Ai-driven supply chain sales forecasting and intelligent inventory optimization system and method

Abstract

The application provides an AI-driven supply chain sales prediction and intelligent inventory optimization system and method, comprising: a data preprocessing module for interfacing with a supply chain order system; a sales prediction module adopting a deep neural network with an adaptive sparsity evolution mechanism; an inventory optimization decision module adopting a stream reinforcement learning architecture based on eligibility traces, abandoning experience replay cache and target network structure; an expert model auxiliary module adopting a three-level intervention mechanism based on confidence; a strategy feedback update module adopting an incremental parameter update strategy, realizing a closed-loop iteration of prediction-optimization-feedback. The application adopts a stream reinforcement learning method, abandons a traditional experience replay structure, realizes real-time decision and continuous optimization in a high-frequency order scenario, introduces a sparse perception network and a confidence adjustment mechanism, and enhances the adaptability of the system to cold start, small sample and jump scenarios.

Description

Technical Field

[0001] This invention relates to the field of intelligent supply chain management technology, specifically to an AI-driven supply chain sales forecasting and intelligent inventory optimization system and method. Background Technology

[0002] In current supply chain management systems, sales forecasting and inventory decisions widely employ traditional time series analysis methods such as moving averages, exponential smoothing, ARIMA, and autoregressive models. While these methods are effective in static environments, they struggle to handle complex scenarios such as drastic market fluctuations, sudden promotional events, or logistical constraints. Furthermore, conventional machine learning methods, such as random forests, gradient boosting trees (e.g., XGBoost), or multilayer perceptron networks, while possessing some predictive accuracy, lack timely updates and fail to adapt to the real-time and responsive requirements of business systems.

[0003] In recent years, deep reinforcement learning (DRL) methods have been increasingly used in supply chain strategy modeling. However, mainstream methods generally rely on experience replay and target network mechanisms, which are essentially backtracking policy iterations based on offline batch data training. This structure suffers from a series of problems when dealing with high-frequency, low-latency, and non-stationary data, such as high computational resource consumption, policy instability, long update cycles, and complex deployment.

[0004] In dynamic, high-frequency supply chain systems, order systems constantly generate new data streams. Without online updates and feedback mechanisms, forecasting systems will quickly fail, and inventory strategies will suffer from lag or structural errors, leading to consequences such as inventory backlog, frequent stockouts, and decreased customer satisfaction.

[0005] Therefore, there is an urgent need for an intelligent sales forecasting and inventory optimization method that has adaptive update capabilities, requires no experience replay, and can be directly deployed in real-time order data stream environments, especially in high-dimensional, multi-category, and highly volatile scenarios, with good stability, scalability, and resource efficiency. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, this invention proposes an AI-driven supply chain sales forecasting and intelligent inventory optimization system and method. Through an instant reward function, the system can uniformly model and optimize inventory holding costs, stockout losses, and replenishment expenses, achieving balanced decision-making across multiple objectives. More importantly, the system successfully addresses the limitations of purely data-driven methods in specific scenarios through a rule-guided mechanism introduced by an expert model module. In typical situations such as holidays, promotional activities, or sales spikes, the system can automatically invoke expert rules for strategy correction and risk avoidance, significantly improving the response speed and business interpretability to structural events.

[0007] To achieve the above objectives, this invention proposes an AI-driven supply chain sales forecasting and intelligent inventory optimization system, comprising:

[0008] The data preprocessing module is used to connect with the supply chain order system, collect structured information such as sales quantity, product identification, inventory status, and sales date, perform missing value processing and standardization transformation, and construct a sliding time window feature sequence with overlapping intervals;

[0009] The sales forecasting module employs a deep neural network with an adaptive sparsity evolution mechanism. The initial sparsity is set to over 90%, and the network connection density is dynamically adjusted by monitoring the convergence speed of the prediction error. It also supports a bufferless streaming online update mechanism.

[0010] The inventory optimization decision module adopts a streaming reinforcement learning architecture based on qualification traces, abandons experience replay caching and target network structure, and solves the gradient vanishing and distribution drift problems under high-frequency data streams through a dual time scale update mechanism.

