Supply chain intelligent scheduling decision optimization method and system based on big data driving

Abstract

The invention relates to the field of supply chain scheduling decisions, in particular to a supply chain intelligent scheduling decision optimization method and system based on big data driving. The method comprises the following steps: collecting supply chain heterogeneous data streams, extracting abnormal disturbance events, carrying out nonlinear disturbance response mining, and constructing a supply chain disturbance response field; identifying dynamic state parameters of all nodes of the supply chain, analyzing correlation characteristics among the parameters, performing multi-node logistics state transfer evolution, and generating logistics state transfer characteristics of each node; and multi-node delay evaluation is carried out based on the supply chain heterogeneous data flow, global distribution mapping is carried out, and an upstream and downstream logistics state delay characteristic spectrum is constructed. According to the invention, adaptive optimization collaboration between multiple nodes and multiple targets is realized, so that the stability, flexibility and economic benefit of the whole supply chain system are effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of supply chain scheduling decision-making, and in particular to a supply chain intelligent scheduling decision optimization method and system driven by big data. Background Art

[0002] With the continuous advancement of globalization and the rapid development of information technology, supply chain management is facing unprecedented complexity and uncertainty. Especially with the widespread adoption of emerging technologies such as big data, the Internet of Things (IoT), cloud computing, and artificial intelligence, traditional supply chain scheduling models are no longer able to meet the current dynamic, ever-changing, and high-frequency business needs. Efficiently integrating multi-source heterogeneous data to achieve intelligent perception, dynamic response, and global optimization of the supply chain has become a core challenge that urgently needs to be overcome in modern supply chain management.

[0003] In traditional supply chain scheduling systems, decision-making relies primarily on static models, historical experience, and periodic manual adjustments. This approach suffers from numerous issues, including delayed response, large forecast errors, and low scheduling resource utilization. In particular, when faced with disruptive factors such as sudden order changes, logistics delays, and imbalances in upstream and downstream collaboration, traditional scheduling methods struggle to promptly identify and intelligently respond to issues, significantly compromising the overall operational efficiency and robustness of the supply chain. Coupled with the changing global trade environment, the frequent occurrence of public emergencies, and the increasing demand for delivery accuracy and timeliness, supply chain scheduling optimization urgently needs to evolve towards greater intelligence, real-time capabilities, and predictive capabilities.

[0004] Some existing supply chain scheduling optimization approaches have attempted to incorporate digital technologies, such as ERP systems and MES scheduling platforms. While these have improved management automation to a certain extent, these systems generally rely on rule-driven process control, making it difficult to deeply explore the complex relationships between multiple nodes, multiple events, and multiple factors and to make dynamic scheduling decisions. Furthermore, current scheduling systems lack a unified analysis and processing mechanism for multi-source, heterogeneous data (such as order data, logistics trajectories, inventory dynamics, weather conditions, and contract terms). This leads to severe information silos and low data utilization, limiting the effectiveness of scheduling optimization strategies. Summary of the Invention

[0005] In order to solve the above technical problems, the present invention proposes a supply chain intelligent scheduling decision optimization method and system based on big data drive to solve at least one of the above technical problems.

[0006] To achieve the above objectives, the present invention provides a supply chain intelligent scheduling decision optimization method based on big data driving, comprising the following steps: Step S1: Collect heterogeneous data streams from the supply chain, extract abnormal disturbance events, and conduct nonlinear disturbance response mining to construct a supply chain disturbance response field; Step S2: Identify the dynamic state parameters of all nodes in the supply chain, analyze the correlation characteristics between the parameters, and perform multi-node logistics state transition evolution to generate the logistics state transition characteristics of each node; Step S3: Perform multi-node delay evaluation based on heterogeneous data streams in the supply chain, perform global distribution mapping, and construct a delay feature map of upstream and downstream logistics status; Step S4: Based on the upstream and downstream logistics state delay characteristic map and the supply chain disturbance response field, the logistics state transition characteristics are analyzed in real time and multi-round game optimization is performed to build a supply chain multi-round game optimization model; Step S5: Identify real-time contract terms information, conduct contract intention change mining, and then calculate the optimal buffer time window to obtain the optimal buffer time window; Step S6: Based on the optimal buffer time window, the supply chain multi-round game optimization model is simulated for contract modification scheduling, and multi-round decision-making with minimum cost is performed to build an order mutation decision optimization engine.

[0007] In this specification, a supply chain intelligent scheduling decision optimization system based on big data driving is provided, which is used to execute the supply chain intelligent scheduling decision optimization method based on big data driving as described above, including: The disturbance response mining module is used to collect heterogeneous data streams from the supply chain, extract abnormal disturbance events, conduct nonlinear disturbance response mining, and construct a supply chain disturbance response field; The state transition evolution module is used to identify the dynamic state parameters of all nodes in the supply chain, analyze the correlation characteristics between parameters, and perform multi-node logistics state transition evolution to generate logistics state transition characteristics for each node; The delay mapping module is used to perform multi-node delay evaluation based on heterogeneous data streams in the supply chain, and perform global distribution mapping to build a delay feature map of upstream and downstream logistics status; The game optimization module is used to conduct real-time supply chain information analysis and multi-round game optimization based on the upstream and downstream logistics status delay characteristic map and supply chain disturbance response field to build a supply chain multi-round game optimization model; The buffer time window calculation module is used to identify real-time contract terms information, conduct contract intention change mining, and then calculate the optimal buffer time window to obtain the optimal buffer time window; The multi-round decision module is used to simulate contract modification scheduling in the supply chain multi-round game optimization model based on the optimal buffer time window, perform multi-round decision-making with minimum cost, and build an order mutation decision optimization engine.

[0008] The present invention achieves the following beneficial effects: It fully integrates multi-source, heterogeneous data streams from procurement, transportation, warehousing, markets, and the external environment, improving the timeliness and multidimensional coverage of data perception. Through nonlinear perturbation recognition, it accurately extracts dynamic response patterns for emergencies such as order delays, material shortages, traffic blockages, and market anomalies. The constructed perturbation response field has the ability to perceive real-time changes and remember historical responses, serving as the steady-state foundation for subsequent forecasting and scheduling models. It effectively improves the scheduling model's sensitivity and response speed to non-structural risks, making the decision-making system more resilient to interference. It extracts multi-dimensional, real-time status parameters (such as throughput, order backlog rate, and transportation efficiency) for each supply chain node, laying the data foundation for refined scheduling. By analyzing the evolutionary correlation characteristics between node parameters, it identifies key bottleneck nodes and interlocking impact paths, achieving a transition from point-based monitoring to networked coordination. The construction of node state transition characteristics facilitates dynamic simulation of the operational evolution trends of logistics channels, improving the accuracy of forecasts of medium- and long-term logistics changes. It supports the construction of scheduling strategies for multi-node, multi-level logistics networks, enhancing the system's dynamic controllability and coordination capabilities. Supported by full-chain data, this system enables quantitative assessment of delays at all key nodes within the supply chain network, enhancing the ability to identify bottlenecks. The delay feature map visualizes upstream and downstream logistics delays in a spatiotemporal manner, facilitating the identification of regional or systemic delay risks. It intuitively reflects the responsiveness and coordination of multi-echelon supply networks, providing a key reference for priority scheduling, resource allocation, path reconstruction, and risk early warning. It significantly improves the system's identification granularity for delay risks and its response speed to scheduling instructions, reducing the probability of overall system link imbalance. The system integrates disturbance response and delay features to form real-time feature inputs for game-based scheduling, enhancing the model's adaptability to the complex and dynamic environments of the supply chain. A multi-round game mechanism allows different scheduling objectives (cost, timeliness, and risk) to dynamically collide within the simulation framework to obtain the optimal or near-optimal strategy for the system. The system dynamically evolves and updates scheduling solutions in real time, effectively avoiding overfitting or local optimality issues under a single strategy. This system enhances the scheduling system's intelligent decision-making capabilities in complex and uncertain environments, achieving a qualitative upgrade from static rule-driven to dynamic game-driven scheduling. Text mining based on contract terms and historical behaviors can identify potential deviations in performance intentions of customers or suppliers in advance; through semantic modeling and behavioral trend analysis, it can obtain future risk indicators and avoid the scheduling system's passive response of "after-the-fact"; the calculation of the optimal buffer time window enables the system to have a reserved adaptation period to cope with contract fluctuations and enhance tolerance for performance uncertainty; it can maximize the balance between contract performance efficiency and customer satisfaction, and reduce operational losses caused by misjudgment or delayed response.Before the contract is changed, multiple response strategies are simulated and a scenario-based minimum cost path map is constructed to achieve dual optimization of scheduling costs and response time. Contract adjustments, supply chain game results, risk characteristics and buffer windows are comprehensively considered to significantly improve the resilience of the order scheduling system to sudden events. The constructed order mutation decision optimization engine has the ability to adapt to scenarios, self-adjust and dynamically optimize. It can achieve rapid response and intelligent scheduling to extreme events (such as emergency orders, supply interruptions, and temporary order returns), and improve the sensitivity and resilience of the entire link system. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 This is a schematic diagram of the steps of a supply chain intelligent scheduling decision optimization method based on big data drive of the present invention; Figure 2 Detailed implementation flow chart of step S1; Figure 3 Detailed implementation flow chart of step S2; Figure 4 Schematic diagram of the detailed implementation steps of step S3. DETAILED DESCRIPTION

[0010] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0011] This application example provides a method and system for optimizing supply chain scheduling decisions based on big data. The execution entities of the method and system include, but are not limited to, the following: mechanical equipment, data processing platforms, cloud server nodes, network upload devices, etc. that are equipped with the system, which can be regarded as general computing nodes of this application. The data processing platform includes, but is not limited to, at least one of an audio and image management system, an information management system, and a cloud data management system.

