 involves monitoring the supply chain execution status and setting triggering and exception response mechanisms, specifically including: Real-time collection of logistics trajectory data and warehouse environment parameters via IoT devices: Logistics trajectory monitoring: Using GPS positioning modules and RFID tags, the real-time location, speed and stationary status of transport vehicles are collected at a preset frequency (e.g. every 30 seconds). The system compares the real-time trajectory with the planned route generated in step S4 and calculates the estimated time of arrival (ETA). Warehouse environment monitoring: For special categories built in S1: fresh produce, pharmaceuticals, and chemicals, temperature and humidity sensors and gas concentration sensors are deployed in the warehouse nodes and transport compartments. The sensors upload environmental parameters in real time, and the system determines whether they are within the preset compliance range (e.g., the temperature in the fresh produce compartment needs to be maintained at 0-4℃). Once the environmental parameters exceed the limit, an alarm is triggered immediately. Real-time monitoring of demand fluctuations, identifying sudden surges or drops in demand. Fluctuation identification: Calculate the deviation ratio between the actual demand in the current period and the predicted demand in step S2. If the deviation ratio exceeds the preset abnormal threshold (e.g., actual demand suddenly increases to 200% of the predicted value, i.e., the "hot item" phenomenon; or suddenly drops to 10%), the system automatically determines it as an abnormal demand and marks the status as a "high priority event". Dynamic triggering mechanism and rescheduling judgment: The system has a preset abnormal trigger threshold library and monitors in real time whether key indicators have reached the threshold. Time-sensitive trigger: When logistics tracking shows that the vehicle delay time exceeds a preset threshold (e.g., exceeding the planned arrival time by 2 hours), it is judged as a serious delay; Inventory trigger: When the real-time inventory level of a certain warehouse node falls below the dynamic safety inventory warning line (calculated in step S2), it is determined to be a stockout risk; Environmental trigger: When the environmental parameters (such as temperature) of a special category continue to be abnormal for more than a preset time, it is judged as a risk of cargo damage. Once any of the above trigger conditions are met, the system will suspend the regular scheduling logic and start the dynamic rescheduling engine. Emergency plans are automatically generated and executed, including emergency replenishment, switching logistics providers, and canceling low-priority orders.

