AI-driven supply chain sales prediction and intelligent inventory optimization system and method

Abstract

The invention provides an AI-driven supply chain sales prediction and intelligent inventory optimization system and method, and the system comprises a data preprocessing module which is used for docking with a supply chain order system; the sales prediction module adopts a deep neural network with an adaptive sparseness evolution mechanism; the inventory optimization decision-making module adopts a qualification trace-based streaming reinforcement learning architecture and abandons experience playback cache and a target network structure; the expert model auxiliary module adopts a confidence-based three-level intervention mechanism; and the strategy feedback updating module adopts an incremental parameter updating strategy to realize prediction-optimization-feedback closed-loop iteration. According to the method, a streaming reinforcement learning method is adopted, a traditional experience playback structure is abandoned, and real-time decision making and continuous optimization in a high-frequency order scene are achieved; a sparse sensing network and a confidence regulation mechanism are introduced, and the adaptability of the system to cold start, small samples and jump scenes is enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent supply chain management, and particularly to an AI-driven supply chain sales prediction and intelligent inventory optimization system and method. BACKGROUND

[0002] In current supply chain management systems, sales prediction and inventory decision-making widely adopt traditional time series analysis methods such as moving average method, exponential smoothing method, ARIMA, autoregressive model, etc. These methods have certain effects in static environment, but are difficult to cope with complex situations such as market environment fluctuation, sudden promotion activities or logistics constraints. In addition, although conventional machine learning such as random forest, gradient boosting tree (such as XGBoost) or multilayer perception network has certain prediction accuracy, it lacks adaptation to real-time and feedback requirements of business systems due to untimely updating.

[0003] In recent years, deep reinforcement learning (DRL) methods have been gradually used in supply chain strategy modeling, but mainstream methods generally rely on experience replay and target network mechanisms, which are essentially retrospective policy iteration based on offline batch data training. This structure has a series of problems such as large computational resource consumption, unstable policy, long update cycle and complex deployment when facing high-frequency, low-latency and non-stationary distribution data.

[0004] In dynamic and high-frequency supply chain systems, the order system continuously generates new data streams. If there is a lack of online updating and feedback adaptation mechanism, the prediction system will quickly fail, and the inventory strategy will also have response lag or structural errors, resulting in consequences such as inventory accumulation, frequent stockouts, and decline in customer satisfaction.

[0005] Therefore, there is an urgent need for an intelligent sales prediction and inventory optimization method with adaptive updating capability, without experience replay, which can be directly deployed in real-time order data stream environment, especially in high-dimensional multi-category and high-volatility scenarios with good stability, scalability and resource efficiency. SUMMARY

[0006] To overcome the shortcomings of the prior art, the AI-driven supply chain sales prediction and intelligent inventory optimization system and method are proposed, which can model and optimize the inventory holding cost, stockout loss and replenishment expenditure through an instant reward function, realizing balanced decision-making of multiple targets. More importantly, the system successfully solves the limitations of pure data-driven methods in special scenarios through the rule guidance mechanism introduced by the expert model module. In typical situations such as holidays, promotional activities or sales jumps, the system can automatically call expert rules for strategy correction and risk avoidance, significantly improving the response speed and business explanation ability to structural events.

[0007] To achieve the above purpose, the AI-driven supply chain sales prediction and intelligent inventory optimization system is proposed, which comprises: A data preprocessing module is used to connect the supply chain order system, collect structured information such as sales quantity, product identification, inventory status, sales date, perform missing value processing and standardization conversion, and build a sliding time window feature sequence with overlapping intervals. A sales prediction module adopts a deep neural network with an adaptive sparsity evolution mechanism, with an initial sparsity rate of more than 90%, dynamically adjusting the network connection density by monitoring the convergence speed of the prediction error, and supporting a cache-free streaming online update mechanism. An inventory optimization decision module adopts a streaming reinforcement learning architecture based on eligibility traces, discarding the experience replay cache and target network structure, and solving the gradient vanishing and distribution drift problems under high-frequency data flow through a double-time-scale update mechanism. An expert model auxiliary module adopts a three-level intervention mechanism based on confidence, including a rule engine unit, a special situation correction unit and a confidence adjustment unit, and realizes adaptive fusion of model output and expert rules through Bayesian confidence scoring. A strategy feedback update module adopts an incremental parameter update strategy, realizes the closed-loop iteration of prediction-optimization-feedback through sliding window error accumulation and adaptive learning rate adjustment.