[0011] The expert model assistance 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. It achieves adaptive fusion of model output and expert rules through Bayesian confidence scoring.

[0012] The strategy feedback update module adopts an incremental parameter update strategy, and achieves a closed-loop iteration of prediction-optimization-feedback through sliding window error accumulation and adaptive learning rate adjustment.

[0013] Furthermore, the adaptive sparsity evolution mechanism of the sales forecasting module includes:

[0014] The sparsity dynamic adjustment strategy adaptively adjusts based on the convergence speed of the prediction error:

[0015] ;

[0016] in, Let be the network sparsity at time t. To minimize sparsity, To achieve maximum sparsity, Let ε be the time decay constant, ε be the error convergence threshold, and ΔE be the error decrease rate within the sliding window.

[0017] The importance assessment of connections is based on a weighted combination of weight magnitude, gradient contribution, and activation frequency. It dynamically retains the top k% of connections by importance score, enabling online adaptive evolution from sparse to dense.

[0018] Furthermore, the inventory optimization decision module employs a dual-timescale streaming reinforcement learning update mechanism:

[0019] Fast timescale updates (1-10 milliseconds): ;

[0020] Slow timescale updates (100-1000 milliseconds): ;

[0021] Where, θ fast and θ slow The parameter sets are updated quickly and slowly, respectively. The value can be [0.01, 0.1]. The value ranges from [0.001, 0.01], where k is the cumulative window size. For time difference error, For qualification traces;

[0022] By separating parameters and updating asynchronously, the system captures instantaneous gradient information on a fast timescale and maintains policy stability on a slow timescale, thus solving the distribution drift problem in streaming learning.

[0023] Furthermore, the immediate reward function of the inventory optimization decision module adopts a piecewise nonlinear design:

[0024] when hour: ;

[0025] when hour: ;

[0026] when hour: ;

[0027] when When out of stock: ;

[0028] in:

[0029] The immediate reward at time t;

[0030] Actual sales volume;

[0031] Ending inventory;

[0032] Safety stock levels;

[0033] Maximum inventory capacity;

[0034] , , Piecewise inventory cost function;

[0035] Unit sales profit;

[0036] Out-of-stock penalty coefficient;

[0037] Over-storage penalty coefficient;

[0038] Forecast sales at time $t$;

[0039] Current inventory;

[0040] Restocking activity.

[0041] Nonlinear modeling of inventory costs is achieved through segmented design, and differentiated cost strategies are adopted under different inventory levels.

[0042] Furthermore, the inventory optimization strategy network enhances credit allocation by introducing a momentum eligibility trace mechanism:

[0043] ;

[0044] ;

[0045] in, To enhance qualification traces, For momentum, μ∈[0.1, 0.3] is the momentum influence coefficient, ν∈[0.9, 0.99] is the momentum decay rate, γ is the reward discount factor, and λ is the qualification trace decay coefficient;

[0046] The introduction of the momentum term accelerates the propagation of credit allocation and enhances the learning ability of long-term dependencies; at the same time, adaptive gradient pruning is used to prevent gradient explosion.

[0047] Furthermore, the expert model assistance module employs a Bayesian confidence fusion mechanism:

[0048] Model confidence is calculated based on prediction uncertainty: ;

[0049] The fused output uses Bayesian weighting: ;

[0050] in, Model confidence;

[0051] Model predicted output value;

[0052] Historical average sales;

[0053] : Standard deviation of historical sales;

[0054] : Predicting uncertainty;

[0055] Maximum permissible uncertainty;

[0056] Rule confidence;

[0057] Model output;

[0058] : Rule output;

[0059] When the model confidence level is lower than the threshold θ c In this way, the rule weights are automatically increased, enabling intelligent integration of the model and expert knowledge.

[0060] Furthermore, the policy feedback update module employs an adaptive learning rate adjustment mechanism:

[0061] ;

[0062] in, Let be the adaptive learning rate at time t. =10 -4 , =10 -2 , The recent error variance To adjust sensitivity;

[0063] By dynamically adjusting the learning rate based on the error variance, the learning rate is reduced to maintain stability when the variance is high, and increased to accelerate convergence when the variance is low, thus achieving adaptive optimization.