[0012] See also Figures 1 to 4 The present invention provides a supply chain intelligent scheduling decision optimization method based on big data driving, comprising the following steps: Step S1: Collect heterogeneous data streams from the supply chain, extract abnormal disturbance events, and conduct nonlinear disturbance response mining to construct a supply chain disturbance response field; Step S2: Identify the dynamic state parameters of all nodes in the supply chain, analyze the correlation characteristics between the parameters, and perform multi-node logistics state transition evolution to generate the logistics state transition characteristics of each node; Step S3: Perform multi-node delay evaluation based on heterogeneous data streams in the supply chain, perform global distribution mapping, and construct a delay feature map of upstream and downstream logistics status; Step S4: Based on the upstream and downstream logistics state delay characteristic map and the supply chain disturbance response field, the logistics state transition characteristics are analyzed in real time and multi-round game optimization is performed to build a supply chain multi-round game optimization model; Step S5: Identify real-time contract terms information, conduct contract intention change mining, and then calculate the optimal buffer time window to obtain the optimal buffer time window; Step S6: Based on the optimal buffer time window, the supply chain multi-round game optimization model is simulated for contract modification scheduling, and multi-round decision-making with minimum cost is performed to build an order mutation decision optimization engine.

[0013] In the embodiment of the present invention, see Figure 1 , is a schematic flow chart of the steps of a big data-driven supply chain intelligent scheduling decision optimization method of the present invention. In this example, the steps of the big data-driven supply chain intelligent scheduling decision optimization method include: Step S1: Collect heterogeneous data streams from the supply chain, extract abnormal disturbance events, and conduct nonlinear disturbance response mining to construct a supply chain disturbance response field; In this embodiment, supply chain data comes from a wide range of sources, including logistics trajectory data (such as GPS location information and shipping status updates), order event streams (order creation, modification, cancellation, and fulfillment status changes), and real-time environmental and meteorological information (temperature, humidity, rainfall, wind speed, etc.). To ensure data timeliness and integrity, a real-time collection system based on message queues (such as Kafka) and distributed stream processing frameworks (such as Apache Flink) is employed to ensure data inflow within seconds. Logistics trajectory data is collected via IoT devices at a frequency of approximately every 5 seconds. Order event streams are event-driven, with update frequency fluctuating depending on order volume, typically reaching hundreds of records per minute. Meteorological information is accessed through a third-party API and updated every 10 minutes. After collection, the data undergoes format standardization, time alignment, and fusion through an ETL process. A unified timestamp verification mechanism is employed to ensure accurate comparison of the three types of data across the timeline, laying the foundation for subsequent anomaly detection. A multimodal anomaly detection method is employed to extract disturbance events from the fused heterogeneous data streams. Specifically, the system employs a variational autoencoder (VAE) based on a time-series neural network combined with an isolation forest approach. First, the VAE model learns the temporal feature distribution of normal supply chain operations and identifies periods with high reconstruction error as potential anomaly areas. Second, an isolation forest approach is applied to these anomaly areas to further identify outliers, filter out noise, and improve accuracy. In experiments, the model, trained on historical supply chain operation data, achieved an accuracy of 91.3% and a recall of 88.7%. Abnormal events included logistics delays, sudden increases or decreases in orders, and drastic changes in weather conditions, with an average capture latency of less than 30 seconds. Combining these extracted abnormal events, the system focuses on identifying abnormal logistics transitions (such as route deviations and abnormal dwell times), order change nodes (sudden changes in order quantity and cancellation rates), and meteorological sudden changes (sudden drops in temperature and sudden rainstorms). By constructing a multidimensional feature space (time, space, and event type), an algorithm based on density peak clustering (DPC) is used to locate disturbance sudden changes and extract multi-dimensional disturbance feature points. Specific parameters were set, such as a cluster radius threshold of 30 minutes and a spatial neighborhood threshold of 5 kilometers, to ensure that the clustering results accurately reflect local disturbances without being overly dispersed. Experiments showed that this method effectively identified 98% of real disturbance events, with an error rate of 4.5%. Based on the identified multi-dimensional disturbance mutation points, the nonlinear response relationship of the supply chain system to disturbances was further explored. A spatiotemporal dynamic network analysis method based on tensor decomposition was used to map the disturbance points and the node state changes (throughput, inventory fluctuations, delays, etc.) they triggered into a high-dimensional nonlinear response tensor. Using non-negative tensor decomposition (NTF) technology, the spatiotemporal propagation path and intensity of the disturbance impact were captured, revealing the complex coupling relationship between the disturbance and the supply chain state variables.In the experiment, a regional supply chain network consisting of 50 nodes and three months of operational data was selected. The nonlinear response model accurately recovered the disturbance propagation path, achieving an explanation rate of 87.4%. The supply chain disturbance response field constructed based on this model can reflect the overall dynamic behavior of the supply chain system under multi-dimensional disturbances in real time, providing accurate decision-making support for intelligent scheduling strategies.

[0014] Step S2: Identify the dynamic state parameters of all nodes in the supply chain, analyze the correlation characteristics between the parameters, and perform multi-node logistics state transition evolution to generate the logistics state transition characteristics of each node; In this embodiment, the dynamic state parameters of supply chain nodes directly reflect the node's operating status and service capabilities. They primarily include: node throughput (the number of orders or goods processed per unit time), wait time (the average dwell time of goods or orders at a node), inventory fullness (the ratio of current inventory to maximum inventory), and logistics inflow and outflow rate (the rate of goods flowing into and out of the node). Data sources include logistics scanning systems, warehouse management systems (WMS), order management systems (OMS), and real-time device data collected by sensors. Raw data is first preprocessed, including outlier filtering, missing value interpolation, and time synchronization, to ensure data accuracy and usability for dynamic state parameter calculation. Dynamic parameters for each node are calculated using a rolling window (e.g., 5-minute or 15-minute window) to ensure temporal continuity and timeliness. In the experiment, the rolling window for node throughput was set to 15 minutes, with a throughput fluctuation frequency of approximately 0.05 Hz. The wait time statistical period was also 15 minutes to capture short-term fluctuations rather than noise. Inventory fullness is updated in real time based on daily inventory count data, and logistics inflow and outflow rate is calculated jointly with sensor and order flow data. After identifying the node's dynamic state parameters, we further explored the correlation characteristics between these parameters, providing a basis for understanding the node's internal and external operating mechanisms. We focused on exploring the temporal correlation characteristics between throughput and waiting time, as well as the joint fluctuation patterns between inventory fullness and logistics inflow and outflow rates. To investigate the temporal correlation between throughput and waiting time, we employed a multi-analysis approach based on Granger causality tests and mutual information to verify whether changes in throughput predict changes in waiting time, thereby revealing potential causal relationships. Experiments demonstrated that the Granger causality relationship between throughput and waiting time was significant in 85% of nodes, and mutual information analysis also showed a correlation exceeding 0.7, reflecting the strong influence of node throughput on waiting time. The correlation analysis between inventory fullness and logistics inflow and outflow rates employed a combination of dynamic time warping (DTW) and a convolutional neural network (CNN) model to identify the synchronous fluctuations and time lag characteristics of the two parameters over time. Experiments at 30 typical warehouse nodes revealed that inventory fullness fluctuations precede changes in inflow and outflow rates by approximately 10-20 minutes, suggesting that inventory adjustments have a predictive effect on logistics flow. Based on the dynamic parameters of each node and their associated characteristics, a logistics state transition model for multiple supply chain nodes was constructed to describe the evolution of node status over time and in response to upstream and downstream influences. A Markov chain-based multi-state transition model was employed to categorize the logistics status of each node into four categories: "normal," "congested," "buffered," and "abnormal." State transition probabilities were dynamically updated using historical time series data statistics and associated parameter adjustments. The model incorporates associated characteristic adjustment factors, such as the causal strength of throughput versus waiting time, as a weighting factor for state transition probabilities, reflecting the sensitivity of node status to parameter changes.In the experiment, the state transition model was tested in an actual supply chain network, with a node state prediction accuracy of 82.5%, and its early warning capability for congestion and abnormal conditions was significantly better than the traditional threshold method. According to the above model, a logistics state transition feature is generated for each node, specifically in the form of a state label and a transition probability matrix in the form of a time series, which is used for subsequent scheduling decision input and dynamic risk assessment. This feature includes the current operating status of the node, future short-term state predictions, and state switching trends, supporting the scheduling system to perform real-time evaluation and scheduling optimization of node loads. Taking a large manufacturing supply chain as an example, the generated logistics state transition feature was continuously monitored for three months, successfully capturing multiple key abnormal congestion points, helping the scheduling system to adjust the transportation plan in a timely manner, reducing the overall delay rate by 12% and shortening inventory turnover time by 8%.

[0015] Step S3: Perform multi-node delay evaluation based on heterogeneous data streams in the supply chain, perform global distribution mapping, and construct a delay feature map of upstream and downstream logistics status; In this embodiment, the heterogeneous data stream of the supply chain contains logistics trajectory timestamps, order event timestamps, and environmental data timestamps. First, these timestamps are systematically collected and preprocessed to ensure the consistency of the time base of multi-source data. In response to time synchronization errors, the heterogeneous timestamps are uniformly calibrated based on the Network Time Protocol (NTP) correction and dynamic time warping (DTW) algorithm to eliminate time deviations between systems. In the experiment, the timestamp calibration accuracy is controlled at the millisecond level, meeting the fine-grained requirements of logistics operations. Then, the timestamps of key dynamic state parameters are extracted from each node, such as the node throughput change point, order status change time, and logistics scanning time, to provide basic data for subsequent delay calculations. Based on the node dynamic state and timestamp data, three types of core delays in the supply chain are identified: Information latency: This refers to the time difference between when order, contract, or scheduling information is sent and when it actually takes effect at the receiving node, reflecting the efficiency of information transmission. By analyzing the time difference between the order event stream and the corresponding logistics trajectory, we use time window technology (such as a 10-minute rolling window) to statistically analyze the distribution of information transmission latency.

