An adaptive AIGC inventory intelligent optimization and replenishment strategy prediction system

Abstract

The application relates to the field of inventory management and discloses a self-adaptive AIGC inventory intelligent optimization and replenishment strategy prediction system, which comprises the following modules: a situation perception and feature extraction module that processes internal and external multi-source data by using a large language model fine-tuned in the field to output a structured situation feature vector; a demand prediction and simulation coupling module that receives the feature vector and candidate strategy parameters from a strategy generation module, performs conditional demand prediction and parallel simulation analysis; a strategy generation and optimization module that internally generates an artificial intelligence model, generates and screens structured replenishment strategies based on prediction and simulation results; and a collaborative optimization control module that coordinates the prediction simulation and strategy generation modules to perform multiple rounds of iterative optimization through a dynamic feature attention modulation unit in the module, and reversely adjusts the weight of the situation feature vector in the system based on the sensitivity analysis results generated by the simulation. The application realizes deep perception of market situations and causal deduction of strategy influences.

Description

Technical Field

[0001] This invention relates to the field of inventory management, specifically to an adaptive AIGC inventory intelligent optimization and replenishment strategy prediction system. Background Technology

[0002] In the field of inventory management, traditional replenishment decision-making systems mainly rely on statistical forecasting models (such as time series analysis) and operations research optimization methods (such as economic order quantity models) based on historical sales data. These methods typically treat forecasting and optimization as two separate processes: the forecasting module outputs a single estimate of future demand, while the optimization module solves for the replenishment plan that minimizes cost or maximizes service level under this fixed input. However, this static, unidirectional decision-making process has some shortcomings.

[0003] First, traditional systems cannot effectively absorb and quantify external unstructured environmental information, such as the impact of social media sentiment, industry news, competitor activities, and macroeconomic events on demand. This information is a key factor leading to sudden fluctuations and structural changes in demand, and its absence causes predictive models to be severely lagging and inaccurate in responding to market changes.

[0004] Secondly, there is a disconnect between the forecasting and optimization stages. The forecasting model does not consider the potential feedback effect of the replenishment strategy on market demand, while the optimization model is based on a fixed demand assumption that does not change with the strategy. This leads to the strategy failing in actual implementation due to market demand deviating from expectations, meaning there is a lack of a strategy-market feedback loop.

[0005] Furthermore, the decision-making process lacks transparency and dynamic adjustment capabilities. The system typically outputs a rigid optimal solution, failing to provide a set of options for weighing different objectives, and lacking mechanisms to calibrate and learn its judgment logic in response to environmental changes. Decision-makers struggle to understand the rationale behind the proposed solutions and the potential risks. Summary of the Invention

[0006] The purpose of this invention is to provide an adaptive AIGC inventory intelligent optimization and replenishment strategy prediction system to solve at least one of the above-mentioned technical problems.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] An adaptive AIGC inventory intelligent optimization and replenishment strategy prediction system includes:

[0009] The collaborative optimization control module, and the context awareness and feature extraction module, demand prediction and simulation coupling module, and strategy generation and optimization module that are connected to it in communication;

[0010] The context awareness and feature extraction module includes a large language model fine-tuned by text in the supply chain domain, used to process internal and external multi-source heterogeneous data in real time, and output a structured supply chain context feature vector with time alignment and source credibility weights.

[0011] The demand prediction and simulation coupling module is used to receive the context feature vector and candidate policy parameters from the policy generation and optimization module, perform conditional demand prediction that integrates context information, and perform parallel simulation deduction and sensitivity analysis on the candidate policies.

[0012] The strategy generation and optimization module includes a generative artificial intelligence model, which is used to generate, filter and sort structured replenishment strategies based on the prediction and simulation results output by the demand prediction and simulation coupling module.

[0013] The collaborative optimization control module includes a dynamic feature attention modulation unit, which coordinates the demand prediction and simulation coupling module and the strategy generation and optimization module to perform multiple rounds of iterative optimization, and adjusts the weight of the context feature vector in prediction and simulation in reverse through the dynamic feature attention modulation unit based on the sensitivity analysis results generated by the simulation.

[0014] In a further embodiment, the demand forecasting and simulation coupling module includes:

[0015] The counterfactual forecasting unit is used to dynamically adjust the input or parameters of the forecasting model based on the future inventory and market action parameters corresponding to the specific candidate replenishment strategy received, and to generate conditional demand forecasts and forecast interval distributions under that strategy.

[0016] The simulation and sensitivity analysis unit is used to perform parallel Monte Carlo simulations of multiple candidate replenishment strategies under various preset market scenarios based on the conditional demand forecast and forecast interval distribution, output the performance probability distribution of each strategy, calculate the contribution of each scenario feature dimension to the performance index based on the simulation data, and generate a sensitivity analysis report.

[0017] In a further embodiment, the strategy generation and optimization module includes:

[0018] The structured strategy generation unit is used to generate a set of candidate replenishment strategies represented by a data structure of inventory units, replenishment time points, and replenishment quantities, based on the current inventory status, business constraints, and forecast inputs.

[0019] The risk perception pre-screening unit, connected to the strategy knowledge base, is used to match the generated candidate replenishment strategies with the high failure risk strategy patterns recorded in the strategy knowledge base, and filter out candidate strategies whose comprehensive similarity on the key dimensions exceeds a preset risk threshold.

[0020] In a further embodiment, the multi-round iterative optimization process executed by the collaborative optimization control module includes:

[0021] Initiate the first round of optimization to obtain an initial Pareto frontier that includes at least one replenishment strategy;

[0022] Determine whether there exists a first strategy and a second strategy in the initial Pareto front that have a difference greater than a threshold on at least two preset performance objectives;

[0023] If so, then trigger the iteration: guide the demand forecasting and simulation coupling module to generate demand forecasting intervals at different quantile levels for the first strategy and the second strategy; guide the strategy generation and optimization module to generate new candidate strategies in the parameter neighborhood of the first strategy and the second strategy; and perform simulation evaluation on the new candidate strategies;

[0024] The Pareto front is updated based on the evaluation results until the number of iterations reaches the upper limit or the overall improvement of the updated Pareto front on all preset performance objectives is lower than the preset tolerance. The iteration is then terminated and the final Pareto front is output as the recommended strategy set.

[0025] In a further embodiment, the operation of the dynamic feature attention modulation unit includes:

[0026] Receive the sensitivity analysis report output by the simulation and sensitivity analysis unit;

[0027] Based on the report, a set of dynamic weight coefficients corresponding one-to-one with each dimension of the context feature vector were calculated.

