Supply chain management real-time decision system based on market data analysis

Abstract

The application relates to the technical field of market data processing, and particularly discloses a supply chain management real-time decision system based on market data analysis, which comprises the following steps: a data sensing module acquires multi-dimensional original data streams in real time, performs disorderly data correction and missing value interpolation based on a time window, and generates a standardized characteristic vector sequence; a deduction prediction module inputs the characteristic vector sequence into a multi-task learning model to output demand prediction values and fluctuation probability distributions; a collaborative decision module constructs a distributed decision network, and based on the demand prediction values and the fluctuation probability distributions, calculates strategy gradients of each intelligent agent through a multi-agent reinforcement learning algorithm according to preset resource constraint conditions, and generates a real-time decision instruction set; and a dynamic feedback module acquires actual inventory dynamic information, calculates deviation statistics between the demand prediction values and real consumption values, and inversely propagates the deviation statistics to a shared feature extraction layer of the multi-task learning model.

Description

Technical Field

[0001] This invention relates to the field of market data processing technology, and more specifically to a real-time decision-making system for supply chain management based on market data analysis. Background Technology

[0002] As global supply chains become increasingly complex, businesses face growing volatility in their external environments. Traditional supply chain management (SCM) systems primarily rely on statistical analysis of historical data from internal ERP (Enterprise Resource Planning) and POS (Point of Sale) systems to develop procurement, production, and distribution plans. However, these systems suffer from the following significant drawbacks: Current supply chain management systems typically employ a single-task learning model architecture for demand forecasting, meaning the model only outputs sales figures for a specific future period, neglecting quantitative modeling of forecast uncertainty. In actual supply chain operations, market demand is influenced by multiple factors, including macroeconomic fluctuations, competitor strategy adjustments, and shifts in consumer preferences, exhibiting significant non-stationarity and randomness. However, traditional forecasting models lack explicit mechanisms to express market volatility risk, rendering their single forecasts ineffective in providing risk references for downstream decision-making. Downstream inventory management, production scheduling, and logistics allocation modules, upon receiving a "deterministic forecast," can only passively respond based on fixed safety stock coefficients or empirical rules, unable to dynamically adjust decision-making strategies according to the severity of market fluctuations. When actual market fluctuations exceed preset empirical ranges, the system often suffers from structural inventory imbalances due to a lack of risk perception, or becomes overly conservative during periods of high uncertainty, leading to resource waste.

[0003] Furthermore, existing supply chain decision-making systems mostly employ a centralized, serial computing architecture. After the forecasting module completes its calculations, it transmits the results to the planning module, which then distributes them sequentially to the procurement, production, and logistics stages. In this architecture, the forecasting model is fixed once trained, and its parameters remain unchanged throughout the entire operating cycle. As the market environment changes, a gradual shift (i.e., model drift) occurs between the output distribution of the forecasting model and the actual demand distribution. The serial architecture lacks a feedback path from the execution end to the forecasting end, making it impossible to dynamically adjust the internal parameters of the forecasting model based on actual performance data. This leads to a continuous degradation in forecast accuracy over time, ultimately affecting the overall decision-making quality of the supply chain.

[0004] Therefore, how to simultaneously output deterministic forecasts and uncertainty quantification results during the demand forecasting stage, and to build a closed-loop feedback mechanism from the execution end to the forecasting end, has become a key technical challenge to improve the adaptability of the supply chain management system. Summary of the Invention

[0005] The purpose of this invention is to provide a real-time decision-making system for supply chain management based on market data analysis, and to solve the following technical problems: How to ensure that the demand forecasting model outputs both deterministic forecasts and uncertainty quantification results, and how to achieve a complete online adaptive update mechanism for the forecasting model based on actual data from the execution end.

[0006] The objective of this invention can be achieved through the following technical solutions: A real-time decision-making system for supply chain management based on market data analysis includes: The data perception module is used to acquire multidimensional raw data streams in real time, and perform time window-based out-of-order data correction and missing value interpolation on the multidimensional raw data streams to generate a standardized feature vector sequence. The inference and prediction module is used to input the feature vector sequence into the multi-task learning model to output the demand prediction value and the fluctuation probability distribution. The multi-task learning model includes a shared feature extraction layer, a first decoding head and a second decoding head. The shared feature extraction layer outputs deep semantic features. The first decoding head regresses based on the deep semantic features to obtain the regionalized demand prediction value within the future preset period. The second decoding head regresses based on the deep semantic features to obtain the fluctuation probability distribution corresponding to the demand prediction value. The collaborative decision-making module is used to construct a distributed decision-making network and, based on demand forecasts and fluctuation probability distributions, calculates the policy gradients of each agent through a multi-agent reinforcement learning framework that optimizes the PPO algorithm for near-end policy, according to preset resource constraints, and generates a real-time decision instruction set. The collaborative decision-making module is equipped with a global reward function, which is determined by a weighted sum of inventory holding costs, order stockout penalties, and logistics transportation costs. The dynamic feedback module is used to obtain dynamic information on actual inventory, calculate the deviation statistics between the demand forecast and the actual consumption, and backpropagate the deviation statistics to the shared feature extraction layer of the multi-task learning model to update the network weight parameters of the shared feature extraction layer online, thereby realizing the decision-making closed loop. The dynamic feedback module is also equipped with a counterfactual reasoning submodule. When the deviation statistic exceeds the drift threshold, a counterfactual simulation scenario similar to the current market environment is constructed. In the counterfactual simulation scenario, the hyperparameter combination of the multi-task learning model is adjusted. Based on the simulation results, the optimal hyperparameter update strategy is determined and applied to the online inference and prediction module.