[0016] Example 2 This embodiment provides a dynamic scheduling and optimization method for a multi-category supply chain based on S2B2C. This method is applied to a supply chain management platform and includes the following specific steps: S1. Multi-source data acquisition and multi-category profile construction: The system builds a multi-source data acquisition interface to access data from three terminals in real time. Consumer data collection: The system uses tracking technology to collect consumers' browsing history, order records, search keywords, and feedback data on the app and mini-program. For example, the system detected that the search popularity for "cherries" in a certain region increased by 200% in the past 24 hours. B-end data collection: The system connects to the distributor's ERP system to collect its inventory level, sales rate, financial status and fulfillment capability. For example, it can obtain information that a distributor's warehouse currently has only 100 items left in stock. S-end data collection: The system connects with upstream supplier systems and logistics TMS systems to collect supplier capacity, raw material inventory, logistics partner transportation capacity and warehouse node status; Multi-category profiling: The system categorizes products based on their attributes (shelf life, volume, value density, turnover rate); Perishable fresh produce, such as cherries and seafood, is categorized as "high-priority" by the system, with a high scheduling weight coefficient to prioritize transportation speed and require mandatory binding to refrigerated vehicles. Bulk standardized products, such as toilet paper and laundry detergent, are classified as "cost priority" and have a lower scheduling weight coefficient, prioritizing carpooling rate and warehousing costs. Chemicals: Classified as "Special Control Category", with rules for mixed loading restrictions, and prohibited from being transported together with food products; S2. Collaborative Demand Forecasting Based on S2B2C: The system uses a time series model (LSTM) combined with the XGBoost machine learning algorithm to predict the demand for each product category in different regions over the next 7 days. The system introduces a B-end sales plan factor for correction. For example, a large supermarket plans to hold an "anniversary promotion" next Monday. After the system captures the plan factor, it automatically adjusts the predicted demand for the corresponding product category in that region, correcting the bias caused by relying solely on historical data. Dynamic inventory setting: Based on forecast error and service level agreement (SLA), the system dynamically calculates the safety stock threshold for each warehousing node and implements ABC classification management for multi-category characteristics: For category A high-value, high-turnover goods, the system sets the inventory warning sensitivity to "an alarm will be triggered when the inventory level is lower than 3 days' sales"; for category C long-tail goods, the warning threshold is relaxed to "an alarm will be triggered when the inventory level is lower than 0.5 days' sales". S3. Dynamic scheduling model construction: The system constructs a multi-objective optimization model, with the objective function set as a weighted combination of minimizing total cost and maximizing service level; Constraint Analysis: The model sets hard constraints, such as supplier capacity limits, warehouse capacity limits, and vehicle load limits. In particular, for multi-category characteristics, the model embeds a mixed loading constraint matrix. For example, when the same vehicle is loaded with both "perfume (chemicals)" and "cake (food)" in the scheduling plan, the plan is judged to violate the hard constraints and is directly eliminated. Problem Transformation: The system transforms the multi-category scheduling problem into a hybrid model of the vehicle routing problem with time windows (VRPTW) and the multi-level inventory control problem; S4. Intelligent Solution and Scheduling Scheme Generation: The system uses a genetic algorithm (GA) combined with reinforcement learning (RL) to solve the above model. First, the system uses a reinforcement learning agent to generate a set of relatively good initial populations (scheduling schemes) based on historical experience. Then, the genetic algorithm performs selection, crossover, and mutation operations on the population. During the iteration process, the reinforcement learning agent monitors the evolution direction in real time and makes local corrections to individuals that violate soft constraints (such as slight timeouts) to accelerate convergence. Generate instructions: The model outputs the optimal solution; specific instructions for generation are: Procurement scheduling: Place an order for 5,000 pieces of fresh produce with supplier A, requiring delivery by Wednesday; Inventory transfer: Generate an instruction to transfer 200 surplus standard products from the North China warehouse to the East China warehouse, which is facing a shortage of inventory. Logistics dispatch: Automatically assign cold chain vehicle V01 to fresh food orders and plan the optimal delivery route; assign ordinary truck V02 to standard product orders and plan carpooling routes; S5. Real-time monitoring and anomaly response mechanism: During execution, the system performs real-time monitoring through IoT devices; Monitoring scenario: The system detected that the temperature inside the cargo compartment of transport vehicle V01 rose from 4°C to 10°C through temperature and humidity sensors, exceeding the compliant range for fresh produce. Dynamic rescheduling: When an environmental anomaly is triggered, the system automatically suspends the current task of the vehicle, recalculates the plan, and generates an emergency plan: immediately notifies the nearest backup warehouse to transfer the vehicle, marks the original vehicle as "under maintenance", and automatically sends a delay notification SMS to the customer; S6. Performance Evaluation and Closed-Loop Optimization: After the scheduling task is completed, the system performs a multi-dimensional evaluation of the scheduling, calculates indicators such as order fulfillment on-time rate (e.g., 98%), inventory turnover days (e.g., 5 days), and logistics cost ratio. The system feeds back the actual execution data (e.g., the deviation between actual demand and forecast) to the data pool of S1, uses comparative analysis to automatically correct the parameters in the demand forecasting model, optimizes the scheduling weight configuration for the next cycle, and achieves PDCA closed-loop optimization.