[0064] An AI-driven supply chain sales forecasting and intelligent inventory optimization method, applicable to the aforementioned system, includes the following steps:

[0065] S1: Construct an overlapping sliding window to collect order data, with a window length of T and an overlap rate greater than 50%;

[0066] S2: Predict sales through an adaptive sparse network and dynamically adjust the sparsity according to the formula;

[0067] S3: Based on dual-timescale streaming reinforcement learning computational replenishment strategy, with fast and slow timescales updated separately;

[0068] S4: Apply Bayesian confidence to fuse expert rules and model output;

[0069] S5: Execute replenishment decisions and collect feedback information;

[0070] S6: Update system parameters through an adaptive learning rate mechanism to form a closed-loop optimization.

[0071] Furthermore, in step S3, the distribution drift problem in streaming learning is addressed by employing importance sampling correction:

[0072] ;

[0073] in, This is the corrected time difference error. This is the original time difference error. and These are the current and historical data distributions, respectively. ∈[1.5, 3.0] is the clipping constant;

[0074] Distribution changes are monitored using KL divergence; when the KL divergence exceeds the threshold θ... KL When this happens, a slow timescale parameter recalibration is triggered to ensure the stability of the strategy in non-stationary environments.

[0075] Furthermore, the system performance indicators are as follows:

[0076] Prediction accuracy: Mean absolute percentage error (MAPE) ≤ 12%, a 40% reduction compared to the traditional ARIMA method;

[0077] Inventory turnover rate: increased by more than 30%;

[0078] Out-of-stock rate: reduced by more than 50%;

[0079] System response: End-to-end latency ≤ 50 milliseconds, supports millisecond-level parameter updates;

[0080] Resource consumption: Memory ≤ 100MB, Model parameter count ≤ 100K;

[0081] Scalability: Supports parallel processing of 10,000+ SKUs, with a single update complexity of O(s·d), where s is the number of sparse connections and d is the network depth.

[0082] Compared with the prior art, the beneficial effects of the present invention are:

[0083] 1. This invention provides an AI-driven supply chain sales forecasting and intelligent inventory optimization system and method, which supports real-time forecasting and strategy updates in scenarios with continuous order data inflow. It adopts a streaming reinforcement learning architecture, abandons the traditional batch processing modeling method, does not rely on experience replay caching and target network structure, has high-frequency order processing capabilities and dynamic strategy adaptation capabilities, and can cope with complex business environments such as intensive promotions, high concurrency in e-commerce, and strong volatility.

[0084] 2. This invention provides an AI-driven supply chain sales forecasting and intelligent inventory optimization system and method. It employs a sparsely initialized multilayer perceptron model structure, combined with sliding window features, periodic time factors, and product embedding vectors, to improve the model's training stability and generalization ability under cold-start, small-sample, and high-dimensional input conditions. Simultaneously, the system has a lightweight structure, low computational resource consumption, and is feasible for direct deployment on edge nodes or in enterprise private cloud environments, adapting to multi-terminal parallel deployment requirements.

[0085] 3. This invention provides an AI-driven supply chain sales forecasting and intelligent inventory optimization system and method, constructing a complete closed-loop path covering sales forecasting, inventory optimization, expert intervention, and feedback updates. It uses an instant reward function to uniformly model inventory holding costs, stockout losses, and replenishment expenses, and introduces a rule-guided mechanism through an expert model module to automatically correct replenishment strategies and mitigate risks in typical scenarios such as holidays, promotional activities, or sales fluctuations, significantly improving the system's response speed to structural events and its business interpretation capabilities. Attached Figure Description

[0086] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0087] Figure 1 This is a schematic diagram of the system architecture of the present invention;

[0088] Figure 2 This is a flowchart of the state-action-reward modeling process for the inventory optimization strategy module;

[0089] Figure 3 This is a schematic diagram of the data processing and forecasting process for the sales forecasting module. Detailed Implementation

[0090] The technical solution of the present invention will be more clearly and completely explained below with reference to the accompanying drawings and through the description of preferred embodiments of the present invention.