[0007] Preferably, the shared feature extraction layer includes a long short-term memory network sublayer and a multi-head attention mechanism sublayer; Furthermore, the long short-term memory network sublayer is used to model the temporal dependency relationship of the feature vector sequence, and outputs a hidden state vector containing long-term trend features; Furthermore, the multi-head attention mechanism sublayer is coupled to the LSTM sublayer and configured to perform multi-head parallel self-attention calculation on the hidden state vector in different feature dimensions to capture the cross-dimensional nonlinear correlation between consumer behavior, competitor dynamics and economic indicators.

[0008] Preferably, the second decoding head is used to output the mean and variance parameters of each time step within a future preset period based on deep semantic features through a fully connected layer, and to construct a fluctuation probability distribution that follows a Gaussian distribution based on the mean and variance parameters, so as to quantify the uncertainty risk of market fluctuations.

[0009] Preferably, the demand forecast is calculated as follows: The first decoder receives the deep semantic features output by the shared feature extraction layer. ; Preset cycle for the future Inner The region in the first Demand forecast at each time step The calculation formula is: ; in, To predict the cycle length, ∈ ,and For the total number of regions, The deep semantic feature vectors output by the shared feature extraction layer, and ∈ , Indicates the current moment. The dimension of deep semantic features; For the first The weight matrix corresponding to each time step ∈ ; Let be the demand bias vector, and ∈ , This is the activation function used to constrain demand. For the first The region in the first Regional demand forecasts for each time period.

[0010] Preferably, the fluctuation probability distribution is calculated as follows: The second decoder receives the deep semantic features output by the shared feature extraction layer. and output the mean parameter. and variance parameter : ; ; in, The weight vector is the mean. The weight vector is the log-variance. It is the mean bias vector. This is the variance bias vector; The second decoder will... The region in the first The demand at any given time is modeled as a random variable following a Gaussian distribution, where the mean parameter is... Take the localized demand prediction value of the corresponding region and time step output by the first decoder head. The calculation formula is: ; in, This is the actual demand value. Variance represents the degree of market uncertainty.

[0011] Preferably, the deviation statistic in the dynamic feedback module is specifically the mean absolute percentage error; the dynamic feedback module is used to trigger the online incremental learning process of the multi-task learning model when the mean absolute percentage error in N consecutive monitoring periods exceeds the preset drift threshold, and to use the actual inventory level and order fulfillment data in the most recent M monitoring periods as training samples to fine-tune the network weight parameters of the shared feature extraction layer in order to suppress model drift.

[0012] Preferably, the data perception module is deployed in an edge computing node cluster. The edge computing node cluster is configured to perform local preprocessing on data within a preset geographical range from the data source, and only upload the processed feature vector sequence to the cloud server to reduce data transmission bandwidth usage and the computing load of the central node.

[0013] Preferably, the dynamic feedback module is further configured with a counterfactual reasoning submodule, which includes: When the deviation statistic exceeds the drift threshold, a counterfactual simulation scenario similar to the current market environment is constructed. The hyperparameter combination of the multi-task learning model is adjusted in the counterfactual simulation scenario, and the optimal hyperparameter update strategy is determined based on the simulation results. The optimal hyperparameter update strategy is then applied to the online inference and prediction module.

[0014] Preferably, the multidimensional raw data stream includes consumer behavior, competitor dynamics, and economic indicators; The collaborative decision-making module is also equipped with a global reward function, which is determined by the weighted sum of inventory holding costs, order out-of-stock penalties, and logistics transportation costs. The collaborative decision-making module is also coupled with a blockchain-based distributed ledger. Preset resource constraints are stored in the distributed ledger, which is configured to record the interaction logs and resource usage status of each agent during the generation of real-time decision instruction sets, ensuring the traceability and tamper-proof nature of the decision-making process.

[0015] The beneficial effects of this invention are: The problem of real-time access and standardization of multi-source heterogeneous market data: How to solve the mismatch between external market data (consumer behavior, competitor dynamics, economic indicators) and internal supply chain data of enterprises in terms of granularity, format and timeliness, and achieve the technical challenge of second-level data perception and cleaning; The coexistence of forecast accuracy and uncertainty quantification: How to ensure the accuracy of regional demand forecasting while outputting a quantifiable probability distribution of fluctuations, so as to provide risk perception capabilities for supply chain decision-making in high-risk scenarios; Multi-department decision-making conflicts and real-time issues: How to break through the efficiency bottleneck of traditional serial decision-making processes, achieve parallel collaborative optimization of procurement, production, and logistics under multiple resource constraints, and at the same time resolve resource competition conflicts among multiple entities; The problem of maintaining the long-term effectiveness of the model: How to construct a closed-loop feedback mechanism of prediction-decision-execution, realize online adaptive updates of the model based on actual supply chain operation data, and suppress model drift.