Claims

1. A method for dynamic scheduling and optimization of multi-category supply chains based on S2B2C, characterized in that, Includes the following steps: S1. Multi-source data collection and multi-category profile construction: Collect multi-source data, access C-end data, B-end data and S-end data, classify according to product attributes and build multi-category profiles, and set differentiated weight coefficients and scheduling priorities for different categories; S2. Collaborative demand forecasting based on S2B2C: Using time series models combined with machine learning algorithms to predict the demand of C-end for each category in different regions and time periods, introducing B-end sales plan factors to correct forecasting deviations, and dynamically setting safety stock thresholds. For multi-category characteristics, an ABC classification management strategy is implemented, and a more sensitive inventory early warning mechanism is set for high-value and high-turnover categories. S3. Dynamic scheduling model construction: Construct a multi-objective optimization model, set the objective function and constraints, and transform the multi-category scheduling problem into a multi-level inventory control problem; S4. Intelligent solution and scheduling scheme generation: The model is solved by combining genetic algorithm with reinforcement learning mechanism to generate procurement scheduling, inventory transfer and logistics dispatch instructions; S5. Real-time monitoring and anomaly response mechanism: Monitor the execution status of the supply chain, set a trigger mechanism, and perform dynamic rescheduling when an anomaly is triggered; S6. Performance Evaluation and Closed-Loop Optimization: Evaluate scheduling performance from multiple dimensions and feed execution data back to the data pool to correct model parameters.

2. The method for dynamic scheduling and optimization of a multi-category supply chain based on S2B2C as described in claim 1, characterized in that, In step S1, data from C-end, B-end, and S-end are accessed, and multi-category profiles are constructed based on product attributes, specifically including: Consumer-side data includes browsing history, historical orders, search keywords, and feedback, which are used to identify real-time fluctuations in demand. B2B data includes distributors' / retailers' inventory levels, sales rates, financial status, and fulfillment capabilities. S-side data includes the platform's upstream supplier capacity, raw material inventory, logistics partner capacity, and warehouse node status; The multi-category profiling system categorizes products into high-frequency standardized products, long-tail non-standardized products, and perishable fresh products based on their shelf life, volume, value density, and turnover rate attributes, and sets corresponding differentiated scheduling priorities.

3. The method for dynamic scheduling and optimization of a multi-category supply chain based on S2B2C as described in claim 1, characterized in that, In step S2, time series models combined with machine learning algorithms are used to predict the demand for each product category by consumers (C-end) in different regions and time periods. A sales plan factor from the business side (B-end) is introduced to correct prediction biases, and a safety stock threshold is dynamically set. The specific steps are as follows: S21. Data Preprocessing and Feature Engineering: Clean the collected historical sales data from C-end users: remove outliers, smooth the data, and extract time features, spatial features, and product features. Structure the sales plan data uploaded from B-end users and extract promotion type, discount level, pre-heating duration, and expected target audience tags. S22. Constructing a Hybrid Forecasting Model: A dual-stream fusion architecture is adopted for basic demand forecasting. For each product category, the ARIMA model is used to capture the linear trend and periodic fluctuations of historical sales, outputting a basic forecast value Yb. The XGBoost algorithm is used to predict demand fluctuation correction values ​​Ye, taking into account external features and C-end behavioral features. Through a weighted fusion strategy, the natural demand forecast value for C-end customers is obtained. ,in α This represents the proportion of the time series model in the final prediction result; S23. Introducing deviation correction for B-end sales plan factors: Constructing a promotion elasticity coefficient model to... To address the issue of forecasts lagging behind planned activities, adjustments are made: Planning factor coefficients are calculated, and key factors are extracted based on the sales plan uploaded by the B-end: discount rate, marketing channel weight, and historical conversion rate of similar activities. An additive or multiplicative model is used to adjust the demand. During the pre-sale period, the final collaborative demand forecast is calculated using a formula. ,in β For B-end plans, confidence weights, For incremental demand; S24. Forecast Error Analysis and Dynamic Safety Stock Setting: Based on the corrected forecast results, calculate the safety stock for each storage node: Calculate the standard deviation of the forecast error: Use a sliding window to statistically analyze the deviation between recent forecast values ​​and actual sales, and calculate the standard deviation. σe ; Determine the service level coefficient: Based on the product profile and SLA agreement constructed in S1, query the Z-value of the normal distribution quantile; Dynamic threshold calculation: combining replenishment lead time L and its variance. Substitute into the dynamic safety stock formula: ,in This represents the average daily demand. S25. ABC Classification Inventory Warning Configuration: Based on the classification results in S1, configure differentiated inventory warning strategies: Category A: Establish a high-frequency inventory counting mechanism; when inventory levels fall below S... A red alert is immediately triggered when the value reaches 1.2 times the normal value, and the S3 scheduling model is activated to prioritize processing. Category B / C: Appropriately relax the warning threshold to S 0.8 times, to reduce inventory holding costs and avoid capital accumulation.