[0016] Logistics delay refers to the amount of time that the actual transportation or handling of goods exceeds the expected time. This is calculated based on logistics trajectory data: the actual transportation time between nodes minus the planned transportation time. The experiment sets a threshold for expected transportation time, based on the 95th percentile of the average transportation route time.

[0017] Decision delay reflects the time difference between the issuance of a scheduling instruction and the start of an action at the execution node. By analyzing scheduling system logs and execution feedback times, decision response lags can be identified. The statistical period for decision delay is generally set at the hourly level. In the implementation, a sliding window approach is combined with anomaly detection to eliminate occasional outliers and ensure the robustness of delay estimation. Experimental results show that the average values ​​for the three types of delay are 3.2 minutes for information delay, 18 minutes for logistics delay, and 12 minutes for decision delay. Each delay exhibits significant time-varying characteristics. Delays between supply chain nodes exhibit non-aligned time series, making traditional synchronization analysis difficult to accurately reflect the true delay state. To address this issue, dynamic time warping (DTW) is used to nonlinearly align inter-node delay series and assess the degree of delay matching and time-lag relationships between multiple nodes. The DTW method can handle delay variations at different time steps and rates across supply chain nodes, achieving optimal matching of delay time series between nodes. The degree of timing misalignment between upstream and downstream nodes is quantified by calculating the DTW distance and alignment path. In an experiment, 10 key nodes in a manufacturing supply chain network were analyzed, and the average DTW distance was reduced by 15%, demonstrating that alignment effectively improves the accuracy of delay time series comparisons. Based on the non-aligned multi-node delay assessment results, a delay characteristic map was constructed to reflect the delay propagation and impact between upstream and downstream nodes in the global supply chain. This map uses nodes as vertices, and delay intensity and transfer relationships as edge weights, described using a weighted adjacency matrix from graph theory. A fusion analysis method based on spectral clustering and graph convolutional networks (GCNs) was used to identify key nodes and bottleneck areas in delay propagation. Using supply chain network data consisting of 50 nodes, the GCN model accurately captured the delay propagation path and key impact areas, achieving an 89% accuracy rate in delay propagation path prediction, significantly outperforming traditional static analysis. The visualization of the delay characteristic map displays the multi-dimensional distribution of information, logistics, and decision delays between nodes, assisting the scheduling system in identifying bottlenecks and implementing targeted optimization adjustments.

[0018] Step S4: Based on the upstream and downstream logistics state delay characteristic map and the supply chain disturbance response field, the logistics state transition characteristics are analyzed in real time and multi-round game optimization is performed to build a supply chain multi-round game optimization model; In this example, the upstream and downstream logistics state delay feature maps constructed in the previous stage are dynamically fused with the nonlinear disturbance response field to form a multidimensional decision input space. This space integrates each node's state transition probability, current delay level (information delay, logistics delay, decision delay), and sensitivity indicators in the disturbance response tensor (such as inventory fluctuation coefficient and response nonlinearity strength), forming a cross-temporal, spatial, and behavioral portrait of supply chain operations. To achieve information fusion, tensor concatenation and dimensionality reduction (such as principal component analysis (PCA) or autoencoders (AE)) are used to uniformly map high-dimensional, heterogeneous features into a unified feature space suitable for decision modeling. In experiments, the fused feature dimensions were compressed to 32. The retained main features covered key elements such as delay propagation paths (edge ​​weights), node disturbance response thresholds, and logistics state transition rates. The cumulative explained variance exceeded 91.2%, ensuring the expressive power of the input features. Based on the fused supply chain state features, a multi-round simulation game model was constructed to simulate node response behavior under different scheduling strategies and the overall system evolution. In the game model, each node is abstracted as an intelligent agent, whose game strategies include cost-prioritized scheduling, service-prioritized scheduling, robustness-enhanced scheduling, and speed-burst response scheduling. Each strategy corresponds to a different reward function weight combination, emphasizing time cost, inventory security, system stability, or response speed.

[0019] To simulate the interactive game relationship between strategies, a multi-agent reinforcement learning (MARL) approach was employed. Each agent determines its scheduling behavior based on the current environment state (i.e., the state transition characteristics of the node) and its selected strategy. The MARL framework uses a deep Q-network (DQN) as the core algorithm, and the reward function is configured with multi-dimensional objectives such as logistics cost minimization, delay control, inventory stability, and order fulfillment rate. During training, offline simulations were conducted using three months of historical supply chain operation data. Each game lasted 8 hours, and a total of 5,000 rounds were trained. An ε-greedy strategy was used to balance exploration and exploitation. After training, the system converged to a set of high-performing game strategy combinations, reducing average scheduling response time by 15.4% and overall inventory fluctuation by 12.8%. Based on the simulation training results, a multi-round game optimization model was further constructed to achieve continuous strategy evolution and feedback updates. This model uses system performance indicators (such as revenue dynamics, adaptability trajectories, and strategic synergy factors) after each round of game play as input to perform an evolutionary fit on the scheduling strategies, constructing a strategy game evolution curve to quantify the relative advantages and synergy potential among different strategies. The evolutionary analysis utilizes a genetic algorithm combined with a clustering method. The strategy space is first encoded and fitness evaluated, and then the strategy evolution process is simulated through crossover, mutation, and selection operations. Across 50 simulations, the average system performance improvement of the evolved strategy combination was 18.7%, significantly higher than that of the fixed-strategy baseline model. Ultimately, the system feeds the evolved optimal strategy combination back into the supply chain scheduling engine as the dynamic strategy set for actual deployment, while retaining multiple alternative strategies to accommodate potential future disruption scenarios. The entire multi-round game optimization model operates through continuous learning and adaptive evolution, supporting efficient and stable intelligent scheduling in supply chains under highly uncertain environments.

[0020] Step S5: Identify real-time contract terms information, conduct contract intention change mining, and then calculate the optimal buffer time window to obtain the optimal buffer time window; In this embodiment, the system accesses the real-time electronic contract system between supply chain enterprises (e.g., an electronic text interface based on an API or optical character recognition technology) and performs structured extraction of the latest contract terms. Key fields include delivery time, quantity, logistics constraints, service level requirements, and change clauses. The BERT (Bidirectional Encoder Representations from Transformers) model, a natural language processing (NLP) technology, is used to embed contract semantics and extract contextual semantic features at the sentence and paragraph levels. Furthermore, in conjunction with the contract version management mechanism, the system compares differences between the current version and previous versions. A semantic matching-based change annotation algorithm (such as SimHash or SBERT semantic distance threshold recognition) is used to identify clauses with substantive changes. In the experimental setup, clause segments with a semantic similarity below 0.85 were identified as having intended modifications, with the system achieving an average recognition accuracy of 92.6%. After identifying contract semantic changes, further structural mining of contract intent is performed. By building a "clause content-intent label" mapping model, natural language descriptions are converted into machine-readable intent categories, such as delivery time compression, order quantity increase, quality improvement, and logistics method adjustment. The system then assesses the intensity of the intent change (e.g., quantified in terms of hours of time compression or percentage increase in quantity). The system then maps the contract intent change information to the corresponding supply chain nodes and analyzes how the change will cause changes in node behavior. Using node behavior logs (such as logistics receipt records, job start times, and scheduling response times) as a reference, behavioral change trends before and after the intent change are compared to identify and mark behavioral transition points. Behavioral transition points are considered key triggers for the transition from a stable state to a disrupted state in the supply chain. After identifying behavioral transition points, the actual impact of these transitions on upstream and downstream nodes needs to be quantified. A perturbation propagation function is used to model the fluctuations in behavioral transitions, including multi-dimensional parameters such as information reception delay, behavioral adjustment lag, and actual execution offset. Ultimately, the intensity of the fluctuations caused by the contract intent change is determined for each node. The behavioral fluctuations are superimposed along the propagation paths between upstream and downstream nodes to form a cumulative time offset curve. Based on this curve, an improved Peak-Holding Buffer Model was used to calculate the optimal buffer time window. This model aims to determine the minimum buffer duration required to ensure that more than 95% of node behaviors can stably converge to the scheduling rhythm of the new contract intent. In the experimental setup, backtesting was performed using node feedback data from the past 10 behavior cycles (4 hours per cycle). The resulting output was a recommended buffer duration for each type of contract intent. For example, the recommended buffer window for a sudden increase in order volume is 6.5 hours, while for a sudden change in logistics methods, it is 9 hours.The system dynamically transmits the calculated optimal buffer time window to the scheduling game engine as input, adjusting the frequency and tempo of different strategies. For example, upon identifying a high-intensity shift in intent within a contract, the system can automatically extend the response window for some scheduling strategies to prevent frequent strategy switching due to incomplete node adaptation. Furthermore, the scheduling system can utilize the buffer time window to implement temporary transition strategies (such as dual-track scheduling and partial resource locking) to ensure a smooth transition.

[0021] Step S6: Based on the optimal buffer time window, the supply chain multi-round game optimization model is simulated for contract modification scheduling, and multi-round decision-making with minimum cost is performed to build an order mutation decision optimization engine.