[0091] like Figure 1 As shown, the entire system consists of five core modules, forming a highly integrated intelligent decision-making system. The data preprocessing module directly connects to the enterprise's supply chain order system, collecting structured information such as sales volume, product identification, and inventory status in real time, and ensuring data quality through missing value handling and standardization transformation. The processed data flows into the feature construction and standardization unit, which uses a sliding window mechanism to construct time series features and ensures the continuity and integrity of the features through high overlap sampling.

[0092] The sales forecasting module is one of the core components of the system, and its detailed data processing and forecasting process is as follows: Figure 3 As shown, the original order data undergoes a systematic multi-layered processing. First, missing values ​​are handled and data standardization is performed using methods such as moving average imputation and linear interpolation. Then, a sliding window builder generates high-dimensional feature vectors. The feature extractor further extracts key information such as periodic features, embedding vectors, moving averages, trend slopes, and price variation coefficients, providing rich input for the prediction model. The system employs a sparse perceptual neural network for sales forecasting. The network structure uses a multi-layered design with progressively decreasing hidden layers, dynamically adjusting the network connection density through an adaptive sparsity evolution mechanism. This design allows the network to adaptively adjust its structure from an initial high sparsity state based on the convergence speed of the prediction error, significantly reducing computational resource consumption while maintaining prediction accuracy. The dynamic sparsity adjustment and error monitoring mechanisms on the side ensure that the model can respond in real-time to changes in data distribution, ultimately outputting high-precision sales forecasts for the next several days.

[0093] 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 2As 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.

[0094] 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.

[0095] 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.

[0096] 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.

[0097] 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.

[0098] 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.

[0099] 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.

[0100] 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.

[0101] ;

[0102] 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.

[0103] 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 These are the input and output dimensions of the layer, respectively, and the bias term is initialized to 0.

[0104] The network training uses the Adam adaptive moment estimation optimizer.

[0105] In the dynamic adjustment strategy of network sparsity, the initial sparsity Set to 0.95, minimum sparsity Set to 0.5, the sparsity time decay constant. Set to 1000, error convergence threshold Set to 0.01. The importance score for connections uses a comprehensive index:

[0106] ;

[0107] in For connection weights, For all weight matrices, For loss function, Activation value This represents the total number of activations. The network structure is re-evaluated and adjusted every 100 iterations.

[0108] The state space of the inventory optimization decision module is designed as a 31-dimensional vector, including the current inventory level $I_t$, the sales forecast for the next 7 days, the actual sales for the past 7 days, the forecast error for the past 7 days, the current order in transit quantity, the supplier's average delivery time LT, the standard deviation of delivery time, and the week code (1-7) and month code (1-31) for the current date. The action space is discretized into 20 levels, including replenishment actions. The replenishment capacity is evenly distributed from 0 to the maximum.

[0109] The Q network employs a duel network architecture, with a shared layer consisting of two fully connected layers (128 and 64 nodes), and a hidden layer of 32 nodes each for the value stream and the advantage stream. In the dual-timescale update mechanism, a fast learning rate is used. Set to 0.05, slow learning rate Set to 0.005, cumulative window size k=50. Fast timescale time difference error. and slow time scale time difference error Based on the Bellman equation, the qualification trace decay coefficient λ is set to 0.92, and the reward discount factor γ is set to 0.95.

[0110] KL divergence threshold for distribution drift detection Set to 0.1, the clipping constant for importance sampling. Set to 2.0. The state distribution is estimated using a circular buffer with a capacity of 10000, employing kernel density estimation, and the bandwidth h is determined using the Silverman rule.

[0111] In the immediate reward function, the safety stock level Set the average sales over 14 days, and use the formula

[0112] ;

[0113] Dynamic calculation, including average daily demand and standard deviation of demand Calculated based on historical 30-day data The standard normal quantile corresponding to a 95% service level. Maximum inventory capacity. Set as the average sales volume over 30 days. Actual sales volume. To predict sales With available inventory Minimum value. Ending inventory. .

[0114] In the piecewise inventory cost function, the basic holding cost Set to 0.02% of the product value per day. Low inventory risk coefficient is set to 0.5, and high inventory penalty coefficient is set to 2.0. Unit sales profit p is dynamically set based on the product's gross profit margin, averaging 35% of the selling price for apparel. Out-of-stock penalty coefficient α is set to 1.5 times the average selling price, and overstock penalty coefficient β is set to... 10 times.