[0008] Further, the adaptive sparsity evolution mechanism of the sales prediction module comprises: A sparsity dynamic adjustment strategy adaptively adjusts according to the prediction error convergence speed: ;

[0009] Wherein, is the network sparsity at time t, is the minimum sparsity, is the maximum sparsity, is the time decay constant, ε is the error convergence threshold, and ΔE is the error reduction rate in the sliding window. The connection importance evaluation is based on the weighted combination of weight amplitude, gradient contribution and activation frequency, and the connections with top k% importance scores are dynamically preserved to realize online adaptive evolution from sparse to dense.

[0010] Further, the inventory optimization decision module adopts a double-time-scale stream reinforcement learning update mechanism: Fast time scale update (1-10 milliseconds): ; Slow time scale update (100-1000 milliseconds): ;

[0011] Where θ fast and θ slow are the fast and slow update parameter sets, respectively, with values [0.01, 0.1], with values [0.001, 0.01], k is the cumulative window size, is the time difference error, is the eligibility trace; Through parameter separation and asynchronous update, the fast time scale captures the immediate gradient information, and the slow time scale maintains the stability of the strategy, solving the distribution drift problem in stream learning.

[0012] Further, the immediate reward function of the inventory optimization decision module adopts a segmented nonlinear design: When : ; When : ; When : ; When (out of stock): ; Where: : the immediate reward at time t; : the actual sales volume; : the end-of-period inventory; : the safety stock level; : the maximum inventory capacity; , , : the segmented inventory cost function; : the unit sales profit; : penalty coefficient of broken goods; : penalty coefficient of overstock; : predicted sales at time t; : current inventory; : replenishment action.

[0013] The nonlinear modeling of inventory cost is achieved by segment design, and different cost strategies are adopted at different inventory levels.

[0014] Further, the inventory optimization strategy network enhances credit allocation by introducing a momentum eligibility trace mechanism: ; ;

[0015] wherein, is an enhanced eligibility trace, is a momentum term, μ∈[0.1, 0.3] is a momentum influence coefficient, ν∈[0.9, 0.99] is a momentum decay rate, γ is a reward discount factor, and λ is an eligibility trace decay coefficient; The introduction of the momentum term accelerates the propagation of credit allocation and improves the learning ability of long-term dependencies; at the same time, adaptive gradient clipping is used to prevent gradient explosion.

[0016] Further, the expert model auxiliary module adopts a Bayesian confidence fusion mechanism: The model confidence is calculated based on the prediction uncertainty: ;

[0017] The fusion output is weighted using Bayesian: ;

[0018] wherein, : model confidence; : model prediction output value; : historical sales mean; : historical sales standard deviation; : prediction uncertainty; : maximum allowed uncertainty; : rule confidence; : model output; : rule output; When the model confidence is lower than the threshold θ c , automatically raise the rule weight, realize the intelligent fusion of model and expert knowledge.

[0019] Further, the policy feedback update module adopts an adaptive learning rate adjustment mechanism: ;

[0020] Wherein, is the adaptive learning rate at time t, =10 -4 , =10 -2 , is the recent error variance, is the adjustment sensitivity; Adjust the learning rate dynamically through the error variance, reduce the learning rate to keep stable when the variance is high, and increase the learning rate to speed up the convergence when the variance is low, realize adaptive optimization.