[0028] The dynamic weight coefficients are sent to the context awareness and feature extraction module to adjust the numerical scale of its output feature vector, and then sent to the counterfactual prediction unit to adjust the weighting of each context feature in the prediction model.

[0029] In a further embodiment, the records stored in the strategy knowledge base include: parameters of the replenishment strategy executed in the past, a snapshot of the market situation before execution, and deviation data between the actual performance and the expected performance after execution;

[0030] When the risk perception pre-screening unit performs matching, the key dimensions calculated include: the combination of inventory level and demand fluctuation, and the combination of promotion intensity and competitive situation.

[0031] In a further embodiment, the predictive model parameters dynamically adjusted by the counterfactual prediction unit include at least one of the following: demand price elasticity coefficient, sales loss rate due to inventory shortage, and demand satisfaction priority across supply chain nodes.

[0032] In a further embodiment, the system also includes:

[0033] The closed-loop learning and calibration module is used to monitor the actual performance of the implemented replenishment strategy and compare it with the performance predicted by simulation.

[0034] The closed-loop learning and calibration module includes a deviation attribution unit, which is used to determine the main source category of performance deviation based on the comparison results and market context data when the strategy is executed;

[0035] The main sources include: basic trend prediction error, unidentified external event impact, event impact estimation bias, and strategy execution environment anomalies.

[0036] The closed-loop learning and calibration module selectively sends parameter update instructions to the context awareness and feature extraction module, the prediction model in the demand prediction and simulation coupling module, or the strategy knowledge base, based on the main source category.

[0037] A further proposed approach is to include at least two types of quantitative indicators in the structured supply chain context feature vector:

[0038] Parameters of the event influence decay function derived from event text recognition and historical impact analysis;

[0039] Estimated coefficients of competitive cross-elasticity derived from analysis of competitor pricing and activity information;

[0040] An index of the intensity of emotion transmission derived from regionalized social media sentiment analysis.

[0041] In a further embodiment, each replenishment strategy in the recommended strategy set is accompanied by a multi-dimensional evaluation label derived by the simulation and sensitivity analysis unit. The label includes: the expected profit under the baseline scenario, the guaranteed profit level under the worst-case credible scenario, and the identifiers of the top N uncertainty factors that contribute the most to profit fluctuations.

[0042] The beneficial effects of this invention are:

[0043] This invention interprets unstructured market information in real time through a context awareness and feature extraction module, and through a dynamic attention modulation unit in the collaborative optimization control module, adjusts the upstream feature weights in reverse according to the sensitivity results of downstream simulation analysis. This enables the system to have the ability to focus on key contradictions similar to human experts, dynamically capture and respond to core environmental variables that affect decision-making, and improve the effectiveness of replenishment. Attached Figure Description

[0044] The invention will now be further described with reference to the accompanying drawings.

[0045] Figure 1 This is a logical schematic diagram of the present invention. Detailed Implementation

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] Please see Figure 1 As shown, this invention is an adaptive AIGC inventory intelligent optimization and replenishment strategy prediction system, comprising:

[0048] The collaborative optimization control module, and the context awareness and feature extraction module, demand prediction and simulation coupling module, and strategy generation and optimization module that are connected to it in communication;

[0049] The context awareness and feature extraction module includes a large language model fine-tuned by text in the supply chain domain, used to process internal and external multi-source heterogeneous data in real time, and output a structured supply chain context feature vector with time alignment and source credibility weights.

[0050] Specific types of multi-source heterogeneous data: Internal and external multi-source heterogeneous data are the core inputs of the system, specifically including: internal data (inventory data, sales data, procurement data, and production data from the enterprise ERP system, such as daily inventory balance, historical replenishment records, and product outbound details); external data (market demand data, competitor pricing / replenishment strategies, industry policies, social media sentiment data, and logistics timeliness data, such as sales volume of similar products on a certain e-commerce platform, regional epidemic prevention and control policies, and product evaluation data on Douyin / Xiaohongshu).

[0051] Example of time-series alignment and source credibility weighting: Time-series alignment unifies data at different time granularities into a unified standard (e.g., aligning hourly sales data, daily inventory data, and weekly policy data to hourly levels, with missing data filled in using linear interpolation); Source credibility weights are assigned based on data reliability. Example: Enterprise ERP system inventory data (reliable source, weight 0.8), third-party market research data (weight 0.6), social media sentiment data (high noise, weight 0.3). In the final structured feature vector, the value of each dimension = original data value × corresponding source weight.

[0052] A large language model fine-tuned from supply chain domain texts: Based on general large language models (such as LLaMA and ChatGLM), it is fine-tuned using supply chain domain text data (such as inventory management manuals, replenishment cases, supply chain policy documents, and industry reports) to accurately identify supply chain domain terms (such as safety stock Pareto frontier Monte Carlo deduction) and avoid the general model's misinterpretation of professional texts. For example, it can extract the contextual features of reduced logistics timeliness from the recent upgrade of regional logistics control and the extension of the replenishment cycle by 3 days, and quantify it as structured data of a 30% extension of the replenishment cycle.

[0053] The demand prediction and simulation coupling module is used to receive the context feature vector and candidate policy parameters from the policy generation and optimization module, perform conditional demand prediction that integrates context information, and perform parallel simulation deduction and sensitivity analysis on the candidate policies.

[0054] The strategy generation and optimization module includes a generative artificial intelligence model, which is used to generate, filter and sort structured replenishment strategies based on the prediction and simulation results output by the demand prediction and simulation coupling module.

[0055] Generative AI Model: This is a generative AI model (adaptable to architectures such as GPT series and diffusion models). It is specifically optimized for inventory replenishment strategy generation scenarios. It can combine demand forecast results, business constraints, and strategy knowledge base data to autonomously generate candidate replenishment strategies that meet the format requirements (inventory unit, replenishment time point, replenishment quantity tuple). It also has the ability to filter and rank strategies. Based on simulation performance results, it can eliminate low-performance and high-risk strategies and retain the optimal strategy set. Its core advantage is that it can quickly generate diverse candidate strategies and can adapt to the dynamic changes of different market scenarios through multiple rounds of iterative optimization.