[0016] Of course, any product implementing this invention does not necessarily need to achieve all the advantages described above at the same time. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a block diagram of the real-time decision-making system for supply chain management based on market data analysis, as described in this invention. Figure 2 This is a schematic diagram of the structure of a multi-task learning model according to the present invention. Detailed Implementation

[0019] 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.

[0020] Please see Figure 1-2 As shown, this invention is a real-time decision-making system for supply chain management based on market data analysis, comprising: The data perception module is used to acquire multidimensional raw data streams in real time, and perform time window-based out-of-order data correction and missing value interpolation on the multidimensional raw data streams to generate a standardized feature vector sequence. The inference and prediction module is used to input the feature vector sequence into the multi-task learning model to output the demand prediction value and the fluctuation probability distribution. The multi-task learning model includes a shared feature extraction layer, a first decoding head and a second decoding head. The shared feature extraction layer outputs deep semantic features. The first decoding head regresses based on the deep semantic features to obtain the regionalized demand prediction value within the future preset period. The second decoding head regresses based on the deep semantic features to obtain the fluctuation probability distribution corresponding to the demand prediction value. The collaborative decision-making module is used to construct a distributed decision-making network and, based on demand forecasts and fluctuation probability distributions, calculates the policy gradients of each agent through a multi-agent reinforcement learning framework that optimizes the PPO algorithm for near-end policy, according to preset resource constraints, and generates a real-time decision instruction set. The collaborative decision-making module is equipped with a global reward function, which is determined by a weighted sum of inventory holding costs, order stockout penalties, and logistics transportation costs. The dynamic feedback module is used to obtain dynamic information on actual inventory, calculate the deviation statistics between the demand forecast and the actual consumption, and backpropagate the deviation statistics to the shared feature extraction layer of the multi-task learning model to update the network weight parameters of the shared feature extraction layer online, thereby realizing the decision-making closed loop. The dynamic feedback module is also equipped with a counterfactual reasoning submodule. When the deviation statistic exceeds the drift threshold, a counterfactual simulation scenario similar to the current market environment is constructed. In the counterfactual simulation scenario, the hyperparameter combination of the multi-task learning model is adjusted. Based on the simulation results, the optimal hyperparameter update strategy is determined and applied to the online inference and prediction module.

[0021] The aforementioned technical solution includes a real-time supply chain management decision-making system based on market data analysis, comprising a data perception module, a prediction and inference module, a collaborative decision-making module, and a dynamic feedback module. The data perception module is deployed in an edge computing node cluster and configured to ingest multi-dimensional raw data streams in parallel from multiple external data sources using a streaming computing framework (such as Apache Flink or Apache Kafka Streaming). The data perception module acquires multi-dimensional raw data streams in real time and performs time-window-based out-of-order data correction and missing value interpolation on the data streams to generate standardized feature vector sequences. The multi-dimensional raw data streams specifically include: consumer behavior data: clickstream logs, shopping cart addition records, search keyword frequencies, etc., from e-commerce platforms; competitor dynamic data: competitor product prices, promotional activities, inventory status, etc., periodically captured via API interfaces; and economic indicator data: published CPI and PMI indices, industry prosperity indices, and social media sentiment analysis results.

[0022] The inference and prediction module is coupled to the data perception module and is configured to input the feature vector sequence into a multi-task learning model. This multi-task learning model includes a shared feature extraction layer, a first decoding head, and a second decoding head. The shared feature extraction layer is used to accurately capture the cross-dimensional nonlinear correlation between consumer behavior, competitor dynamics, and economic indicators, and finally outputs deep semantic feature information. The first decoding head uses the deep semantic feature information to regress and predict the regional demand forecast value within a preset period in the future, and the second decoding head uses the same deep semantic feature information to determine the fluctuation probability distribution corresponding to the demand forecast value, which is convenient for the next step of processing. The collaborative decision-making module is coupled to the inference and prediction module and configured to build a distributed decision-making network including procurement agents, production agents, and logistics agents. This module can be externally connected to a blockchain ledger to store constraints. Each agent uses the Proximal Policy Optimization (PPO) algorithm for reinforcement learning and is configured with an independent Actor network and Critic network. Based on the demand forecast and fluctuation probability distribution, and according to preset resource constraints (including capacity limits, warehouse capacity limits, transportation vehicle quantity constraints, and budget constraints), the collaborative decision-making module calculates the policy gradient of each agent in parallel iteratively through a multi-agent reinforcement learning algorithm. The dynamic feedback module is coupled to the collaborative decision-making module and the extrapolation and prediction module, and is configured to monitor the actual inventory level and order fulfillment delay at the supply chain execution end through an Internet of Things (IoT) sensor network. Specifically, it collects real-time dynamic information on actual inventory and order fulfillment data in each warehouse area through RFID readers, weighing sensors, and GPS tracking devices deployed in the warehouse; the dynamic feedback module calculates the statistical deviation between the demand forecast and the actual consumption value.