[0021] The AI-driven supply chain sales forecasting and intelligent inventory optimization method is applicable to the system, comprising the following steps: S1: Construct overlapping sliding window to collect order data, window length T, overlap rate greater than 50%; S2: Predict sales through adaptive sparse network, dynamically adjust sparsity according to the formula; S3: Calculate the replenishment strategy based on double time scale streaming reinforcement learning, separate update of fast and slow time scales; S4: Apply Bayesian confidence to fuse expert rules and model output; S5: Execute replenishment decision and collect feedback information; S6: Update system parameters through adaptive learning rate mechanism to form a closed loop optimization.

[0022] Further, the distribution drift problem of streaming learning in step S3 is solved by using importance sampling correction: ;

[0023] Wherein, is the corrected time difference error, is the original time difference error, and are the current and historical data distribution respectively, ∈[1.5, 3.0] is the clipping constant; Monitor the distribution change through KL divergence, when the KL divergence exceeds the threshold θ KL , trigger the slow time scale parameter recalibration to ensure the stability of the strategy in non-stationary environment.

[0024] Further, the system performance indicators are: Prediction accuracy: mean absolute percentage error (MAPE) ≤ 12%, reduced by 40% compared to traditional ARIMA methods; Inventory turnover rate: increased by more than 30%; Out-of-stock rate: reduced by more than 50%; System response: end-to-end delay ≤ 50 milliseconds, supporting millisecond-level parameter updates; Resource occupation: memory ≤ 100 MB, model parameter quantity ≤ 100K; Scalability: supporting 10000+ SKU parallel processing, single update complexity O(s·d), where s is the number of sparse connections and d is the network depth.

[0025] Compared with the prior art, the beneficial effects of the present application are: 1. The present application provides an AI-driven supply chain sales prediction and intelligent inventory optimization system and method, which supports real-time prediction and strategy updating in the scenario of continuous order data inflow, adopts a streaming reinforcement learning architecture, discards the traditional batch processing modeling method, does not need to rely on experience replay cache and target network structure, has high-frequency order processing capability and dynamic strategy adaptation capability, and can cope with complex business environments such as intensive promotion, e-commerce high concurrency, and strong volatility.

[0026] 2. The present application provides an AI-driven supply chain sales prediction and intelligent inventory optimization system and method, which adopts a sparse initialized multi-layer perceptron model structure, combines sliding window features, periodic time factors, and product embedding vectors, and improves the training stability and generalization ability of the model under cold start, small sample, and high-dimensional input conditions. At the same time, the system structure is lightweight, the computational resource consumption is low, and it has the feasibility of direct deployment in edge nodes or enterprise private cloud environments, adapting to the needs of multi-terminal parallel deployment.

[0027] 3. The present application provides an AI-driven supply chain sales prediction and intelligent inventory optimization system and method, which builds a complete closed-loop path covering sales prediction, inventory optimization, expert intervention, and feedback updating, models the inventory holding cost, out-of-stock loss, and replenishment expenditure through an immediate reward function, and introduces a rule-guided mechanism through an expert model module to automatically correct and avoid risks in typical situations such as holidays, promotional activities, or sales jumps, significantly improving the system's response speed and business explanation ability to structural events. BRIEF DESCRIPTION OF DRAWINGS

[0028] In order to more clearly illustrate the technical solutions in the specific embodiments or prior art of the present application, the drawings required to be used in the description of the specific embodiments or prior art will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0029] Figure 1 is a schematic diagram of the system architecture of the present application; Figure 2 is a state-action-reward modeling flowchart of the inventory optimization strategy module. Figure 3 is a schematic diagram of the data processing and prediction flow of the sales forecasting module. DETAILED DESCRIPTION

[0030] The technical solutions of the present application will be described in more detail and completely below by combining the drawings and through the description of the preferred embodiments of the present application.