[0022] In this example, contract modification events are introduced as intervention variables based on an existing multi-round game optimization model. Contract modification simulation samples are generated by mixing synthetic data with real order data, including but not limited to delivery time advancement, order quantity expansion, service level upgrade, and node reallocation. In the experiment, five contract modification scenarios were set, each simulated 100 times over a two-week continuous operation cycle. The system maps these modification events to corresponding nodes in the logistics state transition graph and introduces a node state transition simulation mechanism within an "optimal buffer time window." This mechanism applies delayed loading, capacity restrictions, or priority adjustments to node behavior within the buffer window to prevent sudden changes from causing nonlinear impacts on the entire system. The scheduling simulation engine operates based on a multi-agent game system, maintaining the original four policy types (cost priority, service priority, robustness enhancement, and speed response). The engine continuously evolves the system at a set time granularity (e.g., one simulation round per hour). During the contract modification scheduling simulation, the system gradually evaluates the execution performance of different policy combinations, focusing on factors such as order fulfillment cost, node response efficiency, inventory pressure fluctuations, and transportation route utilization. During each round of the game, the system records the payoff performance of the current strategy in the game environment and updates its strategy weights using a reinforcement learning mechanism (such as DQN or PPO) to gradually minimize the overall cost. To ensure fair evaluation, the experimental design introduced a control group (without a buffer window) and a test group (with a buffer window) to conduct comparative analysis under multiple rounds of contract modification. The results show that the buffer mechanism effectively mitigates node response delays caused by sudden order changes. The average order fulfillment cost of the test group decreased by 12.4%, abnormal inventory fluctuations decreased by 19.3%, and overall system stability was significantly improved. After multiple rounds of simulation, the system modeled the optimal combination of strategies under each type of contract change scenario and the game evolution trajectory, establishing a three-dimensional mapping model of "order change-strategy response-result payoff." This model uses a fusion of multivariate regression and decision tree methods for predictive modeling. Based on inputs such as the current contract modification content, supply chain status characteristics, and disturbance response intensity, it can quickly recommend the optimal strategy combination and game path. The resulting "Order Change Decision Optimization Engine" possesses three core functions: real-time analysis of order change triggers and impact paths; automatic matching of buffer windows with scheduling strategy cadence; and output of minimum-cost scheduling decision recommendations. The engine is deeply integrated with the main scheduling system. Upon identifying contract changes, it rapidly invokes the prediction module to generate optimized decision results and drives logistics task reconfiguration and resource reallocation through APIs or scheduling platform interfaces.

[0023] In this embodiment, refer to Figure 2 , is a flowchart of the detailed implementation steps of step S1. In this embodiment, the detailed implementation steps of step S1 include: Collect heterogeneous supply chain data streams, including logistics trajectory data, order event streams, and real-time environmental weather information; Identify abnormal noise in heterogeneous data streams of the supply chain and perform adaptive filtering and noise reduction to obtain filtered and optimized heterogeneous data streams; Based on filtering and optimizing heterogeneous data streams, we can identify abnormal logistics transition trajectories, order change nodes, and meteorological mutation points to obtain multi-dimensional disturbance mutation points. Nonlinear disturbance response mining is performed on multi-dimensional disturbance mutation points to construct a supply chain disturbance response field.

[0024] In this embodiment, intelligent supply chain scheduling optimization requires establishing a multi-source, heterogeneous data collection mechanism to build a comprehensive data support system. "Heterogeneous data streams" refer to data streams from various links in the supply chain, with varying formats and update frequencies. These data primarily include three key categories: The first is logistics trajectory data, collected primarily through GPS devices, RFID readers, and vehicle-mounted terminals, recording dynamic behavioral characteristics such as the location, speed, and docking time of transport vehicles; the second is order event streams, collected from enterprise ERP systems, OMS systems, and other systems, recording various event nodes during order creation, modification, and fulfillment; and the third is real-time environmental weather information, including temperature, humidity, rainfall, wind speed, and road condition warnings. This is typically obtained through accessing third-party APIs such as the National Meteorological Administration platform, AutoNavi, or Baidu Maps' weather modules. During data collection, a highly concurrent and scalable data access framework is required. For example, Apache Kafka can be used for real-time streaming data access, employing a distributed architecture to support data throughput exceeding 1TB per day. To ensure time series alignment and data integration, data format standards must be standardized, such as using a unified data protocol format based on JSON or Protocol Buffers. Furthermore, to ensure data real-time performance and integrity, a multi-level caching mechanism and a resumable transmission strategy are necessary. To address data latency issues in logistics trajectory and meteorological data, a time window mechanism (such as a 5-minute rolling window) is used for synchronization, ensuring that the data has a consistent time dimension before entering the subsequent processing module. This stage is the foundation for subsequent data noise reduction, mutation identification, and scheduling decision optimization. Heterogeneous data collection is often accompanied by a significant amount of abnormal noise, such as trajectory point jumps due to network latency, lost order event records, and frequent jitter in meteorological data. Left unaddressed, these issues can severely impact modeling accuracy. Therefore, anomaly identification and adaptive filtering noise reduction mechanisms are necessary during data preprocessing to improve data quality. To identify abnormal noise, we employed a rule-based and model-based approach. At the rule level, domain knowledge rules were established to identify overt anomalies such as "significant displacement in a short period of time" (e.g., displacement greater than 5 kilometers within a minute) and "illogical order status transitions" (e.g., jumping directly from order placement to delivery completion). At the model level, we combined unsupervised algorithms such as Isolation Forest and Local Outlier Factor (LOF) to perform multi-scale clustering on multidimensional data, identifying more subtle anomalies. During the experiment, we used three months of trajectory and order data from 10 logistics companies as a sample set. A validation set was constructed through manual annotation. The recognition accuracy of various models was compared, and the Isolation Forest model was ultimately selected, achieving a recognition accuracy of 91.3%.

[0025] For noise reduction, an adaptive Kalman filter algorithm is used to smooth dynamic trajectory and time series data. Unlike traditional Kalman filters, this adaptive method dynamically adjusts the prediction covariance matrix based on data noise characteristics, thereby improving robustness in response to sudden changes in the environment. For example, when processing meteorological data, an adaptive threshold adjustment mechanism is implemented for wind speed mutation points to prevent misidentification of actual meteorological fluctuations as anomalies. In actual deployment, the system sets a default sliding window of 10 minutes and a maximum state transition covariance adjustment of 20%. This ensures that the filtered data maintains true signal changes while minimizing high-frequency noise. This step lays a high-quality data foundation for subsequent identification of trajectory anomalies and disturbance nodes. After data cleaning and noise reduction, the deep event recognition phase for heterogeneous data begins. The goal is to extract "mutation points" that may trigger supply chain fluctuations from the three dimensions of trajectory, order, and weather—namely, abnormal logistics transition trajectories, key order change nodes, and meteorological environmental mutation points—to construct a global disturbance map. Logistics abnormal trajectory identification utilizes a dual approach based on trajectory pattern matching and spatiotemporal anomaly detection. First, a normal trajectory template library is constructed by clustering historical trajectory data (for example, K-means clustering yields 20 common route models). Then, cosine similarity matching is performed on the current trajectory. Trajectories below a threshold of 0.85 are identified as anomalous. Next, a Hidden Markov Model (HMM) is combined with temporal trajectory state inference to identify abnormal state transitions, particularly those caused by prolonged vehicle quiescence, circuitous routes, and frequent deviations from main roads. This module achieves recognition accuracy of 89.6%. Order change node identification is based on order event time series modeling. An LSTM model is used to model the order lifecycle and capture nonlinear patterns in state transitions. For example, experiments have found that an abnormal increase in the interval between "order placement" and "collection" often indicates delays in warehousing or human resources. By setting the average state transition time for the baseline model (for example, a standard collection time of 2 hours), any deviation exceeding 3σ is flagged as a sudden event.

[0026] Meteorological mutation points are identified through a combination of statistical learning and mutation detection. For example, the CUSUM algorithm (Cumulative Sum Control Chart) is used to monitor mean shifts in continuous meteorological data and rapidly detect sudden changes in indicators such as rainfall, wind speed, temperature, and humidity. Furthermore, a meteorological level coding system is introduced to standardize the impact of various meteorological events, with level 3 representing a medium-risk mutation. The final step is to systematically model the identified multi-dimensional disturbance mutation points to construct a "supply chain disturbance response field" for scheduling decisions. This response field can be viewed as a four-dimensional tensor structure, with the dimensions representing disturbance type (trajectory, order, weather), geographic location, timestamp, and disturbance intensity. The goal is to quantify the chain reaction triggered by each type of disturbance in the supply chain system. To achieve this, nonlinear dynamic causal modeling methods (such as Granger causal graphs and TCN temporal convolutional networks) are first used to model the response relationship between the mutation points and subsequent order fulfillment indicators (such as delivery delay rate and logistics cost). In an experiment, a causal model was constructed based on three months of regional data from Shandong Province. The model found that the average delivery delay at meteorological mutation points was 3.5 hours, the average delay at trajectory mutation points was 2.1 hours, and the order mutation point reduced the fulfillment success rate by 12.3%. Secondly, a disturbance response function R(p, t, s) was introduced, where p represents the mutation point type, t is the time delay, and s is the geospatial coordinate. Bayesian optimization was used to learn the nonlinear response weights between each disturbance point and the system feedback. This function enables the preemptive simulation of intervention effects when predicting future scheduling tasks, for example, identifying potential high-risk disturbance areas within a city within the next two hours. Finally, the response function was embedded in the supply chain scheduling engine, enabling the generation of intelligent scheduling strategies based on the disturbance field through dynamic route adjustments, order priority changes, and warehouse resource reallocation. After deployment, the model was tested in an A / B test in the East China region, resulting in a 12% reduction in the average delivery time and a 15.7% decrease in the delay rate, validating the core value of disturbance response modeling for intelligent scheduling optimization.