[0023] In summary, in one specific embodiment, the edge computing node cluster of the data awareness module adopts a three-tier collaborative architecture of "cloud-edge-device": Terminal layer: This includes IoT sensors (RFID readers, weighing sensors) deployed in warehouses across various regions, e-commerce platform log collection agents, competitor data crawlers, etc., which are responsible for the collection and initial packaging of raw data. Edge layer: Deploy preprocessing services on edge nodes within a 50-kilometer radius of the data source. Each edge node is configured with a 4-core CPU, 16GB of memory, and a 100GB SSD, running the Ubuntu 20.04 operating system and deploying a Docker containerized environment. Cloud layer: Deployed on a public cloud server (configured with a 32-core CPU, 128GB memory, and 1TB SSD), responsible for computationally intensive tasks such as model training and global policy formulation.

[0024] The edge nodes use the Apache Flink streaming computing framework (version 1.14) with a parallelism of 4. The specific data preprocessing flow is as follows: (1) Data access: Raw data is ingested in parallel from multiple data sources through Flink's Kafka Source connector. The Kafka cluster is deployed on 3 nodes with 12 partitions. (2) Disordered data correction: The late data is reordered using a sliding time window (window size W=5 minutes, sliding step S=1 minute), and the maximum allowed late time is 2 minutes; for missing data that has not yet arrived within the window, multiple imputation based on historical data of the same period is used to complete it - the median of the same time in the past 7 days is taken as the imputation value; (3) Data standardization: Z-Score standardization is performed on continuous features (mean and standard deviation are obtained from the training set statistics), and One-Hot encoding is performed on categorical features; (4) Feature vector generation: A d vector is generated at each time step. x =128-dimensional feature vectors, and a sequence of feature vectors is formed by L=30 consecutive time steps as the model input.

[0025] The data synchronization strategy between edge nodes and cloud servers is as follows: Processed, standardized feature vector sequences are uploaded in batches every 5 minutes, using the gRPC protocol for transmission, with each upload limited to under 10MB. Actual testing shows this reduces bandwidth usage by approximately 85%.

[0026] Fault tolerance mechanism: Edge nodes are configured with Flink's RocksDB state backend to achieve state persistence, with a checkpoint interval of 30 seconds; when the network is interrupted, the preprocessed data is temporarily stored in a local Kafka queue (100GB disk capacity), and is retransmitted in order after the network is restored to ensure that no data is lost.

[0027] As one embodiment of the present invention, please refer to Figure 2 As shown, the shared feature extraction layer includes a long short-term memory network sublayer and a multi-head attention mechanism sublayer; Furthermore, the long short-term memory network sublayer is used to model the temporal dependency relationship of the feature vector sequence, and outputs a hidden state vector containing long-term trend features; Furthermore, the multi-head attention mechanism sublayer is coupled to the LSTM sublayer and configured to perform multi-head parallel self-attention calculation on the hidden state vector in different feature dimensions to capture the cross-dimensional nonlinear correlation between consumer behavior, competitor dynamics and economic indicators.

[0028] In the above technical solution, the shared feature extraction layer is composed of a Long Short-Term Memory (LSTM) sublayer and a multi-head attention mechanism sublayer connected in series. The LSTM sublayer receives a sequence of feature vectors from consecutive time steps. The temporal dependencies in the sequence are modeled through gating mechanisms (input gate, forget gate, output gate), and the output is a hidden state vector containing long-term trend features. The multi-head attention mechanism sublayer is coupled to the LSTM sublayer and configured to... Parallel self-attention computation (m=8) is performed across m feature dimensions to capture cross-dimensional nonlinear relationships between consumer behavior, competitor dynamics, and economic indicators, ultimately outputting deep semantic features. The first decoding head is based on these deep semantic features. ( ).

[0029] The multi-task learning model employs a training strategy that combines offline pre-training with online fine-tuning. The specific training process is as follows: (I) Construction of Training Dataset: Supply chain operation data of a large retail enterprise over the past three years (January 2019 to December 2021) was collected, including: daily outbound records from 32 regional warehouses (approximately 35,040 records), consumer behavior logs from e-commerce platforms (5 million records per day, sampled and stored), historical data on competitor prices and promotional activities (collected daily via API), and macroeconomic indicators (industry indices such as CPI and PMI). The data was divided into a training set (70%), a validation set (15%), and a test set (15%) in chronological order, strictly ensuring the chronological order to avoid future information leakage.

[0030] (ii) Definition of loss function: The model adopts end-to-end joint optimization, and the total loss function is the weighted sum of the demand forecasting task and the uncertainty quantification task.

[0031] The demand forecasting task uses the mean absolute error (MAE): ; The uncertainty quantification task employs negative log-likelihood loss (NLL): ; Total loss: ,in (Determined through a grid search of the validation set).

[0032] (III) Training hyperparameters: The optimizer is Adam, and the initial learning rate is 1×10. -3 The training uses cosine annealing with a learning rate of 256; batch size of 256; 200 training epochs; dropout rate of 0.2; and weight decay (L2 regularization) coefficient of 1×10⁻⁶. -5 An early stopping strategy is adopted, and training is terminated when the validation set loss does not decrease for 20 consecutive rounds.