[0031] As shown in Figure 1 , the entire system is composed of five core modules, forming a highly integrated intelligent decision-making system. The data preprocessing module directly interfaces with the enterprise supply chain order system, collecting structured information such as sales quantity, product identification, and inventory status in real time, and ensuring data quality through missing value processing and standardization conversion. The processed data flows into the feature construction and standardization processing unit, which uses a sliding window mechanism to construct time series features and ensures the continuity and integrity of the features through high overlap rate sampling.

[0032] The sales forecasting module is one of the core components of the system, and its detailed data processing and prediction flow is shown in Figure 3 . The original order data undergoes systematic multi-layer processing, first filling missing values and performing data standardization through methods such as moving average and linear interpolation, and then entering the sliding window builder to generate high-dimensional feature vectors. The feature extractor further extracts periodic features, embedding vectors, moving averages, trend slopes, and price variation coefficients, etc. key information, providing rich inputs for the prediction model. The system uses a sparse perception neural network for sales forecasting, with a network structure designed with multiple hidden layers decreasing layer by layer, and dynamically adjusting the network connection density through an adaptive sparsity evolution mechanism. This design allows the network to adaptively adjust the structure according to the convergence speed of the prediction error from the initial high sparsity state, significantly reducing the computational resource consumption while ensuring the prediction accuracy. The dynamic sparsity adjustment and error monitoring mechanism on the side ensures that the model can respond in real time to changes in data distribution, ultimately outputting high-precision sales prediction results for multiple days in the future.

[0033] The inventory optimization decision module adopts a streaming reinforcement learning architecture based on eligibility traces, and its state-action-reward modeling process is as follows: Figure 2 As shown, this module abandons the traditional experience replay caching and target network structure, and solves the gradient vanishing and distribution drift problems under high-frequency data streams through an innovative dual-timescale update mechanism. Multi-dimensional information such as predicted sales, current inventory, and time labels are input into the state space builder to generate high-dimensional feature vectors for use by the MDP policy model. The system adopts a fast-slow separation dual-timescale update mechanism: the fast timescale captures immediate gradient information, while the slow timescale maintains policy stability. The immediate reward function adopts a piecewise nonlinear design, comprehensively considering inventory costs, sales profits, stockout penalties, and overstocking penalty coefficients, achieving refined management of different inventory levels. The momentum qualification trace mechanism enhances the system's ability to learn long-term dependencies by strengthening credit allocation and accelerating momentum term propagation. The policy update module ensures policy stability in non-stationary environments through importance sampling correction and KL divergence monitoring of distribution drift.

[0034] The expert model assistance module employs a three-tiered intervention mechanism based on confidence levels, including a rule engine, confidence-weighted calculation, and Bayesian fusion. When the model's predictions have high uncertainty or encounter special business scenarios, the system automatically increases the weight of expert rules, achieving intelligent fusion of model output and expert knowledge. This design is particularly suitable for handling special situations with clear business rules, such as holiday promotions, new product launches, and seasonal fluctuations.

[0035] The policy feedback update module employs an incremental parameter update strategy, dynamically optimizing system performance through an adaptive learning rate adjustment mechanism. This module collects policy execution results from the execution layer, analyzes system performance through sliding window error accumulation, and feeds optimization signals back to each module. For example... Figure 1 The feedback loop design shown ensures that the system can continuously learn and improve. The feedback path runs from the execution layer through the policy feedback update module, connecting to the data preprocessing module and the feature construction unit, forming a complete closed-loop optimization mechanism. It is particularly noteworthy that the feedback signal propagates along the right edge of the system, avoiding cross-interference with other modules and ensuring the clarity and reliability of signal transmission.

[0036] The execution layer, acting as the system's output interface, distributes inventory optimization and order response strategies to the purchasing and warehousing systems. Through a dual-timescale rapid update and slow-stabilization mechanism, the system can quickly respond to market changes while maintaining decision-making stability. The low-latency response mechanism ensures that the end-to-end processing time from data input to decision output is controlled at the millisecond level, meeting the real-time decision-making requirements of high-frequency order scenarios.