[0027] In this embodiment, the specific steps of mining the nonlinear disturbance response of multi-dimensional disturbance mutation points and constructing the supply chain disturbance response field are as follows: Calculating the disturbance amplitude and duration of the multi-dimensional disturbance mutation point; Conduct disturbance propagation trend analysis on multi-dimensional disturbance mutation points in the supply chain to obtain the propagation trend characteristics of each disturbance mutation point; Quantifying the impact of the disturbance based on the disturbance amplitude, duration, and propagation trend characteristics to generate a fluctuation state of the supply chain disturbance point; Performing local disturbance correlation evolution on the multi-dimensional disturbance mutation points of the supply chain to identify the local disturbance correlation area; Based on the fluctuation state of the supply chain disturbance point, the nonlinear disturbance response of the local associated area of ​​the disturbance is mined to generate the nonlinear disturbance response tensor of multiple nodes in the supply chain; The nonlinear disturbance response tensor is mapped across the entire graph to construct a supply chain disturbance response field.

[0028] In this embodiment, after identifying multi-dimensional perturbation mutation points, the perturbation characteristics of each perturbation point must first be quantified. These perturbation characteristics primarily include two key metrics: perturbation magnitude and perturbation duration. This step aims to construct a quantitative profile of the perturbation point's intensity, which serves as the basis for subsequent perturbation impact analysis. Specifically, the perturbation magnitude reflects the instantaneous degree of deviation of the perturbation point from key system variables. For example, in logistics trajectory data, the magnitude can be defined as the deviation of displacement velocity from the normal velocity per unit time. In meteorological perturbations, the magnitude can be calculated as the maximum jump in meteorological indicators (such as rainfall and wind speed) within 5 minutes. In actual experiments, the intervention threshold for wind speed perturbations was set at ±6 m / s. If this change was exceeded within 1 minute, it was marked as a high-intensity perturbation, with a magnitude level of 3 (high). The perturbation duration is defined as the time span during which the perturbation remained within the abnormal range. For example, a storm impact lasted from 1:45 PM to 4:10 PM, a duration of 145 minutes. To more accurately model the characteristics of persistent disturbances, a sliding time window mechanism (10-minute window size, 1-minute sliding step) is introduced. Within the window, segments where the variable is continuously non-stationary are evaluated, thereby extracting valid disturbance segments. Furthermore, in the order data, the event interval is used as the reference duration. For example, if the time span from "payment to delivery" exceeds 3σ, it is considered a slow disturbance. All disturbance amplitudes and durations are normalized to the interval [0, 1] and stored as disturbance feature vectors, providing basic data support for the subsequent propagation trend modeling. The core of disturbance propagation trend analysis lies in identifying the outward diffusion mode and path of each sudden event in the spatial and temporal dimensions, similar to the mapping of infectious source diffusion models in supply chain networks. The goal of this step is to generate a "propagation trend profile" for each disturbance point, which includes spatiotemporal dynamic characteristics. Specifically, the response delay and peak propagation path of the disturbance are modeled in the temporal dimension. A combination of time series autoregressive analysis (ARIMA) and variational autoencoders (VAE) is used to extract the dynamic evolution trajectory of the disturbance point's impact on the subsequent system response variable. For example, the logistics delays caused by a trajectory mutation point on subsequent order nodes can reach their peak impact at T+30 minutes and then decay to normal at T+90 minutes. In the experiment, the time step was set to 10 minutes, and the system response was sampled over 120 minutes. By constructing a disturbance response trajectory curve, its propagation slope, impact half-life, peak response time, and other characteristics were calculated.

[0029] In the spatial dimension, graph-structured propagation modeling is used to identify the path along which a disturbance propagates from its source to other nodes. A weighted directed graph is constructed, using warehouses, distribution centers, and stores in the supply chain network as nodes and order flows and transportation routes as edges. The PageRank propagation scoring algorithm is then used to analyze the response strength and lag time of each node to the initial disturbance. Experimental analysis reveals that when the initial disturbance point is located in a first-tier urban distribution center, its propagation depth and breadth are significantly higher than those at edge warehouses (with an average of 17.3 propagation nodes vs. 6.1). Finally, a propagation trend feature vector is generated for each disturbance point, including information such as propagation depth, speed, spatial sweep coefficient, and response lag time. After obtaining the basic and propagation characteristics of the disturbance point, the impact of the disturbance is comprehensively quantified to form a "disturbance point fluctuation state" to more scientifically guide the scheduling system in identifying high-risk nodes. The core of this step is to construct a disturbance intensity scoring model and transform the disturbance point into a decision input unit with a quantitative risk label. The specific method is to define a disturbance impact function based on the disturbance feature vector: ,in M represents the disturbance amplitude, D represents the duration, T_s represents the propagation speed, and T_p represents the spatial radius. An XGBoost model was used to conduct supervised learning modeling of historical disturbance events and their actual impact on key supply chain metrics (such as delivery delay rate and customer satisfaction decline rate). The training sample size was approximately 120,000 disturbance instances, and a mean square error (RMSE) of 0.036 was achieved through five-fold cross-validation. The disturbance impact score was normalized to a fluctuation index (fluctuation index) between 0 and 100 and categorized into five levels (no fluctuation, slight fluctuation, moderate fluctuation, significant fluctuation, and extreme fluctuation), corresponding to different levels of scheduling intervention. This fluctuation state was ultimately associated with the disturbance point and became an important reference parameter for prioritizing the scheduling system's response. In a real-world deployment test, the application of this fluctuation state classification enabled the system to identify nearly 48% of high-impact disturbance points in advance and initiate route adjustments 18 minutes in advance. Disturbances in supply chain networks often do not exist in isolation, but rather co-evolve in localized regions with specific spatiotemporal patterns. Therefore, it is necessary to cluster all disturbance points and perform regional modeling to identify local hotspots or secondary propagation areas of disturbances to improve response accuracy. This approach employs a disturbance evolution identification mechanism based on spatiotemporal clustering. First, all disturbance points are clustered according to their geographic coordinates and timestamps using a modified DBSCAN (Density-Based Spatial Clustering with Noise) algorithm, with a distance threshold of 3 km and a time threshold of 30 minutes. This algorithm effectively eliminates isolated disturbance points and identifies multiple disturbance evolution clusters. Second, the relationships between disturbance points within each cluster are analyzed to construct a co-evolution map of disturbance points and assess their propagation coordination (e.g., Jaccard coefficient and temporal drift alignment). In an experiment, data from the past six months in Jiangsu Province were analyzed, and a total of 416 local disturbance correlation areas were identified. The core area contained an average of 12.6 disturbance points, with the longest evolution duration reaching 5.2 hours. These areas are often high-incidence areas for scheduling failures and contract fulfillment obstacles. Within each associated region, key nodes (such as core transmission points and secondary outbreak points) are further extracted, and a disturbance channel model is established to provide a structural foundation for subsequent disturbance response optimization. After identifying the disturbance-associated regions, the nonlinear responses of multiple supply chain nodes within these regions to the disturbance are further characterized. This involves establishing a nonlinear tensor structure of disturbance input and response output, which serves as a mathematical representation of the dynamic responses between multiple nodes. The constructed response tensor is a three-dimensional structure R(i, j, t), representing the response value of the i-th disturbance point to the j-th node at the t-th time. This construction method, based on a fusion model of a temporal convolutional network (TCN) and an attention mechanism, can capture the dynamic weight relationship between long-term temporal dependencies and disturbances.The input is the perturbation point's fluctuating state vector and historical node operating indicators (such as fulfillment rate and inventory changes). The output is the perturbation response value for each node's future state. The model was trained using 300,000 sample pairs, using the Adam optimizer with an initial learning rate of 0.001 and 100 training iterations, ultimately achieving a MAE of 0.027 on the validation set. This tensor can be used to simulate supply chain system responses to different perturbation input scenarios. It is widely used in scheduling simulations and is particularly suitable for predicting the overall evolution of the system when some nodes are disturbed.

[0030] In this embodiment, refer to Figure 3 , is a flowchart of the detailed implementation steps of step S2. In this embodiment, the detailed implementation steps of step S2 include: Identify dynamic state parameters of all supply chain nodes, including node throughput, waiting time, inventory fullness, and logistics in and out rates; Calculating the intra-cycle variation of the node throughput and performing potential correlation mining on the waiting time to obtain a time series throughput-waiting time correlation feature; Identify the correlation fluctuations between inventory fullness and logistics inflow and outflow rates, and obtain the correlation characteristics between inventory fullness and logistics inflow and outflow rates; The multi-node logistics state transition evolution is performed based on the fullness and in / out rate correlation characteristics and the time series throughput-waiting time correlation characteristics to generate the logistics state transition characteristics of each node.