[0033] (iv) Convergence criterion: When the total loss L on the validation set... total If the change in the model over 10 consecutive training rounds is less than 0.1% of the initial loss value, the model is considered to have converged, and the current parameters are saved as a pre-trained model.

[0034] (v) Model evaluation: Calculate MAPE and RMSE indicators on the test set as benchmarks for subsequent online monitoring.

[0035] As one embodiment of the present invention, the second decoding head is used to output the mean and variance parameters of each time step within a future preset period through a fully connected layer based on deep semantic features, and to construct a fluctuation probability distribution that follows a Gaussian distribution based on the mean and variance parameters, so as to quantify the uncertainty risk of market fluctuations.

[0036] In the above technical solution, the second decoder obtains the fluctuation probability distribution corresponding to the demand forecast value based on the same deep semantic feature regression. First, the second decoder outputs the mean parameter and the log-variance parameter, where are the weight vectors of the mean and variance branches, respectively, and is the corresponding bias term. Subsequently, the second decoder models the demand of the r-th region at time t+k as a random variable following a Gaussian distribution. In practical applications, the mean of the prediction distribution is set to be equal to the point prediction value output by the first decoder. The variance characterizes the degree of uncertainty of market fluctuations. When it is large, it indicates that the region faces higher market risk at that time step.

[0037] As one embodiment of the present invention, the demand forecast is calculated as follows: The first decoder receives the deep semantic features output by the shared feature extraction layer. ; Preset cycle for the future Inner The region in the first Demand forecast at each time step The calculation formula is: ; in, To predict the cycle length, ∈ ,and For the total number of regions, The deep semantic feature vectors output by the shared feature extraction layer, and ∈ , Indicates the current moment. The dimension of deep semantic features; For the first The weight matrix corresponding to each time step ∈ ; Let be the demand bias vector, and ∈ , This is the activation function used to constrain demand. For the first The region in the first Regional demand forecasts for each time period.

[0038] In the above technical solution, the first decoding head obtains the regional demand forecast value within a preset period based on deep semantic feature regression; for example, the preset period is the next 14 days (H=14), covering 32 sales regions nationwide (R=32); the calculation formula for the demand forecast value of the r-th region at time t+k is: , ∈ and ∈ These are the weight matrix and bias vector corresponding to the k-th time step, respectively; activation function. Use the Softplus function, i.e. This is to ensure that the output is a non-negative real number, which conforms to the physical constraints of commodity demand.

[0039] As one embodiment of the present invention, the fluctuation probability distribution is calculated as follows: The second decoder receives the deep semantic features output by the shared feature extraction layer. and output the mean parameter. and variance parameter : ; ; in, The weight vector is the mean. The weight vector is the log-variance. It is the mean bias vector. This is the variance bias vector; The second decoder will... The region in the first The demand at any given time is modeled as a random variable following a Gaussian distribution, where the mean parameter is... Take the localized demand prediction value of the corresponding region and time step output by the first decoder head. The calculation formula is: ; in, This is the actual demand value. Variance represents the degree of market uncertainty.

[0040] In the above technical solution, the second decoding head obtains the fluctuation probability distribution corresponding to the demand prediction value based on the same deep semantic feature regression. First, the second decoding head outputs the mean parameter. and variance parameter : ; These are the weight vectors for the mean and variance branches, respectively; then, the second decoder will... The region in the first The demand at any given time is modeled as a random variable following a Gaussian distribution. In practical applications, setting That is, the mean of the predicted distribution is equal to the predicted point value output by the first decoder head; variance Indicates the degree of uncertainty in market fluctuations, when A larger value indicates that the region faces higher market risk at that time step.

[0041] The specific configuration of the distributed decision-making network of the collaborative decision-making module is as follows: (a) Types and number of intelligent agents: There are three types of intelligent agents, totaling 40. Among them, there are 32 procurement intelligent agents (1 per sales region), 3 production intelligent agents (1 per production base), and 5 logistics intelligent agents (1 per logistics center).

[0042] (ii) Definition of state space: Procurement agent (state dimension 64): includes current inventory level (normalized), average daily consumption over the past 7 days, and current demand forecast. {r,t} and fluctuation variance σ {r,t}² Supplier's current price and historical 30-day average, supplier's on-time delivery rate (statistics of the past 90 days), seasonal factor (1-4), and promotional activity intensity (0-1).

[0043] Production intelligence agent (state dimension 48): includes the utilization rate of each production line (0-1), work-in-process inventory level, the summary of demand in each region for the next 7 days, the number of days of raw material inventory available, equipment maintenance status (0 / 1), and the remaining capacity in the current scheduling cycle.

[0044] Logistics intelligent agent (state dimension 40): includes the current location and available capacity of each transportation vehicle, the length of the order queue to be delivered, the inter-regional transportation distance matrix, the real-time traffic congestion index (0-1), and the fuel price index.

[0045] (III) Definition of Action Space: Procurement agent (dimension 3): Discrete action {0, standard replenishment quantity, 1.5 × standard replenishment quantity}, where the standard replenishment quantity is calculated based on demand forecast and safety stock factor.