[0115] In the momentum qualification 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. Calculated via backpropagation. In importance sampling correction, the policy probability ratio... Calculated using the Boltzmann distribution:

[0116] ;

[0117] Among them temperature parameters Set it to 1.0 to control the exploration level.

[0118] The expert model assistance module contains 120 business rules. In the Bayesian confidence fusion, the historical sales average is used. and standard deviation Calculated based on 90 days of data, updated daily. Prediction uncertainty. Estimated using Monte Carlo Dropout, with M=10 forward propagations performed. Maximum allowable uncertainty. Set to 2 Model confidence threshold Set to 0.6. Rule confidence. The initial value is 0.7, which is dynamically adjusted based on historical accuracy, with the range limited to [0.3, 0.95].

[0119] 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.

[0120] 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.

[0121] 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.

[0122] 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.

[0123] 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.

[0124] 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.

[0125] 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.

[0126] In response to supply chain disruptions, the threshold for backup suppliers is a 3-day delay from the primary supplier. Unstable suppliers are defined as those whose delivery time standard deviation exceeds 2 days in the last 30 days; in this case, safety stock is increased by 30%, i.e., multiplied by 1.3.

[0127] In actual operation, the system ran continuously for 30 days in a test environment containing 1000 SKUs, processing an average of 1.5 million orders per day. The prediction accuracy (MAPE) remained stable at 11.8% ± 1.2%, with 10.5% on weekdays, 13.8% on weekends, and 15.2% during promotional periods. Inventory turnover increased from 8.5 times / year to 11.2 times / year. The stockout rate decreased from 4.2% to 2.0%. The average end-to-end latency was 45 milliseconds, P95 latency was 47 milliseconds, and P99 latency was 48 milliseconds. Memory usage was 85MB, including 38MB for model parameters, 35MB for data cache, and 12MB for runtime. The total number of model parameters was 98K, with 61K for the prediction network and 37K for the decision network.

[0128] The above-described specific embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Various modifications, substitutions, and improvements made by those skilled in the art to the technical solutions of the present invention based on the provided textual description and drawings, without departing from the design concept and spirit of the present invention, should all fall within the scope of protection of the present invention. The scope of protection of the present invention is determined by the claims.

Claims

1. An AI-driven supply chain sales forecasting and intelligent inventory optimization system, characterized in that: include: The data preprocessing module is used to connect with the supply chain order system, collect structured information, perform missing value processing and standardization transformation, and construct a sliding time window feature sequence with overlapping intervals; The sales forecasting module employs a deep neural network with an adaptive sparsity evolution mechanism. The initial sparsity is set to over 90%, and the network connection density is dynamically adjusted by monitoring the convergence speed of the prediction error. It also supports a bufferless streaming online update mechanism. The inventory optimization decision module adopts a streaming reinforcement learning architecture based on qualification traces, abandons experience replay caching and target network structure, and solves the gradient vanishing and distribution drift problems under high-frequency data streams through a dual time scale update mechanism. The expert model assistance 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. It achieves adaptive fusion of model output and expert rules through Bayesian confidence scoring. The strategy feedback update module adopts an incremental parameter update strategy, and achieves a closed-loop iteration of prediction-optimization-feedback through sliding window error accumulation and adaptive learning rate adjustment.

2. The AI-driven supply chain sales forecasting and intelligent inventory optimization system according to claim 1, characterized in that, The adaptive sparsity evolution mechanism of the sales forecasting module includes: The sparsity dynamic adjustment strategy adaptively adjusts based on the convergence speed of the prediction error: ； in, Let be the network sparsity at time t. To minimize sparsity, To achieve maximum sparsity, Let ε be the time decay constant, ε be the error convergence threshold, and ΔE be the error decrease rate within the sliding window. The importance assessment of connections is based on a weighted combination of weight magnitude, gradient contribution, and activation frequency. It dynamically retains a preset proportion of connections with the highest importance scores, enabling online adaptive evolution from sparse to dense.