[0031] In this embodiment, before implementing intelligent supply chain scheduling optimization, the dynamic state parameters of all key nodes in the supply chain network (such as warehouses, transshipment centers, stores, and distribution points) must be fully identified during operation. This step aims to construct a node state profile based on real-time operational data, providing foundational data for subsequent state analysis and evolution modeling. Node dynamic state parameters primarily encompass the following four aspects: node throughput (the number of orders processed per unit time); wait time (the average time an order or goods remain at the node); inventory fullness (the ratio of the current inventory level to maximum capacity); and logistics inflow and outflow rate (the frequency of material inflow and outflow per unit time). Data sources include enterprise ERP, WMS (Warehouse Management System), TMS (Transportation Management System), and IoT sensor acquisition platforms. A real-time data stream processing framework is built using Kafka and Flink to ensure that status updates from each node are synchronized every five minutes. During the identification process, a multi-source data fusion mechanism is used to standardize the format of status data collected from different systems (using a unified JSON structure). A time alignment module is used to ensure that all parameters have synchronized timestamps. A boundary determination mechanism is also introduced to eliminate outliers. For example, if a node's inventory fullness is continuously recorded as 0 or 1 for more than 12 hours, data validation rules are triggered and retroactive corrections are performed. In a regional experiment in Jiangsu Province, 286 core nodes were monitored, with an average of 18,000 data points sampled daily per node. A complete 24-hour operational status profile encompassing four dynamic parameters was constructed using a state recognition module. To gain a deeper understanding of the interaction between node throughput characteristics and waiting behavior from a temporal perspective, it is necessary to analyze the cyclical variation of node throughput and explore its potential correlation with waiting time, thereby revealing the causes of bottlenecks and systemic delays. First, to analyze the cyclical variation of node throughput, a seasonal decomposition method (such as STL decomposition) was used to decompose the daily throughput series into a trend term, a cyclical term, and a residual term. 30 consecutive days of throughput records for each node were analyzed at a daily granularity, and the average processing capacity and volatility during key periods (e.g., the morning peak period of 7:00–10:00 AM and the evening peak period of 4:00–19:00 PM) were calculated. In the experimental setup, a 24-hour cycle decomposition window and a 1-hour sliding window were used to identify high-load segments and potential congestion periods at each node within a 24-hour period. Secondly, the potential correlation between wait time and throughput was assessed using a combination of Granger causality testing and Pearson correlation analysis to assess the temporal coupling between the two. Using an hourly time step, a causal testing model with lags of 1 to 5 hours was constructed. A significant Granger causal relationship (p-value < 0.05) was found on multiple nodes regarding throughput changes and wait time, with correlation coefficients reaching as high as 0.79, particularly pronounced at distribution centers and edge warehouses.The cyclical fluctuation characteristics and waiting correlation characteristics of each node are combined into a time-series correlation vector (including dimensions such as average peak lag time, cyclical volatility, and correlation slope). This serves as a key structural feature describing the node's logistics load capacity and buffering capacity, and is used to predict its state evolution under complex perturbations. Inventory management efficiency and logistics flow rate are closely coupled. Especially when supply chain nodes operate under pressure, the coordination between inventory fullness and logistics inflow and outflow rhythms directly determines the system's scheduling flexibility and stability. This step aims to identify potential bottleneck nodes and inventory stagnation trends by modeling the dynamic fluctuation relationship between inventory fullness and logistics inflow and outflow rates. The specific method is a cross-correlation fluctuation identification algorithm based on bivariate time series analysis. First, the inventory fullness series I(t) and the logistics inflow and outflow rate series F(t) are constructed, and their time-varying cross-correlation coefficient curve R(τ) is calculated, where 𝜏 is the lag time. In the experiment, a lag time window of ±12 hours and a resolution of 15 minutes were used. A sliding window was used to calculate the maximum correlation value across time lags and the corresponding time point, thereby identifying the time delay characteristics of inventory responses to logistics changes. To further identify fluctuation coordination patterns, a dynamic time warping (DTW) algorithm was used to dynamically align the two series to determine whether there were any abnormal offsets or lags between them. In field measurements at multiple nodes, if the inflow and outflow rates were consistent with the inventory fluctuation trend and showed no significant lag, it indicated that the node had good inventory adjustment capabilities. Conversely, if inventory fullness was chronically high while logistics inflow and outflow rates were sluggish, this was labeled as an "inventory congestion" state. Inventory health thresholds were set between 0.4 and 0.8, and if they exceeded the upper and lower bounds for more than six consecutive hours, it was recorded as a fluctuation anomaly. Each node was ultimately assigned a set of fluctuation-related characteristic parameters, including the optimal lag time, maximum correlation coefficient, fluctuation consistency coefficient, and inventory adjustment sensitivity, which provided input for modeling node state transition trends. In a network application in a city in East China, this method identified 21 high-risk inventory-flow imbalance points, issuing replenishment warnings an average of 3.6 hours ahead of actual stockouts. Combining the characteristic results output by the first two analysis modules, the final step is to model the dynamic evolution of the logistics state of each node in the supply chain over successive moments. This involves constructing logistics state transition features, which provide a decision-making basis for the scheduling system to predict future state change trends.

[0032] State transition modeling utilizes a fusion of Markov chains and deep time series networks. First, the current node state is classified into multiple state labels (such as "high throughput - low wait" and "low inventory - low flow rate") based on combined features. A state transition probability matrix is ​​constructed based on historical data to identify common evolution paths of node states under different disturbances. Second, sequence modeling methods such as LSTM or Transformer are introduced to perform deep learning training on the nonlinear and long-term dependencies in the state transition process, ultimately outputting state transition predictions. During the model training phase, data from the past 90 days is sampled hourly, resulting in a training sample size of 450,000 node-moment state instances. The model outputs the state transition probability distribution for each node within the next 1, 3, and 6 hours, and generates transition feature vectors, including state stability (a stability coefficient < 0.4 is considered volatile), mutation probability, and state hysteresis. These features not only support the prediction of the future state of a single node but can also be used to build collaborative scheduling models between nodes. For example, when the probability of state mutation at an upstream node is high, the system can pre-allocate resources to the buffer zone of downstream nodes to avoid chain interruptions or warehouse overflows. In the deployed application, the accuracy of state transition prediction reached 88.2%, and the proportion of early prediction of scheduling anomalies increased by 31.5%, laying a solid data foundation for intelligent scheduling strategies in complex supply chain environments.

[0033] In this embodiment, reference Figure 4 The above is a flowchart of the detailed implementation steps of step S3. In this embodiment, the detailed implementation steps of step S3 include: Extract timestamps of different supply chain nodes based on filtering optimization of heterogeneous data streams; Calculating the timestamp deviation to obtain time deviation characteristics of different nodes; identifying information delays, logistics delays, and decision delays at multiple nodes based on the dynamic state parameters; Performing a time series non-aligned delay evaluation on the information delay, logistics delay, and decision delay based on the time deviation feature, thereby obtaining non-aligned delay evaluation information between each node; Based on the non-aligned delay evaluation information, the upstream and downstream node delays of the logistics state transfer characteristics of each node are calculated, and global distribution mapping is performed to construct an upstream and downstream logistics state delay feature map.

[0034] In this embodiment, before implementing delay assessment and scheduling response, it is necessary to accurately extract event timestamps for each node at a specific logistics link from the filtered and optimized heterogeneous supply chain data. Timestamps are fundamental to describing the temporal characteristics of supply chain behavior. They include three types: event occurrence time, data generation time, and system receipt time. Their accuracy directly impacts the accuracy of delay assessment. At this stage, leveraging high-quality, heterogeneous data streams (including logistics trajectories, order events, and environmental data) that have undergone adaptive filtering, the following key timestamps are extracted for each node (such as warehouses, distribution centers, and stores): ① Logistics arrival timestamp (triggered by GPS trajectory points matching node boundaries); ② Order event timestamps (such as order acceptance, shipment, and receipt times); and ③ System receipt timestamp (i.e., the time data is written to the scheduling platform). To ensure uniformity, all timestamps are formatted in GMT+8 with millisecond accuracy and converted to UNIX time format for ease of subsequent numerical analysis. To improve the accuracy of timestamp extraction, an event recognition model (based on rules and neural networks) is introduced to calibrate behavioral sequences. For example, if a vehicle stops at a node for more than 15 minutes, combined with the order's "shipped" status, the trajectory point and the order event are timestamped and confirmed. In actual experiments, three months of data were collected from 1,500 nodes in the Shanghai, Jiangsu, and Zhejiang regions, with an average daily record of over 1.8 million items. After verification, the accuracy of timestamp extraction increased to 97.8%. This step provides solid time series data support for delay calculation and misalignment assessment. After obtaining key timestamps for each supply chain node, the next step is to calculate the deviations between these timestamps to construct a "time deviation signature" to characterize asynchrony and time transmission efficiency issues within the system. The core of this step is to discover inconsistencies in temporal behavior between nodes, which is a prerequisite for identifying the root causes of delays. The calculation of time deviation is based on a "theoretical event chain," which predefines the expected temporal progression and intervals between nodes under a standard logistics process. For example, in a standard distribution chain, there should be a fixed expected time difference ΔT0 between shipment from node A (warehouse) and receipt at node B (distribution center). In real data, if A's delivery time is T1 and B's receipt time is T2, then the time offset ΔT = T2 − T1 − ΔT0. This is used to construct the time offset matrix between all node pairs.