[0046] Production intelligence agent (dimensional 2+): output allocation ratio for each product line (continuous vector) and whether to start overtime production (binary).

[0047] Logistics intelligent agent (dimension 2): transportation route selection (discrete, the number of available routes is determined by the road network) and departure delay time (continuous value, 0-24 hours).

[0048] (iv) Actor and Critic network structure: Each agent is configured with independent Actor and Critic networks, both of which are fully connected neural networks.

[0049] Actor Network: Input Layer (Dimension d) s ) → Fully connected layer (256-dimensional, ReLU) → Fully connected layer (128-dimensional, ReLU) → Output layer (dimensional d) a The activation function is either Softmax or Tanh, depending on the action type.

[0050] Critic Network: Input Layer (Dimension d) s → Fully connected layer (256-dimensional, ReLU) → Fully connected layer (128-dimensional, ReLU) → Output layer (1-dimensional, linear, output state value).

[0051] (v) PPO algorithm hyperparameters: pruning parameter ε=0.2, discount factor γ=0.99, GAE parameter λ=0.95, Actor learning rate 3×10 -4 Critic learning rate 1×10 -3 (All using Adam optimizer), batch size 2048, updated 10 times after each sampling, entropy coefficient 0.01.

[0052] (vi) Communication and Coordination Mechanism between Agents: The "Centralized Training, Distributed Execution" (CTDE) paradigm is adopted. During training, each Critic can access the global state, while during execution, it only relies on local observations. At the same time, a conflict resolution mechanism based on the ContractNet Protocol is configured: when the production agent's capacity is insufficient, it broadcasts a resource request as the bidding party, and logistics and other auxiliary agents return quotations (estimated delivery time and cost) as bidders. The production agent selects the winning bidder based on historical fulfillment rate and on-time delivery rate.

[0053] As one embodiment of the present invention, the deviation statistic in the dynamic feedback module is specifically the mean absolute percentage error. The dynamic feedback module is used to trigger the online incremental learning process of the multi-task learning model when the mean absolute percentage error in N consecutive monitoring periods exceeds the preset drift threshold. It uses the actual inventory level and order fulfillment data in the most recent M monitoring periods as training samples to fine-tune the network weight parameters of the shared feature extraction layer in order to suppress model drift.

[0054] In the above technical solution, in this embodiment, the deviation statistic uses the mean absolute percentage error (MAPE), calculated using the formula... ,in, For the first The actual consumption within each monitoring period This represents the demand forecast for the corresponding period. For example, the number of periods in the statistical window. The value is 7, meaning 7 consecutive days. When the MAPE value exceeds the preset drift threshold (e.g., 10%) within N consecutive monitoring periods, the dynamic feedback module 140 triggers the online incremental learning process of the multi-task learning model. The system uses the actual inventory level and order fulfillment data within the most recent M monitoring periods (M=30 days in this embodiment) as new training samples to fine-tune the network weight parameters of the shared feature extraction layer, instead of retraining from scratch, thereby suppressing model drift without interrupting system services.

[0055] The detailed steps for implementing online incremental learning are as follows: (1) Triggering conditions: When the MAPE value exceeds the preset drift threshold θ=10% for 7 consecutive monitoring cycles (one cycle per day), the process is triggered.

[0056] (2) Training sample preparation: Extract the actual inventory level and order fulfillment data for the most recent 30 days from the database to form a new sample set D. new ={(x i ,y i )}, and i∈30.

[0057] (3) Fine-tuning hyperparameters: The Adam optimizer is used, and the learning rate is set to 1 / 10 of the initial training learning rate, i.e., 1×10. -4 The batch size is 64, and the fine-tuning rounds are 10 epochs. During the fine-tuning process, all parameters of the first and second decoders are frozen, and only the weight parameters of the shared feature extraction layer are updated to avoid catastrophic forgetting.

[0058] (4) Validation and deployment: After fine-tuning, calculate MAPE on the validation data of the most recent 7 days; if MAPE drops below the threshold, deploy the updated model parameters to the online inference and prediction module; if MAPE is still higher than the threshold after 3 consecutive fine-tunings, or if MAPE improves by less than 1 percentage point after a single fine-tuning, trigger full retraining - train the model from scratch using all data of the most recent 180 days.

[0059] In one embodiment of the present invention, the data perception module is deployed in an edge computing node cluster. The edge computing node cluster is configured to perform local preprocessing on data within a preset geographical range from the data source, and only upload the processed feature vector sequence to the cloud server to reduce data transmission bandwidth usage and the computing load of the central node.

[0060] In the above technical solution, the data perception module is deployed in an edge computing node cluster. The edge computing node cluster is configured to perform local preprocessing on data within a preset geographical range from the data source, ensuring that out-of-order data correction and missing value interpolation are performed on the multi-dimensional raw data stream based on time windows. Specifically, a sliding time window (the window size can be configured to 5 minutes) is used to reorder late data. For missing data that has not yet arrived within the window, a multiple interpolation method based on historical data from the same period is used to complete it, and finally a standardized feature vector sequence is generated. The edge computing nodes only upload the processed feature vector sequence to the cloud server, which greatly reduces the data transmission bandwidth usage.