[0056] Example: The model architecture should preferably use GPT-4o-mini (lightweight adaptation) or GPT-4o (high accuracy requirement), the generation strategy iteration efficiency should be ≤30 seconds / round (time spent on single round of candidate strategy generation and preliminary screening), the number of candidate strategies generated can be configured (5-20 / round), the strategy screening and ranking accuracy should be ≥95% (the proportion of high-performing strategies selected), and parallel generation should be supported (up to 10 candidate strategies can be generated at the same time to improve efficiency).

[0057] The collaborative optimization control module includes a dynamic feature attention modulation unit, which coordinates the demand prediction and simulation coupling module and the strategy generation and optimization module to perform multiple rounds of iterative optimization, and adjusts the weight of the context feature vector in prediction and simulation in reverse through the dynamic feature attention modulation unit based on the sensitivity analysis results generated by the simulation.

[0058] Example of reverse adjustment in dynamic feature attention modulation unit: If the sensitivity analysis results show that the contribution of market demand fluctuation features to the performance of replenishment strategy is 60%, while the contribution of raw material price features is only 5%, then this unit calculates dynamic weight coefficients (demand fluctuation 0.6, raw material price 0.05) for reverse adjustment: In the context feature vector, the value of the demand fluctuation dimension is multiplied by 0.6, and the value of the raw material price dimension is multiplied by 0.05 to adjust the feature scale; In the prediction model, the weight of demand fluctuation features is increased from 0.4 to 0.6, and the weight of raw material price is decreased from 0.2 to 0.05, prioritizing high-contribution features.

[0059] By working collaboratively across four modules, the system achieves precise processing of multi-source heterogeneous data, deep coupling of demand and strategy, multi-round iterative optimization, and dynamic feature weight adjustment. This addresses the pain points of traditional inventory systems, such as messy feature processing, disconnect between forecasting and strategy, and unemotional optimization. Leveraging a large-scale supply chain fine-tuning model and a generative AI model, the system improves feature extraction accuracy and strategy generation efficiency, ensuring that strategies adapt to dynamic changes in the supply chain. This lays the foundation for the accurate operation of subsequent modules and enhances the overall intelligence and adaptability of inventory optimization and replenishment forecasting.

[0060] The demand forecasting and simulation coupling module includes:

[0061] The counterfactual forecasting unit is used to dynamically adjust the input or parameters of the forecasting model based on the future inventory and market action parameters corresponding to the specific candidate replenishment strategy received, and to generate conditional demand forecasts and forecast interval distributions under that strategy.

[0062] Counterfactual forecasting elucidates the demand generated if a candidate strategy is implemented, focusing on eliminating interference from other strategies. Example: A candidate replenishment strategy involves replenishing 1000 cases on the 5th of each month, simultaneously launching an 80% discount promotion. The corresponding future inventory parameters are (initial inventory after replenishment: 1500 cases) and market action parameters are (promotion intensity: 80%). The counterfactual forecasting unit adjusts the forecasting model parameters (e.g., adjusting the price elasticity of demand from -0.5 to -0.7 to better align with the promotional scenario) to generate a conditional demand forecast under this strategy: an average daily demand of 120 cases over the next 30 days, with a forecast interval distribution (90th quantile: 140 cases / day, 50th quantile: 120 cases / day, 10th quantile: 100 cases / day). This means there is a 90% probability that the average daily demand will not exceed 140 cases.

[0063] The simulation and sensitivity analysis unit is used to perform parallel Monte Carlo simulations of multiple candidate replenishment strategies under various preset market scenarios based on the conditional demand forecast and forecast interval distribution. It outputs the performance probability distribution of each strategy, calculates the contribution of each scenario feature dimension to the performance indicators based on the simulation data, and generates a sensitivity analysis report.

[0064] Example of preset market scenarios: Three typical market scenarios are preset to cover different uncertainties: Baseline scenario (stable demand, fluctuation ±10%, no external interference); Optimistic scenario (peak season arrives, demand increases by 30%, competitors have no promotions); Pessimistic scenario (off-season, demand decreases by 20%, competitors lower prices by 10%).

[0065] Sensitivity Analysis Report Example: The core content of the report is the contribution of each situational feature to the performance indicator (such as stockout rate, expected profit). Example: When the performance indicator is expected profit, the contribution of situational features is ranked as follows: market demand fluctuation (60%) > promotion intensity (20%) > logistics timeliness (10%) > raw material price (10%). That is, 60% of the profit fluctuation is caused by market demand fluctuation, which provides a basis for subsequent weight adjustment.

[0066] By using a counterfactual prediction unit, the system achieves strategy-driven demand forecasting, addressing the issues of disconnect between traditional demand forecasting and replenishment strategies, and the lack of specificity in forecast results, thereby improving the accuracy and adaptability of demand forecasting. Parallel Monte Carlo simulations enable efficient simulation evaluation across multiple strategies and scenarios, rapidly outputting the performance probability distribution of each strategy and avoiding the limitations of single-scenario simulations. Sensitivity analysis accurately identifies core scenario characteristics affecting strategy performance, providing a core basis for the dynamic weight adjustment of the collaborative optimization control module, while also providing quantitative support for strategy selection, reducing subsequent iterations of ineffective strategies, and improving system optimization efficiency.

[0067] The strategy generation and optimization module includes:

[0068] The structured strategy generation unit is used to generate a set of candidate replenishment strategies represented by a data structure of inventory units, replenishment time points, and replenishment quantities, based on the current inventory status, business constraints, and forecast inputs.

[0069] Example of a structured replenishment strategy tuple: The tuple format is uniformly (inventory unit, replenishment time, replenishment quantity). Combined with business constraints (such as maximum warehouse capacity of 2000 boxes, replenishment cycle ≥ 7 days, minimum replenishment quantity of 500 boxes), a set of candidate strategies is generated. Examples: (boxes, 3rd working day of the month, 800 boxes); (boxes, 10th working day of the month, 1000 boxes); (boxes, 17th working day of the month, 1200 boxes), generating a total of 10-20 candidate strategies (the number is configurable).

[0070] The risk perception pre-screening unit, connected to the strategy knowledge base, is used to match the generated candidate replenishment strategies with the high failure risk strategy patterns recorded in the strategy knowledge base, and filter out candidate strategies whose comprehensive similarity on the key dimensions exceeds a preset risk threshold.