[0035] Furthermore, all deviations were normalized (using the Z-score method) and a multidimensional time deviation feature vector was constructed, including mean deviation, standard deviation, maximum deviation, and deviation direction (lead / lag). The analysis also considered the dynamic changes in deviation during different time periods (such as peak hours and nighttime) to extract time-specific deviation characteristics. In a real-world test, using data from a Shanghai logistics company as an example, time deviation modeling was performed for 20 key nodes. It was found that over 95% of deviations were concentrated within a ±3-hour range, but some nodes exhibited long-term lags exceeding 6 hours, indicating potential scheduling lags or information desynchronization. The construction of time deviation features lays the data structure foundation for subsequent identification of multiple types of delays. Identifying delay types is a core step in delay assessment. This step aims to identify three key types of delays at each supply chain node in actual operation based on the previously extracted time deviation features and dynamic state parameters: information delay, logistics delay, and decision delay, thereby comprehensively characterizing node response lags. Information delay refers to the time difference between the occurrence of an event and the receipt of that information by the system. Information delay is calculated by subtracting the event occurrence time (T_event) from the timestamp reception time (T_recv). If the difference exceeds a system-defined threshold (e.g., 10 minutes), information delay is considered present. Information delay primarily reflects issues with system access efficiency and data chain integrity. Secondly, logistics delay refers to the difference between the actual completion time and the planned time of a node's physical operation. For example, if the planned delivery time is 10:00 and the actual delivery time is 11:30, the logistics delay is 90 minutes. This delay is derived by quantifying node throughput, waiting time, and inventory pressure, and then analyzed comparatively with order scheduling plans. Third, decision delay reflects the response lag between the issuance and execution of scheduling instructions. This is typically achieved through integration with log data from the TMS / scheduling system. For example, if the system generates the optimal route at T_opt and the actual route adjustment occurs at T_act, the delay T_decision = T_act − T_opt. This type of delay requires cross-verification with the intelligent scheduling system and logistics behavior events.

[0036] In practice, multidimensional delay identification was conducted on 110 logistics nodes in a city in Zhejiang Province. The median information delay was 8.2 minutes, logistics delay was 47 minutes, and the average decision delay was 35 minutes. The classified delay data served as input for subsequent temporal misalignment modeling. After identifying the three types of delay, a comprehensive temporal misalignment assessment was conducted. This analysis analyzed how delay differences between nodes contribute to temporal asynchrony in the logistics chain. The goal of this step was to construct a multi-node misalignment delay assessment framework to identify potential coordination barriers and system bottlenecks. Dynamic Time Warping (DTW) and a multi-layer temporal alignment analysis model were used to dynamically assess inter-node delay behavior. Using the three-dimensional delay sequences of information flow, logistics flow, and decision flow as input, the misalignment metrics for each pair of upstream and downstream nodes within a specific time period were calculated, including delay offset, synchronization lag, and local variation amplitude. A delay weighting factor model is also introduced to comprehensively assess the three types of delays. For example, weights are assigned based on their impact on delivery success rate (for example, 0.3 for information delay, 0.5 for logistics delay, and 0.2 for decision delay). This constructs a comprehensive delay score and generates a non-aligned delay score matrix between nodes. Experiments demonstrate the non-aligned evaluation model for 184 logistics nodes in Jiangsu Province. Results show an average node-pair misalignment of 0.42, with a maximum exceeding 0.8, primarily concentrated between third-tier distribution stations and edge warehouses. This reveals key delay bottlenecks and asymmetric response paths within the supply chain system. The final output of this stage is "non-aligned delay evaluation information," which includes timing offset curves between node pairs, delay distribution maps, and delay correlation networks. This information is then used to model delay propagation in upstream and downstream logistics. After identifying and quantifying non-aligned delays between nodes, the final step is to combine this delay information with previously generated logistics state transition features to further model delay propagation paths between upstream and downstream nodes and construct a global delay feature map. This map characterizes the spatial distribution and impact chain of delay behavior across the entire supply chain network. The specific method is to construct a delay propagation graph model based on non-aligned evaluation information: each node is represented as a point in the graph, the delay propagation direction is represented as an edge (directed edge), and the edge weight represents the comprehensive delay score between node pairs. Based on this graph, a delay-aware state transition network is constructed by combining the logistics state transition characteristics of each node (such as state transition probability and fluctuation intensity). Subsequently, graph embedding techniques (such as GraphSAGE or Node2Vec) are used to vectorize the node state and delay propagation characteristics. Based on this, a delay heat map and delay diffusion path map are generated to visualize and track the delay distribution in the network. By analyzing the node centrality and propagation path length in the graph, the core nodes of delay diffusion and key link break points can be identified.In a real-world deployment test, the map was applied to a supply chain network in three provinces in East China. The system discovered that the primary path of latency propagation was concentrated along the "central warehouse - city distribution center - store" chain. The central warehouse's response latency exceeded the average by three times, making it a priority node for scheduling optimization. After the map was constructed, the overall system's scheduling response speed increased by 19.7%, effectively alleviating link congestion under high-concurrency orders. Ultimately, this latency signature map became a key input module for the supply chain's intelligent scheduling engine, enabling precise optimization driven by latency awareness.

[0037] In this embodiment, step S4 includes the following steps: Define cost-priority scheduling decision-makers, service-priority scheduling decision-makers, robustness enhancement decision-makers, and speed emergency response decision-makers to build a scheduling decision strategy pool; Based on the upstream and downstream logistics status delay characteristic map and supply chain disturbance response field, real-time supply chain information analysis is carried out to build a simulation game model; The scheduling decision strategy pool is input into the simulation game model for multi-strategy agent game training, thereby extracting the profit dynamics, adaptive trajectories and strategic synergy factors of different strategies; Based on the profit dynamics, adaptive trajectory and strategic synergy factors, a comprehensive strategy evolution analysis and fitting effect are performed to obtain the scheduling game effect evolution curve; According to the evolution curve of the scheduling game effect, the simulation game model is optimized through multiple rounds of game, and a supply chain multi-round game optimization model is constructed.

[0038] In this embodiment, the first step in building a scheduling decision-making system is to design a diverse set of strategic entities, called "scheduling agents." Each agent simulates different scheduling objective preferences, representing the supply chain's scheduling behavior tendencies under different operational scenarios. To achieve dynamic, scenario-driven intelligent optimization, four typical agents are defined and instantiated in this phase: a cost-first agent, a service-first agent, a robustness agent, and a speed-response agent. Together, they form a "scheduling decision-making pool." The cost-first agent prioritizes minimizing transportation costs, with its decision logic favoring low-cost routes, consolidating transportation nodes, and reducing empty trips. In scheduling simulations, this agent prioritizes large, concentrated orders and tends to avoid highly volatile logistics routes. The service-first agent focuses on maximizing customer satisfaction, with its objective function emphasizing delivery timeliness, integrity, and service-level assurance. In simulations, this agent adopts a weighted response to urgent orders and volatile nodes, proactively allocating resources with high service levels. The robustness-enhancing dispatcher is designed with the system's resilience to risk as its core design focus. Its policy logic emphasizes path redundancy, multi-point deployment, and dynamic replanning, making it particularly suitable for scenarios with frequent disturbances (such as severe weather and node congestion). The speed-sensitive emergency response dispatcher focuses on extreme response speed. In the event of an emergency (such as a surge in orders or a sudden lockdown), it prioritizes the mobilization of resources available in the shortest possible time to implement rapid transit. Each decision-making body operates as an independent agent, equipped with an autonomous policy evaluation function and feedback mechanism. Parameters are customized through empirical simulation and expert research (for example, the service priority body sets an objective constraint of on-time delivery of at least 95% of orders). A policy pool design provides multi-strategy input for subsequent game learning, laying the foundation for dynamic intelligent scheduling. To conduct multi-strategy intelligent scheduling simulation game training, a high-fidelity simulation-gaming model with spatiotemporal dynamics, disturbance propagation, and policy feedback mechanisms must be constructed. This model integrates the delay characteristic maps of upstream and downstream logistics states with the supply chain disturbance response field to simulate the behavioral responses and model strategic confrontation across the entire supply chain network.

[0039] The model construction consists of three parts: First, the delay feature map input module inputs the logistics state evolution trend, non-aligned delay index, and delay propagation path of each node into the simulation model, generating a multi-node, time-sequential logistics network environment. Second, the disturbance response field module simulates the spatial transmission and intensity changes of sudden disturbances in the system. Based on the nonlinear response tensor of the disturbance point, it generates a disturbance factor flow field, dynamically influencing network behavior. Third, the game agent embedding mechanism embeds each scheduling agent in the strategy pool as an independent agent in the simulation environment, participating in the scheduling game through an "observation-decision-feedback" cycle. Regarding experimental parameter setting, a virtual network based on 100 logistics nodes in Jiangsu Province was constructed, covering five order demand scenarios (regular, e-commerce peak, sudden transfer, inventory imbalance, and weather disturbance). The disturbance intensity was set to 10 per 24 hours, with an average of 5 response nodes, a simulation step length of 15 minutes, and a period of 7 days. Through this model, the system can evaluate the decision-making behavior and payoff changes of different strategy agents under specific disturbance and delay structures in real time. The simulation results will serve as a behavioral data source and strategy evaluation basis for subsequent multi-strategy learning. After completing the game simulation model, the strategy training phase begins. This step centers on multi-agent simulation game training, which aims to extract the reward curve, adaptation path, and synergy index of each agent through repeated game play in a multi-perturbation, dynamic delay environment. Game training utilizes the Multi-Agent Reinforcement Learning (MARL) framework, with each agent independently updating its strategy. Its objective function is defined based on the agent's characteristics (e.g., the cost agent minimizes total transportation cost). In each simulation round, the system records each agent's reward changes, resource usage, scheduling accuracy, response time, and other metrics, continuously constructing a reward curve. Profit dynamics: Record the changes in the unit time profit of the strategy body under different disturbance scenarios (such as cost savings per hour or shortened delivery time), and analyze its high-efficiency area and failure boundary; Adaptive trajectory: Record the strategy change frequency, behavioral path offset and learning speed of the strategy body in a continuous disturbance environment, and evaluate its response flexibility; Strategy synergy factor: Calculate the synergy benefits between different strategy bodies under joint operation. If the combination of two strategies improves global efficiency, the synergy factor is greater than 1, otherwise it is less than 1.