[0061] In one embodiment of the present invention, the dynamic feedback module is further configured with a counterfactual reasoning submodule, which includes: When the deviation statistic exceeds the drift threshold, a counterfactual simulation scenario similar to the current market environment is constructed. The hyperparameter combination of the multi-task learning model is adjusted in the counterfactual simulation scenario, and the optimal hyperparameter update strategy is determined based on the simulation results. The optimal hyperparameter update strategy is then applied to the online inference and prediction module.

[0062] In the above technical solution, the dynamic feedback module is also configured with a counterfactual reasoning submodule. When the deviation statistic exceeds the drift threshold, the counterfactual reasoning submodule constructs a counterfactual simulation scenario similar to the current market environment. Specifically, the system retrieves several historical periods from the historical database that are similar to the current market characteristics (such as seasonal factors, promotional activity intensity, competitor price levels, etc.), and adjusts the hyperparameter combination of the multi-task learning model in these counterfactual simulation scenarios. The hyperparameter combination includes the learning rate (adjusted from 0.1 times to 1 times the initial value, for example from 1×10). -3 Adjusted to 1×10 -4 Batch size (adjustable from 32 to 256, e.g., from 256 to 64) and L2 regularization coefficient (adjustable from 1×10⁻⁶). -6 Up to 1×10 -4 For example, from 1×10 -5 Adjusted to 5×10 -5 The system runs simulations in each counterfactual simulation scenario based on the verification dataset of the most recent 30 days. The mean absolute percentage error (MAPE) in the simulation results is used as the evaluation metric. The hyperparameter combination that minimizes MAPE is selected as the optimal hyperparameter update strategy, which is then applied to the online inference and prediction module. This avoids high-risk trial and error in the online environment, ensuring the stability of the system. Furthermore, resource constraints are stored in a distributed ledger built using blockchain technology. This ledger is configured to record the interaction logs and resource usage status of each agent during the generation of real-time decision instruction sets, ensuring the traceability and tamper-proof nature of the decision-making process and providing a basis for enterprise auditing and compliance checks.

[0063] As one embodiment of the present invention, the multidimensional raw data stream includes consumer behavior, competitor dynamics, and economic indicators. The collaborative decision-making module is also equipped with a global reward function, which is determined by the weighted sum of inventory holding costs, order out-of-stock penalties, and logistics transportation costs. The collaborative decision-making module is also coupled with a blockchain-based distributed ledger. Preset resource constraints are stored in the distributed ledger, which is configured to record the interaction logs and resource usage status of each agent during the generation of real-time decision instruction sets, ensuring the traceability and tamper-proof nature of the decision-making process.

[0064] In the above technical solution, the collaborative decision-making module is also configured with a global reward function. : ; in, This includes inventory holding costs (including warehouse rent, capital occupation costs, and goods damage). The cost of stockout penalties (including losses from lost orders, decreased customer satisfaction, etc.). For logistics and transportation costs; , , The weighting coefficients for each cost item can be configured according to the company's strategic preferences. Each agent maximizes the global reward function. The strategy gradient is updated to achieve optimal global resource allocation, rather than just maximizing local interests. In addition, in practical cases, the collaborative decision-making module is also configured with a conflict resolution mechanism based on the Contract Net Protocol. When the available production capacity resources of the production agent cannot meet the raw material needs of the procurement agent, the production agent broadcasts a resource request message as the bidding party, and the logistics agent and other auxiliary production agents return quotation information (including estimated delivery time, cost, etc.) as the bidding parties. The production agent selects the winning bidder based on a preset reputation scoring model (comprehensively considering indicators such as historical performance rate and on-time delivery rate), thereby efficiently resolving resource competition conflicts in a distributed environment.

[0065] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and non-volatile computer storage media are basically similar to the method embodiments, and therefore described more simply; relevant parts can be referred to the descriptions of the method embodiments.

[0066] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined in this application, they should all fall within the protection scope of the present invention.

Claims

1. A real-time decision-making system for supply chain management based on market data analysis, characterized in that: include: The data perception module is used to acquire multidimensional raw data streams in real time, and perform time window-based out-of-order data correction and missing value interpolation on the multidimensional raw data streams to generate a standardized feature vector sequence. The inference and prediction module is used to input the feature vector sequence into the multi-task learning model to output the demand prediction value and the fluctuation probability distribution. The multi-task learning model includes a shared feature extraction layer, a first decoding head and a second decoding head. The shared feature extraction layer outputs deep semantic features. The first decoding head regresses the deep semantic features to obtain the regionalized demand prediction value within a future preset period. The second decoding head regresses the deep semantic features to obtain the fluctuation probability distribution corresponding to the demand prediction value. The collaborative decision-making module is used to construct a distributed decision-making network and, based on the demand forecast and the fluctuation probability distribution, calculate and obtain the policy gradient of each agent through a multi-agent reinforcement learning framework using the near-end policy optimization (PPO) algorithm according to preset resource constraints, and generate a real-time decision instruction set. The collaborative decision-making module is equipped with a global reward function, which is determined by a weighted sum of inventory holding costs, order out-of-stock penalties, and logistics transportation costs. The dynamic feedback module is used to obtain actual inventory dynamic information, calculate the deviation statistics between the demand forecast value and the actual consumption value, and backpropagate the deviation statistics to the shared feature extraction layer of the multi-task learning model to update the network weight parameters of the shared feature extraction layer online, thereby realizing a decision-making closed loop. The dynamic feedback module is also equipped with a counterfactual reasoning submodule. When the deviation statistic exceeds the drift threshold, a counterfactual simulation scenario similar to the current market environment is constructed. In the counterfactual simulation scenario, the hyperparameter combination of the multi-task learning model is adjusted. Based on the simulation results, the optimal hyperparameter update strategy is determined and applied to the online inference and prediction module.