[0071] The generative AI model in the strategy generation and optimization module employs a constraint embedding and performance-oriented generation mechanism. The specific application process is as follows: First, the model reads input data such as the current inventory status, business constraints (warehouse capacity, replenishment cycle, etc.), and demand forecast results, embedding these into the model to generate constraints. Second, it combines historical effective strategy patterns from the strategy knowledge base to generate an initial set of candidate replenishment strategies (10-20, configurable). During generation, strategies that clearly do not meet the constraints are automatically avoided (e.g., replenishment quantity exceeding the maximum warehouse capacity, replenishment cycle shorter than required). Finally, it receives simulation performance results from the demand forecasting and simulation coupling module, and uses the model's built-in ranking algorithm (e.g., weighted ranking based on expected profit and stockout rate) to screen and rank the candidate strategies, eliminating underperforming strategies. Simultaneously, it combines risk perception pre-screening results to filter high-failure-risk strategies, ultimately outputting an optimized set of candidate strategies to support subsequent multiple rounds of iterative optimization.

[0072] Example of a strategy knowledge base and high-failure-risk strategy pattern: The strategy knowledge base stores historical execution data in the following format:

[0073] Strategy ID Replenishment Parameters (Tuples) Execution Market Situation Actual Performance Expected Performance High Failure Risk S001 (Boxes, 5th of each month, 1500 boxes) Off-season, demand fluctuation 15% Inventory backlog rate 30%, loss of 50,000 Inventory backlog rate ≤10%, profit of 80,000 Yes S002 (Boxes, 10th of each month, 600 boxes) Peak season, demand fluctuation 25% Stockout rate 12%, profit of 30,000 Stockout rate ≤5%, profit of 100,000 Yes High Failure Risk Strategy Mode: Off-season and replenishment quantity ≥1200 boxes Peak season + replenishment quantity ≤800 boxes.

[0074] Example of key dimension matching and risk threshold: The key dimensions are the combination of inventory level and demand fluctuation, and the combination of promotion intensity and competitive situation; the overall similarity is calculated using cosine similarity (value 0-1), and the preset risk threshold is 0.8 (configurable). Example: A newly generated candidate strategy is (boxes, 5th of each month, 1400 boxes), the execution scenario is off-season, demand fluctuation is 12%, and the overall similarity with the high failure risk mode S001 is 0.85, which exceeds the threshold of 0.8, so this candidate strategy is filtered out.

[0075] The structured strategy generation unit ensures that candidate strategies are formatted correctly and comply with business constraints, avoiding the problems of disorganized and unsuitable strategy generation in traditional approaches, thus improving the standardization and effectiveness of strategy generation. By leveraging the linkage between the risk perception pre-screening unit and the strategy knowledge base, high-failure-risk strategies are filtered in advance, reducing the ineffective workload of subsequent simulation and iterative optimization, and lowering the probability of strategy execution failure. Combined with the efficient generation and ranking capabilities of generative AI models, a high-quality candidate strategy set is quickly output, improving strategy generation efficiency and screening accuracy, and providing reliable input for multiple rounds of iterative optimization.

[0076] The multi-round iterative optimization process executed by the collaborative optimization control module includes:

[0077] Initiate the first round of optimization to obtain an initial Pareto frontier containing at least one replenishment strategy. Pareto frontier and initial optimization example: The Pareto frontier is a set of strategies without disadvantages, meaning that no single strategy in the set can improve another performance objective without lowering one. Preset performance objectives (configurable): Expected profit (higher is better); stockout rate (lower is better); inventory turnover rate (higher is better). First round optimization example: From 15 candidate strategies, select 3 strategies without disadvantages to form the initial Pareto frontier.

[0078] Strategy A has an expected profit of 1 million, a stockout rate of 6%, and an inventory turnover rate of 3.5 times per month.

[0079] Strategy B: Expected profit of 900,000 (RMB), stockout rate of 4%, and inventory turnover rate of 3.2 (times / month).

[0080] Strategy C: Expected profit of 800,000 (ten thousand), stockout rate of 3% (%), and inventory turnover rate of 3.0 (times / month).

[0081] Determine whether there exist a first strategy and a second strategy in the initial Pareto front that differ from at least two preset performance objectives by a threshold; examples of strategy difference determination and thresholds are provided.

[0082] Preset difference thresholds (configurable): Expected profit difference ≥ 100,000, stockout rate difference ≥ 2%, inventory turnover rate difference ≥ 0.3 times / month. Judgment example: Strategy A (profit 1 million, stockout rate 6%) and Strategy B (profit 900,000, stockout rate 4%) have a profit difference of 100,000 (equal to the threshold) and a stockout rate difference of 2% (equal to the threshold). Since at least two performance targets have differences greater than the threshold, iteration is triggered.

[0083] If so, then trigger the iteration: guide the demand forecasting and simulation coupling module to generate demand forecasting intervals at different quantile levels for the first strategy and the second strategy; guide the strategy generation and optimization module to generate new candidate strategies in the parameter neighborhood of the first strategy and the second strategy; and perform simulation evaluation on the new candidate strategies;

[0084] The Pareto front is updated based on the evaluation results until the number of iterations reaches the upper limit or the overall improvement of the updated Pareto front on all preset performance objectives is lower than the preset tolerance. The iteration is then terminated and the final Pareto front is output as the recommended strategy set.

[0085] Example of iterative process:

[0086] Demand Forecasting: Generate 90th, 50th, and 10th percentile demand ranges for Strategy A (140 boxes / day, 120 boxes / day, and 100 boxes / day), and generate corresponding quantile ranges for Strategy B (130 boxes / day, 110 boxes / day, and 90 boxes / day).

[0087] New candidate strategies are generated within the neighborhood of parameters A and B. Examples: neighborhood strategies for A (boxes, 3rd of each month, 900 boxes) and (boxes, 5th of each month, 1100 boxes), neighborhood strategies for B (boxes, 8th of each month, 850 boxes) and (boxes, 12th of each month, 950 boxes).

[0088] Simulation evaluation: Monte Carlo simulations were performed on the four new strategies to evaluate their performance;

[0089] Update Pareto Frontier: If the new strategy (boxes, 3rd of each month, 900 boxes) performs better than the original strategy (profit of 950,000, stockout rate of 5%, turnover rate of 3.4), then replace the corresponding inferior strategy and update the Pareto Frontier.

[0090] Example of termination condition: The maximum number of iterations is set to 5; the preset tolerance is set to 5% (i.e., if the weighted average improvement of each performance objective in two consecutive Pareto fronts is <5%, then the iteration terminates). Example: After the 4th iteration, the expected profit of the Pareto front is 980,000 on average, and after the 5th iteration it is 990,000. The improvement is 1.02% < 5%, so the iteration terminates, and the final Pareto front (3 strategies) is output as the recommended strategy set.