4. The method for dynamic scheduling and optimization of a multi-category supply chain based on S2B2C as described in claim 1, characterized in that, In step S3, a multi-objective optimization model is constructed, and the objective function and constraints are set, as follows; The objective function is set as a weighted combination of minimizing total cost and maximizing service level; where total cost includes procurement cost, warehousing cost, transportation cost, and stockout penalty cost. Service level includes order fulfillment rate and on-time delivery rate; Constraints include hard constraints and soft constraints; Hard constraints include: supplier capacity limits, warehouse capacity limits, vehicle load / volume limits, and restrictions on mixed loading of multiple product categories; Soft constraints include: B-end merchants' preference for specifying logistics providers and C-end customers' expected delivery time windows.

5. The method for dynamic scheduling and optimization of a multi-category supply chain based on S2B2C as described in claim 1, characterized in that, In step S4, a genetic algorithm combined with reinforcement learning is used to solve the model and generate procurement scheduling, inventory transfer, and logistics dispatch instructions. The specific steps are as follows: S41. Chromosome Encoding and Intelligent Initialization: A segmented real-number encoding strategy is used to construct a chromosome structure. The chromosome includes a procurement gene segment, an inventory allocation gene segment, and a logistics route gene segment, which correspond to the supplier selection, allocation quantity, and vehicle route in the decision variables, respectively. A pre-trained reinforcement learning agent is used to perceive the current environmental state and generate an initial solution population containing prior knowledge to replace random initialization, thereby accelerating the convergence of the algorithm. S42. Fitness Function Construction and Evaluation: Based on the objective function constructed in step S3, establish a fitness evaluation function; handle soft constraints through the penalty function method, transform violations of time windows and B-end preference constraints into penalty terms included in the total cost, calculate the fitness value of each individual in the population, and evaluate the merits of the scheduling scheme; S43. Genetic Operations and Iterative Evolution: Perform selection, crossover, and mutation operations on the current population; the selection operation uses a tournament strategy to retain individuals with high fitness; the crossover operation uses a segmented crossover method to exchange procurement and logistics plans for different individuals while ensuring the logical integrity of gene segments; the mutation operation uses an adaptive mutation operator to randomly perturb gene positions to maintain population diversity. S44. Local search guided by reinforcement learning mechanism: During the iteration process of genetic algorithm, reinforcement learning agent is introduced to intervene. The agent perceives the current environmental state in real time. The agent makes targeted corrections to individuals in genetic algorithm according to the trained policy network. When the agent finds that a vehicle is overloaded, it directly issues an action command to adjust the loading order. When the algorithm is close to the optimal solution but the oscillation does not converge, the reinforcement learning agent directly locks the key parameters based on historical experience and quickly outputs the final result. S45. Termination Decision and Instruction Decoding Output: Determine if the maximum number of iterations and fitness improvement threshold have been reached; when the optimal individual with the highest fitness is output, perform a decoding operation on the optimal individual and map it to a specific business execution instruction: Procurement scheduling instructions: Determine the order quantity and order time for each supplier; Inventory transfer instructions: Determine the transfer routes and quantities of goods between various storage nodes; Logistics dispatch instructions: determine vehicle allocation, delivery route sequence, and carrier selection.

6. The method for dynamic scheduling and optimization of a multi-category supply chain based on S2B2C as described in claim 1, characterized in that, Step S5 involves monitoring the supply chain execution status and setting triggering and exception response mechanisms, specifically including: Real-time data collection of logistics trajectory and warehousing environment parameters is achieved through IoT devices, including temperature and humidity data for specific product categories. Real-time monitoring of demand fluctuations to identify sudden surges and drops in demand; Set abnormal trigger conditions. When the delay exceeds the preset time threshold, the inventory level falls below the safety stock warning line, or the environmental parameters exceed the preset compliance range, a partial or global recalculation is triggered. Emergency plans are automatically generated and executed, including emergency replenishment, switching logistics providers, and canceling low-priority orders.

Images