2. The real-time decision-making system for supply chain management based on market data analysis according to claim 1, characterized in that, The shared feature extraction layer includes a long short-term memory network sublayer and a multi-head attention mechanism sublayer; Furthermore, the long short-term memory network sublayer is used to model the temporal dependency relationship of the feature vector sequence and output a hidden state vector containing long-term trend features; Furthermore, the multi-head attention mechanism sublayer is coupled to the LSTM sublayer and configured to perform multi-head parallel self-attention calculation on the hidden state vector in different feature dimensions to capture the cross-dimensional nonlinear correlation between consumer behavior, competitor dynamics, and economic indicators.

3. The real-time decision-making system for supply chain management based on market data analysis according to claim 1, characterized in that, The second decoding head is used to output the mean and variance parameters of each time step within a future preset period through a fully connected layer based on the deep semantic features, and to construct the fluctuation probability distribution that follows a Gaussian distribution based on the mean and variance parameters, so as to quantify the uncertainty risk of market fluctuations.

4. The real-time decision-making system for supply chain management based on market data analysis according to claim 3, characterized in that, The demand forecast is calculated as follows: The first decoder receives the deep semantic features output by the shared feature extraction layer. ; Preset cycle for the future Inner The region in the first Demand forecast at each time step The calculation formula is: ； in, To predict the cycle length, ∈ ,and For the total number of regions, The deep semantic feature vectors output by the shared feature extraction layer, and ∈ , Indicates the current moment. The dimension of deep semantic features; For the first The weight matrix corresponding to each time step ∈ ; Let be the demand bias vector, and ∈ , This is the activation function used to constrain demand. For the first The region in the first Regional demand forecasts for each time period.

5. The real-time decision-making system for supply chain management based on market data analysis according to claim 3, characterized in that, The fluctuation probability distribution is calculated as follows: The second decoder receives the deep semantic features output by the shared feature extraction layer. and output the mean parameter. and variance parameter : ； ； in, The weight vector is the mean. The weight vector is the log-variance. It is the mean bias vector. This is the variance bias vector; The second decoding head will the first The region in the first The demand at any given time is modeled as a random variable following a Gaussian distribution, wherein the mean parameter... Take the localized demand prediction value of the corresponding region and time step output by the first decoder head. The calculation formula is: ； in, This is the actual demand value. Variance represents the degree of market uncertainty.

6. The real-time decision-making system for supply chain management based on market data analysis according to claim 1, characterized in that, The deviation statistic in the dynamic feedback module is specifically the mean absolute percentage error. The dynamic feedback module is used to trigger the online incremental learning process of the multi-task learning model when the mean absolute percentage error in N consecutive monitoring periods exceeds a preset drift threshold. It uses the actual inventory level and order fulfillment data in the most recent M monitoring periods as training samples to fine-tune the network weight parameters of the shared feature extraction layer in order to suppress model drift.

7. The real-time decision-making system for supply chain management based on market data analysis according to claim 1, characterized in that, The data perception module is deployed in an edge computing node cluster. The edge computing node cluster is configured to perform local preprocessing on data within a preset geographical range from the data source, and only upload the processed feature vector sequence to the cloud server to reduce data transmission bandwidth usage and the computing load of the central node.

8. The real-time decision-making system for supply chain management based on market data analysis according to claim 1, characterized in that, The dynamic feedback module is also configured with a counterfactual reasoning submodule, which includes: When the deviation statistic exceeds the drift threshold, a counterfactual simulation scenario similar to the current market environment is constructed. In the counterfactual simulation scenario, the hyperparameter combination of the multi-task learning model is adjusted, and the optimal hyperparameter update strategy is determined based on the simulation results. Then, the optimal hyperparameter update strategy is applied to the online inference and prediction module.

9. The real-time decision-making system for supply chain management based on market data analysis according to claim 1, characterized in that, The multidimensional raw data stream includes consumer behavior, competitor dynamics, and economic indicators; The collaborative decision-making module is also configured with a global reward function, which is determined by the weighted sum of inventory holding costs, order out-of-stock penalties, and logistics transportation costs. The collaborative decision-making module is also coupled to a blockchain-based distributed ledger. The preset resource constraints are stored in the distributed ledger, which is configured to record the interaction logs and resource usage status of each agent during the generation of real-time decision instruction sets, ensuring the traceability and tamper-proof nature of the decision-making process.

Images