[0091] Through a multi-round iterative optimization mechanism, the Pareto front is gradually optimized, addressing the pain point of traditional strategy optimization which generates strategies in a single round without subsequent optimization, ensuring that the final output of the recommended strategy set consists of strategies without disadvantages. Iteration is triggered by judging the performance differences of strategies, avoiding ineffective iterations and improving optimization efficiency. Combining quantile demand prediction and neighborhood strategy generation, better strategies are accurately discovered, achieving a balance between multiple performance objectives (profit, stockout rate, turnover rate), adapting to the different operational priorities of enterprises, and the final output of the recommended strategy set has stronger practicality and optimality.

[0092] The operation of the dynamic feature attention modulation unit includes:

[0093] Receive the sensitivity analysis report output by the simulation and sensitivity analysis unit;

[0094] Based on the report, a set of dynamic weight coefficients corresponding one-to-one with each dimension of the context feature vector were calculated; example of dynamic weight coefficient calculation:

[0095] In the sensitivity analysis report, the contribution of each scenario feature to the expected profit is as follows: market demand fluctuation (60%), promotion intensity (20%), logistics timeliness (10%), and raw material price (10%). The dynamic weight coefficients are calculated using contribution normalization, i.e., the weight coefficient of each dimension = contribution of that dimension / sum of the contributions of all dimensions (in this example, the sum is 100%). The calculation results are: demand fluctuation 0.6, promotion intensity 0.2, logistics timeliness 0.1, and raw material price 0.1, which correspond one-to-one with each dimension of the scenario feature vector (feature vector dimensions: [demand fluctuation, promotion intensity, logistics timeliness, raw material price]).

[0096] The dynamic weight coefficients are sent to the context-aware and feature extraction module to adjust the numerical scale of its output feature vector; an example of adjusting the numerical scale of the feature vector is as follows:

[0097] The original feature vector (unadjusted) output by the context awareness and feature extraction module is: [120 (demand fluctuation value), 0.8 (promotion intensity), 7 (logistics timeliness, days), 5000 (raw material price, yuan / ton)]. After multiplying by the dynamic weight coefficient, the adjusted feature vector is: [120×0.6=72, 0.8×0.2=0.16, 7×0.1=0.7, 5000×0.1=500]. This achieves the amplification of high-contribution feature values ​​and the reduction of low-contribution feature values, highlighting key information.

[0098] The data is then sent to the counterfactual prediction unit to adjust the weighting of each situational feature in the prediction model.

[0099] This system enables dynamic adaptive adjustment of contextual feature weights, addressing the problem of fixed feature weights in traditional systems that cannot adapt to different market scenarios and strategy requirements. By combining sensitivity analysis results to calculate weight coefficients, it ensures the targeted nature of weight adjustments, prioritizing core features that have a significant impact on strategy performance while mitigating the interference of low-contribution features. Simultaneously, it links contextual awareness and counterfactual prediction modules to achieve a closed loop of sensitivity analysis, weight adjustment, feature optimization, and improved prediction accuracy. This further enhances the accuracy of demand forecasting and the reliability of strategy simulation, providing higher-quality support for multi-round iterative optimization.

[0100] The strategy knowledge base stores records including: parameters of historically executed replenishment strategies, a snapshot of the market situation before execution, and deviation data between actual and expected performance after execution. The strategy knowledge base records are stored in a structured database, with each record containing six core fields. Example:

[0101] Record ID replenishment strategy parameters (tuples) Market scenario snapshot (1 day before execution) Actual performance Expected performance deviation data R001 (boxes, 2024-05-05, 1200 boxes) Inventory level: High (1800 boxes); Demand fluctuation: 20%; Promotion intensity: 0 (no promotion); Competitive situation: Mild (competitors have no action) Stockout rate 2%, profit 850,000, turnover rate 3.1 times / month Stockout rate 5%, profit 900,000, turnover rate 3.3 times / month Stockout rate deviation -3%, profit deviation -50,000, turnover rate deviation -0.2 times / month;

[0102] Among them, the market scenario snapshot is complete scenario feature data at a certain point in time before execution (such as 1 day or 3 days before), which is used to reconstruct the environment when the strategy is executed.

[0103] When the risk perception pre-screening unit performs matching, the key dimensions calculated include: the combination of inventory level and demand fluctuation, and the combination of promotion intensity and competitive situation.

[0104] Example of key dimension combination states:

[0105] Combinations of inventory levels and demand fluctuations (6 typical combinations, expandable):

[0106] Combination 1: High inventory level + low demand fluctuation (e.g., 1800 cases inventory, 10% fluctuation); Combination 2: High inventory level + high demand fluctuation (e.g., 1800 cases inventory, 30% fluctuation); Combination 3: Moderate inventory level + low demand fluctuation (e.g., 1200 cases inventory, 10% fluctuation); Combination 4: Moderate inventory level + high demand fluctuation (e.g., 1200 cases inventory, 30% fluctuation); Combination 5: Low inventory level + low demand fluctuation (e.g., 600 cases inventory, 10% fluctuation); Combination 6: Low inventory level + high demand fluctuation (e.g., 600 cases inventory, 30% fluctuation).

[0107] Combinations of promotional intensity and competitive landscape (4 typical combinations, expandable):

[0108] Combination A: High promotion intensity (discount ≤ 80%) + intense competition (competitors' discounts ≤ 75%); Combination B: High promotion intensity + moderate competition (competitors have no promotions); Combination C: Low promotion intensity (discount ≥ 90%) + intense competition; Combination D: Low promotion intensity + moderate competition.

[0109] By standardizing the stored content of the strategy knowledge base, the execution environment, parameters, and performance deviations of historical strategies are fully restored, providing authentic and comprehensive historical evidence for risk screening and strategy optimization. This avoids the problems of incomplete data and inability to support accurate matching in traditional knowledge bases. The key dimensions of risk matching are clarified to achieve accurate matching of strategy scenarios, improve the accuracy of risk perception pre-screening, and avoid misjudgments caused by matching based on a single dimension. Through comprehensive similarity calculation, high-failure-risk strategies are accurately identified, further reducing the risk of strategy execution. At the same time, historical references are provided for generative AI models, improving the reliability of candidate strategies.