[0037] The entire system can continuously optimize performance in complex and ever-changing supply chain environments. This is achieved through adaptive evolution of sparse neural networks, dual-timescale updates via streaming reinforcement learning, intelligent fusion of expert knowledge, and incremental parameter optimization.

[0038] As a specific example, this document describes the implementation of a supply chain order system-driven adaptive sales forecasting and inventory optimization system in the apparel category of a large e-commerce platform. The system processes real-time order data for 10,000 SKUs, is deployed on a server cluster equipped with Intel Xeon Gold 6248R processors and 128GB DDR4 memory, uses PostgreSQL to store historical data, and Redis to cache real-time streaming data.

[0039] The data preprocessing module directly interfaces with the enterprise ERP system's order data interface, fetching new order data in batches every 5 minutes. The collected raw data includes nine core fields: order number, SKU identifier, sales quantity, sales timestamp, unit price, current inventory level, promotional flag, seasonal code, and holiday identifier. For missing values, missing sales quantities are filled with the moving average of the previous and next three days; missing inventory data is filled with linear interpolation; and missing prices are filled with the most recent valid price for the same SKU.

[0040] The sliding window feature construction uses a time window of length T=14 days, sliding in 7-day increments to achieve a 50% overlap rate. Within each window, a 15-dimensional feature vector is extracted, including a 14-dimensional daily sales volume sequence, and mean features of three scales: moving averages MA(3), MA(7), and MA(14), as well as sales volume variance, trend slope, weekend sales percentage, promotional day percentage, average inventory level, inventory turnover rate, price variation coefficient, seasonal factors, and holiday impact factors. The trend slope is calculated using the least squares method. ;

[0041] Where n is the window length, which is 14. For the first Daily sales volume. Weekend sales percentage: the proportion of total sales volume on Saturdays and Sundays within the calculation window. Price coefficient of variation: defined as the ratio of the price standard deviation to the mean, reflecting the degree of price volatility.

[0042] The sales forecasting module employs an adaptive sparse deep neural network. The network structure includes an input layer that receives a 210-dimensional flattened feature vector, three hidden layers with 256, 128, and 64 nodes respectively, using the ReLU activation function, and an output layer that generates sales forecasts for the next 7 days using a linear activation function. Network initialization uses the Xavier method, with weights following a mean of 0 and a variance of 0. The normal distribution, where and are the input and output dimensions of the layer, respectively, and the bias term is initialized to 0.

[0043] The network is trained using the adaptive moment estimation optimizer Adam.

[0044] In the network sparsity dynamic adjustment strategy, the initial sparsity is set to 0.95, the minimum sparsity is set to 0.5, the sparsity time decay constant is set to 1000, and the error convergence threshold is set to 0.01. The connection importance score uses a comprehensive index: ;

[0045] where is the connection weight, is all the weight matrices, is the loss function, is the activation value is the total number of activations. The network structure is re-evaluated and adjusted every 100 iterations.

[0046] The state space of the inventory optimization decision module is designed as a 31-dimensional vector, including the current inventory level $I_t$, future 7-day sales forecast, historical 7-day actual sales, historical 7-day prediction error, current order in transit, supplier average lead time LT, lead time standard deviation, current date's week encoding (1-7), and month encoding (1-31). The action space is discretized into 20 levels, with replenishment actions uniformly distributed from 0 to the maximum replenishment capacity.

[0047] The Q-network uses a duel network architecture, with shared layers containing two fully connected layers (128 and 64 nodes), and value and advantage streams each containing a 32-node hidden layer. In the double time scale update mechanism, the fast learning rate is set to 0.05, and the slow learning rate is set to 0.005, and the cumulative window size k = 50. The fast time scale time difference error and the slow time scale time difference error According to the Bellman equation, the eligibility trace decay coefficient λ is set to 0.92, and the reward discount factor γ is set to 0.95.

[0048] The KL divergence threshold for distribution drift detection is set to 0.1, and the importance sampling clipping constant is set to 2.0. The state distribution is estimated through a circular buffer with a capacity of 10000, using kernel density estimation, with bandwidth h determined by the Silverman rule.