3. The AI-driven supply chain sales forecasting and intelligent inventory optimization system according to claim 1, characterized in that, The inventory optimization decision module adopts a dual-timescale streaming reinforcement learning update mechanism: Fast timescale updates, i.e., 1-10 milliseconds: ; Slow timescale updates, i.e., 100-1000 milliseconds: ; Where, θ fast and θ slow The parameter sets are updated quickly and slowly, respectively. The value can be [0.01, 0.1]. The value ranges from [0.001, 0.01], where k is the cumulative window size. For time difference error, For qualification traces; By separating parameters and updating asynchronously, the system captures instantaneous gradient information on a fast timescale and maintains policy stability on a slow timescale, thus solving the distribution drift problem in streaming learning.

4. The AI-driven supply chain sales forecasting and intelligent inventory optimization system according to claim 1, characterized in that, The immediate reward function of the inventory optimization decision module adopts a piecewise nonlinear design: when hour: ; when hour: ; when hour: ; when When out of stock: ; in: Instant reward at time t Actual sales volume Ending inventory Safety stock level Maximum inventory capacity , , Piecewise inventory cost function Unit sales profit Out-of-stock penalty coefficient Over-storage penalty coefficient Forecast sales at time t Current inventory Restocking activity; Nonlinear modeling of inventory costs is achieved through segmented design, and differentiated cost strategies are adopted under different inventory levels.

5. The AI-driven supply chain sales forecasting and intelligent inventory optimization system according to claim 1, characterized in that, The inventory optimization decision-making module enhances credit allocation by introducing a momentum qualification trace mechanism: ； ； in, To enhance qualification traces, Let μ be the momentum term, μ ∈ [0.1, 0.3] be the momentum influence coefficient, and ν ∈ [0.9, 0.99] be the momentum decay rate. As a reward discount factor, The qualification trace decay coefficient, This is a restocking action; The introduction of the momentum term accelerates the propagation of credit allocation and enhances the learning ability of long-term dependencies; at the same time, adaptive gradient pruning is used to prevent gradient explosion.

6. The AI-driven supply chain sales forecasting and intelligent inventory optimization system according to claim 1, characterized in that, The expert model assistance module employs a Bayesian confidence fusion mechanism. Model confidence is calculated based on prediction uncertainty: ; The fused output uses Bayesian weighting: ; in, Model confidence; Model predicted output value; Historical average sales; : Standard deviation of historical sales; : Predicting uncertainty; Maximum permissible uncertainty; Rule confidence; Model output; : Rule output; When the model confidence level is below the threshold At that time, the rule weights are automatically increased to achieve the integration of model and expert knowledge.

7. The AI-driven supply chain sales forecasting and intelligent inventory optimization system according to claim 1, characterized in that, The policy feedback update module employs an adaptive learning rate adjustment mechanism: ； in, Let be the adaptive learning rate at time t. , , The recent error variance To adjust sensitivity; By dynamically adjusting the learning rate based on the error variance, the learning rate is reduced to maintain stability when the variance is high, and increased to accelerate convergence when the variance is low, thus achieving adaptive optimization.

8. An AI-driven supply chain sales forecasting and intelligent inventory optimization method, applicable to the AI-driven supply chain sales forecasting and intelligent inventory optimization system as described in any one of claims 1-7, characterized in that, Includes the following steps: S1: Construct an overlapping sliding window to collect order data, with a window length of T and an overlap rate greater than 50%; S2: Predict sales through an adaptive sparse network and dynamically adjust the sparsity according to the formula; S3: Based on dual-timescale streaming reinforcement learning computational replenishment strategy, with fast and slow timescales updated separately; S4: Apply Bayesian confidence to fuse expert rules and model output; S5: Execute replenishment decisions and collect feedback information; S6: Update 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 according to claim 8, characterized in that, Step S3 addresses the distribution drift problem in streaming learning by employing importance sampling correction: ； in, This is the corrected time difference error. This is the original time difference error. and These are the current and historical data distributions, respectively. This is the trimming constant; Distribution changes are monitored using KL divergence; when the KL divergence exceeds the threshold θ... KL When this happens, a slow timescale parameter recalibration is triggered to ensure the stability of the strategy in non-stationary environments.

Images