[0110] The prediction model parameters dynamically adjusted by the counterfactual prediction unit include:

[0111] The price elasticity of demand; the degree to which changes in quantity and price affect demand. The formula is: rate of change in demand / rate of change in price. A negative value indicates that as prices rise, demand falls. The larger the absolute value, the more significant the impact of price on demand. Adjustment example:

[0112] If the candidate strategy has no promotion (price remains unchanged), the coefficient is taken as the base value -0.5 (for every 1% increase in price, demand decreases by 0.5%).

[0113] If the candidate strategy is an 80% discount (price reduction of 20%), the coefficient is adjusted to -0.7 (in promotional scenarios, price has a more significant impact on demand; for every 1% decrease in price, demand increases by 0.7%).

[0114] If the candidate strategy is a 10% price increase, the coefficient is adjusted to -0.4 (user sensitivity is slightly lower in the price increase scenario).

[0115] Sales loss rate due to inventory shortages: When inventory is out of stock, the proportion of unmet demand that ultimately leads customers to abandon their purchase (rather than waiting for restocking) directly impacts performance calculations (the higher the sales loss rate, the greater the profit loss from inventory shortages). Adjustment example:

[0116] If the candidate strategy is to prioritize replenishment for core customers, the sales loss rate will be adjusted to 30% (core customers wait for replenishment, while ordinary customers give up, with an abandonment rate of 30%).

[0117] If the candidate strategy is no-priority replenishment, the sales loss rate is adjusted to 50% (50% of all customers abandon the purchase).

[0118] If the candidate strategy is to provide a replenishment notification when stockouts occur, the sales loss rate will be adjusted to 20% (some customers will wait for replenishment).

[0119] Demand fulfillment priority across supply chain nodes: When there are multiple nodes in the supply chain (such as central warehouse, regional warehouses, and stores), and total inventory is insufficient, prioritizing the fulfillment of demand from any node affects the stockout rate and overall performance of each node. Adjustment example:

[0120] If the candidate strategy is profit-oriented, the priority is adjusted to high-end stores > regional warehouses > central warehouses > ordinary stores (high-end stores have higher profits and should be prioritized).

[0121] If the candidate strategy is market share-oriented, the priority will be adjusted to core regional stores > ordinary regional stores > regional warehouses > central warehouses (prioritizing coverage of core markets).

[0122] If the candidate strategy is inventory turnover-oriented, the priority is adjusted to nodes with low inventory turnover rate > nodes with high turnover rate (prioritizing the digestion of backlogged inventory).

[0123] By dynamically adjusting the core parameters of the forecasting model, demand forecasting can accurately adapt to the different scenarios of various candidate strategies, solving the problems of fixed parameters and disconnect between forecast results and strategy scenarios in traditional counterfactual forecasting. For key scenarios such as price, stockouts, and supply chain nodes, the corresponding parameters are precisely adjusted to improve the accuracy and relevance of conditional demand forecasting, ensuring that the forecast results can truly reflect the demand situation under the strategy. At the same time, parameter adjustments are aligned with the company's operational priorities (profit, market share, inventory turnover), making the forecast results and subsequent simulation performance more closely match the company's actual needs, providing a more reliable basis for strategy selection and optimization.

[0124] The system also includes:

[0125] The closed-loop learning and calibration module monitors the actual performance of the implemented replenishment strategy and compares it with the simulated predicted performance. Example of comparing actual performance with simulated predicted performance:

[0126] The implemented strategy is (boxes, 5th of each month, 1000 boxes). Simulation predicted performance: expected profit of 1 million, stockout rate of 5%, and inventory turnover rate of 3.5 times / month; actual performance: expected profit of 850,000, stockout rate of 12%, and inventory turnover rate of 2.8 times / month; comparison deviation: profit deviation of -150,000, stockout rate deviation of +7%, and turnover rate deviation of -0.7 times / month, triggering deviation attribution.

[0127] The closed-loop learning and calibration module includes a deviation attribution unit, which is used to determine the main source category of performance deviation based on the comparison results and market context data when the strategy is executed;

[0128] The main sources include: basic trend prediction error, unidentified external event impact, event impact estimation bias, and strategy execution environment anomalies.

[0129] Categories of deviation sources and examples of judgment:

[0130] Basic trend forecasting error: This category is defined as a deviation caused solely by the forecasting model's failure to accurately capture long-term / short-term demand trends, without any external unforeseen events. Example: If the seasonal increase in demand (e.g., summer beverage demand being 30% higher than predicted) was not predicted, and there were no external events such as typhoons or policy changes, this is the correct category.

[0131] Unidentified External Event Impact: This occurs because the system failed to capture a specific external event (such as a sudden policy change, natural disaster, or unexpected actions by a competitor), leading to a discrepancy between predictions and actual results. Example: During strategy execution, a sudden regional logistics control measure (which the system failed to identify) caused replenishment delays and increased stockout rates, thus falling into this category.

[0132] Event Impact Estimation Bias: The system identifies an external event, but the estimated impact differs from the actual impact, resulting in bias. Example: If the system identifies a Double Eleven promotional event and estimates a 40% increase in demand, but the actual increase is only 20%, leading to overstocking and inventory backlog, this event would be classified as this type.

[0133] Abnormal Strategy Execution Environment: The strategy itself is reasonable, but anomalies occur during execution (such as warehouse operation errors, logistics delays, or data statistics errors), leading to deviations in actual performance. Example: The strategy requires replenishing 1000 boxes on the 5th of each month, but the warehouse misses sending 200 boxes, causing an increase in the stockout rate. This would be classified as this category.

[0134] The deviation attribution unit judgment logic is as follows: first, check if there are any unidentified external events; then, check if there are any abnormalities in the execution environment; next, check if the impact of the event is biased; and finally, determine the error of the basic trend prediction.

[0135] The closed-loop learning and calibration module selectively sends parameter update instructions to the context awareness and feature extraction module, the prediction model in the demand prediction and simulation coupling module, or the strategy knowledge base, based on the main source category.

[0136] Example of parameter update command:

[0137] If the deviation originates from the error in the basic trend forecast, then a command is sent to the demand forecasting and simulation coupling module to adjust the seasonality factor of the forecasting model (e.g., adjust the summer demand factor from 1.2 to 1.5) to improve the trend capture capability.

[0138] If the deviation originates from the unidentified impact of external events, an instruction is sent to the context awareness and feature extraction module: add a new regional logistics policy data source to improve the external event identification rate (e.g., connect to the local logistics supervision platform to obtain control information in real time).