[0049] Safety stock level in the immediate reward function Set to 14-day average sales, calculated by formula ;

[0050] Dynamically calculated, where daily average demand and demand standard deviation are calculated based on historical 30-day data, with a 95% service level corresponding to the standard normal quantile. Maximum inventory capacity Set to 30-day average sales. Actual sales are the minimum of forecasted sales and available inventory . End-of-period inventory .

[0051] In the piecewise inventory cost function, the base holding cost is set to 0.02% of the item's value per day. The low inventory risk coefficient is set to 0.5, and the high inventory penalty coefficient is set to 2.0. The unit sales profit p is dynamically set according to the item's gross margin, with an average of 35% of the selling price for clothing categories. The out-of-stock penalty coefficient a is set to 1.5 times the average selling price, and the overstock penalty coefficient b is set to 10 times the average selling price.

[0052] In the momentum eligibility trace mechanism, the momentum influence coefficient μ is set to 0.2, and the momentum decay rate is set to 0.95. The gradient of the Q function with respect to the parameters is calculated by backpropagation. In the importance sampling correction, the policy probability ratio is calculated by the Boltzmann distribution: ;

[0053] where the temperature parameter is set to 1.0, controlling the degree of exploration.

[0054] The expert model assistance module contains 120 business rules. In the Bayesian belief fusion, the historical sales mean and standard deviation are calculated based on 90-day data, updated daily. The prediction uncertainty is estimated by Monte Carlo Dropout, performing M=10 forward propagations. The maximum allowed uncertainty is set to 2 . The model confidence threshold is set to 0.6. The rule confidence has an initial value of 0.7, dynamically adjusted according to historical accuracy, with a range limit of [0.3, 0.95].

[0055] The minimum value of the adaptive learning rate in the policy feedback update module. Set to 10 -4 maximum value Set to 10 -2 The sensitivity parameter ρ for learning rate adjustment r Set to 2.0. Recent Error Variance Calculated based on the most recent 100 samples. The target value for the gradient norm is set to 10, and pruning is performed when the actual gradient norm exceeds this value.

[0056] New product processing employs Bayesian Beta-Binomial conjugate priors, prior parameters , indicating weak prior information. In similarity calculation, price similarity uses a Gaussian kernel with a bandwidth set to 0.5 times the standard deviation of prices; category similarity uses the Jaccard coefficient; and attribute similarity uses cosine similarity. The weights for the three are 0.3, 0.4, and 0.3, respectively. The predicted value for the first 30 days is multiplied by a safety factor of 0.7.

[0057] The system adopts an event-driven architecture, with a message queue batch size of 32 messages and a timeout of 100 milliseconds. The exponentially weighted moving average of the monitoring system is set to a weight coefficient of 0.95, and the 3-sigma control limit is calculated based on 1000 historical data points.

[0058] During model updates, the incremental learning batch size is set to 1000 samples, with a trigger interval of 24 hours. The temperature parameter for knowledge distillation is set to 3.0, and the distillation loss weights are... Set to 0.3. The teacher model is the best performing model in the last 7 days, and the student model is a newly trained model.

[0059] Concept drift detection uses the ADWIN algorithm with a significance level of 0.05 and a sliding window size of 500. When drift is detected, the learning rate is temporarily increased by 50% (i.e., multiplied by 1.5), and then restored after 100 iterations.

[0060] During promotional periods, the predictive model switches to a special parameter set, specifically trained based on historical promotional data. The promotional effectiveness enhancement factor is determined according to the promotion type: 1.8 times for discounts exceeding a certain amount, 2.2 times for full reductions, and 1.5 times for buy-one-get-one-free promotions. The model update frequency is increased to three times the normal value, i.e., updated every 8 hours.