[0139] If the deviation originates from an event impact estimation bias, then send an instruction to the context awareness and feature extraction module: adjust the impact decay function parameter of the promotional event (e.g., adjust it from 0.8 to 0.9 to extend the impact duration).

[0140] If the deviation originates from an abnormal strategy execution environment, an instruction is sent to the strategy knowledge base: add an execution exception record field, record the warehouse missed delivery event in the knowledge base, and add execution verification constraints when generating subsequent strategies (such as providing outbound confirmation information within 24 hours after replenishment).

[0141] A new closed-loop learning and calibration mechanism addresses the pain points of traditional systems, such as the lack of self-optimization capabilities and persistent discrepancies between predictions and actual results. By attributing deviations, the core sources of performance deviations are accurately identified, avoiding blind parameter adjustments and improving the targeted nature of system calibration. Update instructions are sent to the corresponding modules based on the source of deviations, achieving a closed-loop iteration of performance monitoring, deviation attribution, parameter calibration, and performance improvement. This continuously optimizes feature extraction accuracy, prediction model accuracy, and the completeness of the strategy knowledge base, enabling the system to adapt to long-term dynamic changes in the supply chain and gradually improve the reliability and stability of inventory optimization and replenishment forecasting.

[0142] The structured supply chain context feature vector includes at least two of the following quantitative indicators:

[0143] The parameters of the event influence decay function are derived from event text recognition and historical impact analysis. The influence of external events (such as promotions, policies, and natural disasters) on supply chain demand will gradually weaken over time. The decay function is used to quantify this change. The core parameter is the decay coefficient (with a value of 0-1, the closer to 1, the slower the decay).

[0144] Example:

[0145] For promotional events (such as the 618 shopping festival), analysis of historical data (demand changes over the past three promotions) yields a decay function of y = 100 × 0.8. t (t is the number of days), then the decay coefficient is 0.8 (core parameter), which means that the influence of the promotion is 100 on the first day, 80 on the second day, 64 on the third day, and decays sequentially.

[0146] The logistics control event has an attenuation coefficient of 0.9, indicating a slow decay of its impact (due to the long duration of control). The function is y = 80 × 0.9. t .

[0147] The competitive cross-elasticity estimation coefficient is derived from competitor pricing and activity information analysis. It measures the impact of competitors' price / activity changes on the company's product demand. The formula is: Company's demand change rate / Competitor's price / activity change rate. A positive value indicates that an increase in competitor's price / decrease in activity leads to an increase in the company's demand. Example:

[0148] If competitor A lowers its price by 10% (price change rate -10%), and the demand for our product increases by 3% (demand change rate +3%), then the cross elasticity coefficient is 3% / (-10%) = -0.3 (a negative value indicates that the competitor lowers its price and our demand decreases).

[0149] If competitor B stops its promotion (activity change rate -100%), and the company's demand increases by 15% (demand change rate +15%), then the cross elasticity coefficient is 15% / (-100%) = -0.15. The larger the absolute value of the coefficient, the more significant the competitor's influence on the company.

[0150] The sentiment transmission strength index, derived from regional social media sentiment analysis, quantifies the emotional inclination (positive, negative, neutral) of users towards a company's products / industry on regional social media platforms (such as Douyin, Xiaohongshu, and local lifestyle platforms), as well as the speed and scope of its transmission within the region, into a value between 0 and 1 (the closer to 1, the stronger the transmission strength). Example:

[0151] On a certain region's Xiaohongshu platform, positive reviews of this company's products account for 80%, negative reviews 10%, and neutral reviews 10%, with a 20% repost rate. Therefore, the emotional transmission intensity index is: positive emotional percentage × repost rate + neutral emotional percentage × 0.5 × repost rate, i.e., 0.8 × 0.2 + 0.1 × 0.5 × 0.2 = 0.17.

[0152] If negative emotions erupt (60% of comments are negative and 30% are forwarded), the indicator is 0.3 (negative emotions are transmitted, the intensity is low but vigilance is needed).

[0153] The core quantitative indicators of the structured context feature vector are clearly defined, which solves the problems of traditional feature vectors being single-dimensional, insufficiently quantified, and unable to accurately reflect the supply chain context. Through three core indicators—event influence, competitive relationship, and user sentiment—the system comprehensively covers key internal and external influencing factors of the supply chain, improving the comprehensiveness and accuracy of context features. The quantitative indicators can be directly used for demand forecasting, simulation, and weight adjustment, providing high-quality and computable feature inputs for subsequent modules, further improving the accuracy of system prediction and optimization, and enabling strategies to better adapt to complex and ever-changing supply chain contexts.

[0154] Each replenishment strategy in the recommended strategy set is accompanied by a multi-dimensional evaluation label derived from the simulation and sensitivity analysis unit. These labels include: the expected profit under the baseline scenario, the guaranteed profit level under the worst-case scenario, and identifiers of the top N uncertainty factors that contribute most to profit volatility. These labels help users (such as inventory managers and supply chain decision-makers) quickly determine the risk-reward characteristics of the strategy.

[0155] The baseline scenario is a normal market situation with stable demand and no sudden external interference. The worst-case scenario is an unfavorable situation that is considered possible but has a low probability, based on historical data and scenario analysis. The guaranteed profit level is the minimum profit that the strategy can achieve under this scenario (avoiding losses), and it is used for risk assessment. Uncertainty factor identifiers: Various scenario characteristics that affect profit fluctuations are uniformly coded and identified. Examples: F1 = Market demand fluctuations, F2 = Competitor prices, F3 = Raw material prices, F4 = Logistics timeliness, F5 = Promotional intensity, F6 = Policy changes, facilitating users to quickly identify risk points. The value of N: Can be configured according to business needs. Common values ​​are 3 or 5 (the top 3 / 5 high-contribution factors). Small enterprises can set it to 2, while large enterprises (multiple categories, multiple regions) can set it to 5-8.

[0156] By using multi-dimensional evaluation tags, we provide comprehensive and quantitative benefit-risk assessment for recommended strategies, addressing the pain points of traditional recommendation strategies that only provide basic performance data, lack risk warnings, and make decision-making difficult. The tags cover regular benefits, extreme risks, and core influencing factors, helping decision-makers quickly grasp the advantages, risks, and applicable scenarios of each strategy, adapting to the risk tolerance and operational priorities of different enterprises. At the same time, the standardized coding of uncertainty factors makes it easy for decision-makers to quickly locate the core risks, formulate countermeasures in advance, improve the efficiency and scientific nature of replenishment strategy decisions, and ensure that the final selected strategy can balance benefits and risks and meet the actual operational needs of enterprises.