[0061] The seasonal factor matrix is ​​12×50 dimensions (12 months × 50 product subcategories), with each element representing the sales adjustment coefficient for that product category in that month. For down jackets, the factor is 2.5 from October to February and 0.2 from June to August; for T-shirts, the factor is 1.8 from June to August and 0.6 from December to February.

[0062] The trigger threshold for the backup supplier is 3 days of delay for the main supplier. The unstable supplier criterion is a standard deviation of delivery time over the last 30 days greater than 2 days, in which case the safety stock is increased by 30%, i.e. multiplied by 1.3.

[0063] In actual operation, the system runs continuously for 30 days in a test environment containing 1000 SKUs, processing 1.5 million orders per day on average. The prediction accuracy MAPE is stable at 11.8% ± 1.2%, 10.5% on weekdays, 13.8% on weekends, and 15.2% during promotional periods. The inventory turnover rate has increased from 8.5 times per year to 11.2 times per year. The out-of-stock rate has decreased from 4.2% to 2.0%. The system's end-to-end average delay is 45 milliseconds, with a P95 delay of 47 milliseconds and a P99 delay of 48 milliseconds. The memory usage is 85 MB, of which 38 MB is model parameters, 35 MB is data cache, and 12 MB is runtime memory. The total number of model parameters is 98K, with 61K in the prediction network and 37K in the decision network.

[0064] The above detailed description merely describes the preferred embodiment of this application, and is not intended for limiting the scope of protection of the application. Any variations, substitutions and improvements of the technical solutions of the application made by those of ordinary skill in the art based on the description and drawings provided by the application should fall within the protection scope of the application. The protection scope of the application is determined by the claims.

Claims

1. An AI driven supply chain sales forecast and intelligent inventory optimization system, characterized in that, The application relates to a sales forecasting and inventory optimization system based on deep learning and reinforcement learning, which comprises the following modules: A data preprocessing module for connecting a supply chain order system, collecting structured information, performing missing value processing and standardization conversion, and constructing a sliding time window feature sequence with overlapping intervals; A sales forecasting module adopting a deep neural network with an adaptive sparsity evolution mechanism, wherein an initial sparsity rate is set to be higher than 90%, a network connection density is dynamically adjusted by monitoring the convergence speed of a prediction error, and a cache-free streaming online updating mechanism is supported; An inventory optimization decision module adopting a streaming reinforcement learning architecture based on eligibility traces, wherein an experience replay cache and a target network structure are abandoned, a double-time-scale updating mechanism is adopted to solve the problems of gradient disappearance and distribution drift under a high-frequency data stream; An expert model auxiliary module adopting a three-level intervention mechanism based on confidence, comprising a rule engine unit, a special situation correction unit and a confidence adjustment unit, and realizing adaptive fusion of model output and expert rules through Bayesian confidence scoring; A strategy feedback updating module adopting an incremental parameter updating strategy, realizing a closed-loop iteration of prediction-optimization-feedback through sliding window error accumulation and adaptive learning rate adjustment.

2. The Al driven supply chain sales forecasting and intelligent inventory optimization system of claim 1, wherein, The adaptive sparsity evolution mechanism of the sales forecasting module comprises the following steps: A sparsity dynamic adjustment strategy is adopted to adaptively adjust the sparsity according to the convergence speed of the prediction error; ； wherein, is the network sparsity at time t, is the minimum sparsity, is the maximum sparsity, is the time decay constant, ε is the error convergence threshold, and ΔE is the error drop rate within the sliding window. The connection importance evaluation is based on the weighted combination of weight amplitude, gradient contribution and activation frequency, important connections with a score in the top k% are dynamically retained, and online adaptive evolution from sparsity to density is realized.

3. The Al driven supply chain sales forecasting and intelligent inventory optimization system as claimed in claim 1, wherein, The inventory optimization decision module adopts a double-time-scale streaming reinforcement learning updating mechanism: Fast time scale updates (1-10 milliseconds): ​ Slow time scale updates (100-1000 milliseconds): ​ where θ fast and θ slow are fast and slow update parameter sets, respectively, taking values in [0.01, 0.1], taking values in [0.001, 0.01], k is the cumulative window size, is the time-difference error, is the eligibility trace; Through parameter separation and asynchronous updating, the fast time scale captures instant gradient information, the slow time scale maintains the stability of the strategy, and the distribution drift problem in streaming learning is solved.

4. The Al driven supply chain sales forecasting and intelligent inventory optimization system of claim 1, wherein, The instant reward function of the inventory optimization decision module adopts a segmented nonlinear design: When Time: ; When Time: ; When Time: ; When (out of stock) : ; Wherein: : instant reward at time t : actual sales volume : ending inventory : safe stock level : maximum inventory capacity , , : piecewise inventory cost function : unit sales profit : penalty coefficient for a broken order : overstock penalty coefficient : predicted sales at time t : Current inventory : replenishment action; The segmented design realizes nonlinear modeling of inventory cost, and a differentiated cost strategy is adopted at different inventory levels.

5. The Al driven supply chain sales forecasting and intelligent inventory optimization system of claim 1, wherein, The inventory optimization strategy network enhances credit allocation by introducing a momentum eligibility trace mechanism: ； ； wherein, is an enhanced eligibility trace, is a momentum term, μ ∈ [0.1, 0.3] is a momentum influence coefficient, v ∈ [0.9, 0.99] is a momentum decay rate, γ is a reward discount factor, and λ is an eligibility trace decay coefficient; The introduction of the momentum term accelerates the propagation of credit allocation and improves the learning ability of long-term dependencies; meanwhile, adaptive gradient clipping is adopted to prevent gradient explosion.

6. The Al driven supply chain sales forecasting and intelligent inventory optimization system of claim 1, wherein, The expert model auxiliary module adopts a Bayesian confidence fusion mechanism: Model confidence is computed based on prediction uncertainty: ; The fusion output adopts Bayesian weighting: ; wherein, : model confidence; : model predicted output value; : historical sales average; : historical sales standard deviation; : prediction uncertainty; : maximum allowed uncertainty; : rule confidence; : model output; : rule output; When the model confidence is lower than the threshold θ c The rule weight is automatically promoted to realize the fusion of the model and the expert knowledge.

7. The Al driven supply chain sales forecasting and intelligent inventory optimization system of claim 1, wherein, The strategy feedback updating module adopts an adaptive learning rate adjustment mechanism: ； wherein, is the adaptive learning rate for time t, = 10 -4 , = 10 -2 , is the recent error variance, is the adjustment sensitivity; Through dynamic adjustment of the learning rate according to the error variance, the learning rate is reduced to maintain stability when the variance is high, and the learning rate is increased to accelerate convergence when the variance is low, so that adaptive optimization is realized.

8. The AI-driven supply chain sales forecasting and intelligent inventory optimization method according to any one of claims 1 to 7, characterized in that, The application comprises the following steps: S1: constructing an overlapping sliding window to collect order data, wherein the window length T is greater than 50%; S2: predicting sales through an adaptive sparse network, and dynamically adjusting the sparsity according to the formula; S3: calculating a replenishment strategy based on a double-time-scale streaming reinforcement learning, and separating and updating the fast and slow time scales; S4: applying Bayesian confidence to fuse expert rules and model output; S5: executing a replenishment decision and collecting feedback information; S6: updating system parameters through an adaptive learning rate mechanism to form a closed-loop optimization. 9.The AI-driven supply chain sales forecasting and intelligent inventory optimization method of claim 8, wherein, In step S3, the distribution drift problem in streaming learning is solved by adopting importance sampling correction. ； where, is the corrected time-difference error, is the original time-difference error, and are the current and historical data distributions, respectively, is a clipping constant; The distribution change is monitored by KL divergence, when KL divergence exceeds threshold θ KL Slow time scale parameter recalibration is triggered to guarantee the stability of the policy under non-stationary environment.

Images