[0157] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. An adaptive AIGC inventory intelligent optimization and replenishment strategy prediction system, characterized in that, include: The collaborative optimization control module, and the context awareness and feature extraction module, demand prediction and simulation coupling module, and strategy generation and optimization module that are connected to it in communication; The context awareness and feature extraction module includes a large language model fine-tuned for the supply chain domain, used to process internal and external data and output structured supply chain context feature vectors. The demand prediction and simulation coupling module is used to receive the situation feature vector and candidate policy parameters from the policy generation and optimization module, and to perform conditional demand prediction and parallel simulation and sensitivity analysis of candidate policies. The strategy generation and optimization module includes a generative artificial intelligence model, which is used to generate and filter structured replenishment strategies based on prediction and simulation results. The collaborative optimization control module includes a dynamic feature attention modulation unit, which coordinates the demand prediction and simulation coupling module and the strategy generation and optimization module to perform iterative optimization, and adjusts the weight of the context feature vector in the system decision-making in reverse through the modulation unit based on the sensitivity analysis results obtained from the simulation.

2. The adaptive AIGC inventory intelligent optimization and replenishment strategy prediction system according to claim 1, characterized in that, The demand forecasting and simulation coupling module includes: The counterfactual forecasting unit is used to dynamically adjust the input or parameters of the forecasting model based on the future inventory and market action parameters corresponding to the specific candidate replenishment strategy received, and to generate conditional demand forecasts and forecast interval distributions under that strategy. The simulation and sensitivity analysis unit is used to perform parallel Monte Carlo simulations of multiple candidate replenishment strategies under various preset market scenarios based on the conditional demand forecast and forecast interval distribution, output the performance probability distribution of each strategy, calculate the contribution of each scenario feature dimension to the performance index based on the simulation data, and generate a sensitivity analysis report.

3. The adaptive AIGC inventory intelligent optimization and replenishment strategy prediction system according to claim 2, characterized in that, The strategy generation and optimization module includes: The structured strategy generation unit is used to generate a set of candidate replenishment strategies represented by a data structure of inventory units, replenishment time points, and replenishment quantities, based on the current inventory status, business constraints, and forecast inputs. The risk perception pre-screening unit, connected to the strategy knowledge base, is used to match the generated candidate replenishment strategies with the high failure risk strategy patterns recorded in the strategy knowledge base, and filter out candidate strategies whose comprehensive similarity on the key dimensions exceeds a preset risk threshold.

4. The adaptive AIGC inventory intelligent optimization and replenishment strategy prediction system according to claim 1, characterized in that, The multi-round iterative optimization process executed by the collaborative optimization control module includes: Initiate the first round of optimization to obtain an initial Pareto frontier that includes at least one replenishment strategy; Determine whether there exists a first strategy and a second strategy in the initial Pareto front that have a difference greater than a threshold on at least two preset performance objectives; If so, then trigger the iteration: guide the demand forecasting and simulation coupling module to generate demand forecasting intervals at different quantile levels for the first strategy and the second strategy; guide the strategy generation and optimization module to generate new candidate strategies in the parameter neighborhood of the first strategy and the second strategy; and perform simulation evaluation on the new candidate strategies; The Pareto front is updated based on the evaluation results until the number of iterations reaches the upper limit or the overall improvement of the updated Pareto front on all preset performance objectives is lower than the preset tolerance. The iteration is then terminated and the final Pareto front is output as the recommended strategy set.

5. The adaptive AIGC inventory intelligent optimization and replenishment strategy prediction system according to claim 2, characterized in that, The operation of the dynamic feature attention modulation unit includes: Receive the sensitivity analysis report output by the simulation and sensitivity analysis unit; Based on the report, a set of dynamic weight coefficients corresponding one-to-one with each dimension of the context feature vector were calculated. The dynamic weight coefficients are sent to the context awareness and feature extraction module to adjust the numerical scale of its output feature vector, and then sent to the counterfactual prediction unit to adjust the weighting of each context feature in the prediction model.

6. The adaptive AIGC inventory intelligent optimization and replenishment strategy prediction system according to claim 3, characterized in that, The strategy knowledge base stores records including: parameters of historically executed replenishment strategies, snapshots of market conditions before execution, and deviation data between actual and expected performance after execution. When the risk perception pre-screening unit performs matching, the key dimensions calculated include: the combination of inventory level and demand fluctuation, and the combination of promotion intensity and competitive situation.

7. The adaptive AIGC inventory intelligent optimization and replenishment strategy prediction system according to claim 2, characterized in that, The predictive model parameters dynamically adjusted by the counterfactual prediction unit include at least one of the following: demand price elasticity coefficient, sales loss rate due to inventory shortage, and demand satisfaction priority across supply chain nodes.

8. The adaptive AIGC inventory intelligent optimization and replenishment strategy prediction system according to claim 1, characterized in that, The system also includes: The closed-loop learning and calibration module is used to monitor the actual performance of the implemented replenishment strategy and compare it with the performance predicted by simulation. The closed-loop learning and calibration module includes a deviation attribution unit, which is used to determine the main source category of performance deviation based on the comparison results and market context data when the strategy is executed; The main sources include: basic trend prediction error, unidentified external event impact, event impact estimation bias, and strategy execution environment anomalies. The closed-loop learning and calibration module selectively sends parameter update instructions to the context awareness and feature extraction module, the prediction model in the demand prediction and simulation coupling module, or the strategy knowledge base, based on the main source category.

9. The adaptive AIGC inventory intelligent optimization and replenishment strategy prediction system according to claim 1, characterized in that, The structured supply chain context feature vector includes at least two of the following quantitative indicators: Parameters of the event influence decay function derived from event text recognition and historical impact analysis; Estimated coefficients of competitive cross-elasticity derived from analysis of competitor pricing and activity information; An index of the intensity of emotion transmission derived from regionalized social media sentiment analysis.

10. The adaptive AIGC inventory intelligent optimization and replenishment strategy prediction system according to claim 4, characterized in that, Each replenishment strategy in the recommended strategy set is accompanied by a multi-dimensional evaluation label derived by the simulation and sensitivity analysis unit. The label includes: the expected profit in the baseline scenario, the guaranteed profit level in the worst-case scenario, and the identifiers of the top N uncertainty factors that contribute the most to profit fluctuation.

Images