[0040] The experiment employed a five-round training mechanism with increasing perturbation intensity, simulating 1,000 game rounds per round, generating over 30,000 strategy behavior trajectories. Results showed that the service-prioritized strategy achieved a 34% increase in revenue dynamics under sudden customer orders, but exhibited uneven performance under high logistics delays. In contrast, the robust strategy significantly improved its synergy factor (reaching 1.26) under large-scale perturbations, demonstrating the system's ability to maintain stable control. After completing multi-strategy game training, in-depth analysis and fitting of the strategy's behavior and performance under different perturbation scenarios is required to reveal its operational advantages, adaptability, and evolutionary trends. Ultimately, a strategy evolution curve (Strategy Evolution Curve) is generated to provide data support for strategy combination and selection. First, through multi-dimensional indicator normalization, the revenue dynamics, adaptability trajectories, and synergy factors are converted into standardized input variables. Next, nonlinear regression fitting algorithms (such as Bayesian Gaussian process regression and B-spline fitting) are used to construct response curves of strategy performance as a function of perturbation intensity, delay level, and order type, analyzing the evolutionary trends of different strategies in a multidimensional parameter space. The evolution curve focuses on the performance of strategies along three dimensions: ① disturbance intensity (from low to high disturbance); ② delay structure (from synchronization to severe misalignment); and ③ demand complexity (from stable orders to sudden, high-priority orders). For example, in high-disturbance, high-misalignment scenarios, the evolution curve of the robust strategy shows an upward trend, while the service-first strategy's profit curve significantly decreases under these conditions, reflecting its applicability limit. Finally, a three-dimensional evolutionary graph and strategy effectiveness heat map are generated for each strategy. In test data, the strategy game effect fitting accuracy reached 92.4%, providing precise data for the next stage of strategy selection and optimization. Multiple rounds of evolutionary optimization were performed on the simulation game model to construct a supply chain multi-round evolutionary game model with continuous learning and self-adjustment capabilities. This model not only enables static strategy evaluation but also supports long-term evolutionary learning and adaptive scheduling in complex scenarios. A strategy optimization mechanism was constructed based on the strategy evolution curves obtained in the previous step. Before each round of game play begins, a strategy combination (e.g., coordinated scheduling of robustness agents and speed agents) is dynamically selected and initial weights are set based on the current disturbance field, latency profile, and historical execution trajectories. After each round, the strategy combination weights are adjusted and the structure is iterated based on the actual returns and synergistic effects of each strategy, implementing a closed-loop "game-feedback-regame" mechanism. Evolutionary Game Theory models are introduced to construct equations for adaptive competition and cooperation between strategy agents. Strategy agents continuously adjust their willingness to participate and resource competition weights based on the payoff function, thereby simulating the collaborative game behavior of multiple strategy agents in a real supply chain.During the simulation process, the model continuously utilized disturbance response data for feedback training, adjusting the strategy combination and scheduling path, ultimately converging to the optimal strategy map. In actual testing, after 10 rounds of game play, the model reduced the average delivery time of the entire supply chain by 11.5% and decreased the abnormal order rate by 17.3%, fully verifying the effectiveness of the multi-round game optimization mechanism and the scheduling benefits.

[0041] In this embodiment, step S5 includes the following steps: Identify real-time contract terms information; identify user modifications to the real-time contract terms information, and perform deep semantic analysis to identify semantic features of contract modifications; Mining intention changes based on the semantic features of contract modifications, analyzing supply chain behavior changes, and marking behavior change points; Quantify the upstream and downstream related behaviors of the behavior transition points to obtain the fluctuation value of the upstream and downstream node behavior impact; The optimal buffer time window is calculated based on the fluctuation value of the upstream and downstream node behaviors to obtain the optimal buffer time window.

[0042] In this embodiment, in the intelligent supply chain scheduling system, contract terms not only represent transactions and performance obligations between nodes but also directly determine key operational parameters such as logistics time, allocation priority, and resource access rights. Therefore, after obtaining authorization from the platform and the enterprise, the first step is to identify contract terms information between nodes at all levels in real time, serving as a crucial input for the constraints of the dynamic scheduling game model. Contract information originates from internal enterprise ERP systems, supplier platforms, blockchain smart contract platforms, and other sources, and is diverse and heterogeneous. This stage utilizes a combination of structured text parsing and rule-driven extraction. First, the contract texts from various sources are formatted in a standardized manner (e.g., unified into a JSON structure). Entity recognition technology is then used to extract the core elements of the terms, including delivery SLAs, service levels, penalties, payment terms, and resource locking windows. To enhance recognition accuracy, a BERT-BiLSTM-CRF joint entity annotation model was trained on a dataset of approximately 34,000 self-constructed contract terms in Chinese and English, achieving an F1 score of 94.2% for identifying term elements. The system updates contract information every 15 minutes to ensure the latest performance constraints are used during the current scheduling process. This phase outputs a structured contract terms data table, providing foundational information for subsequent semantic change identification and intent mining. Contract modifications often occur in response to order fluctuations, supply changes, or service anomalies, reflecting potential shifts in business strategy or performance intent. Therefore, accurately identifying and understanding user modifications to contract terms is a prerequisite for dynamically updating scheduling strategies. The first step in this phase is contract modification detection, which involves comparing the current contract version with previous versions to identify field-level changes, such as "delivery deadline changed from 24 hours to 12 hours" or "a penalty added to a previously non-breach of contract clause." A version comparison algorithm, combined with Levenshtein distance and field hash comparison, precisely locates the modified fields and their locations. After identifying the modified fields, the deep semantic parsing phase begins. A pretrained language model (such as the RoBERTa-Contract version) is used to understand the context and identify the semantic intent of the modifications. This model, fine-tuned based on contract language data, is capable of identifying semantic features of modifications such as "expedited delivery," "resource allocation priority," and "breach of contract risk aversion." The output is a set of "contract modification semantic labels," such as [improved timeliness], [increased responsibilities], and [resource reductions]. In actual testing, across 5,000 enterprise contracts, the model achieved 92.6% accuracy in identifying contract modification behaviors and 88.7% accuracy in semantic parsing labels. The results of this phase are not only used for inferring intent change but also directly impact the parameter configuration of scheduling models for adjusting resource priorities and risk compensation mechanisms.This phase aims to identify potential behavioral changes in customers or partners through in-depth analysis of contract modification semantics, thereby assessing their impact on supply chain operational strategies and promptly identifying and labeling behavioral transition points. First, an "intent classification model" based on contract semantic labels is introduced to map contract modification semantics to behavioral intent. Based on a training corpus of intent transitions (covering eight categories, including changes in order timeliness, risk control, cost reduction, and resource reallocation), the model employs a TextCNN + Transformer architecture for classification, outputting intent labels such as [allocation priority shift], [service level rebalancing], and [performance assurance improvement]. Subsequently, a comparative analysis of intent transitions with historical supply chain scheduling behavior is conducted to identify indicators such as actual scheduling deviation paths, changes in resource utilization, and changes in node response strategies. These indicators are then used to label behavioral transition points. Transition point identification relies on the superposition of three behavioral signals: Sudden changes in time cues (e.g., shortening or speeding up the average scheduling interval); Sudden path deviation (e.g., frequent changes in allocation paths); Resource weight changes (such as priority node changes).

[0043] For example, in a typical scenario, a retailer modified its "city distribution center → store" path contract to "service level priority + resource exclusivity." The system identified the corresponding scheduling strategy shift from "on-demand allocation" to "fixed window advance release," marking it as a behavioral transition point: "Path priority shift + enhanced resource lock." In a real-world experiment involving 8,000 contract changes, 5,720 valid behavioral transition points were identified, with an average lead time of 5.3 hours, providing ample buffer for system scheduling responses. Behavioral transitions often affect not only the contract initiator but also ripple effects across upstream and downstream supply chain nodes. Therefore, it is necessary to model the upstream and downstream propagation of identified behavioral transition points and quantify their impact on the behavioral patterns of related nodes, outputting a "behavioral impact fluctuation value." The system first constructs a node behavior state map, representing each node's behavior state using dimensions such as "dispatching frequency, path selection rate, resource response time, and resource utilization ratio." Then, with the transition point as the central node, an influence propagation model (such as a Markov jump model based on diffusion weights) is used to simulate the propagation of behavioral changes across upstream and downstream nodes.

[0044] The system introduces a "behavior shift vector," defined as the difference between various behavioral indicators before and after a node's transition. This vector is calculated by combining the Euclidean distance of these vectors with an adjustment factor (such as node sensitivity weight) to produce a volatility index (ΔB). For example, if node A's scheduling frequency increases by 28% and its resource response latency decreases by 21% after a contract change, its ΔB is 0.34, indicating moderate behavioral volatility. This volatility index is further used to identify key volatility channels and highly sensitive nodes. In actual operational experiments, the model was deployed across a 57-node supply chain network in central China and successfully identified 12 nodes highly sensitive to behavioral volatility, providing early warning of potential resource scheduling congestion. Finally, based on the behavioral impact volatility of upstream and downstream nodes, the system dynamically calculates the optimal buffer window, providing a basis for accurate time redundancy design and path planning in subsequent scheduling models. The buffer window calculation model considers three input factors: ① node behavior volatility ΔB, ② historical delay distribution, and ③ resource scheduling bandwidth. The system uses a weighted nonlinear regression model to construct a "behavior fluctuation-time window requirement" function, outputting the minimum buffer redundancy required for each path or node. In the mathematical expression, a correction factor α based on the behavior index is introduced, and the final time window calculation formula is: Where α is the behavioral fluctuation sensitivity coefficient (depending on the size of ΔB), μ_"delay" and σ_"delay" are the average delay and standard deviation of delay, respectively. In real-world scheduling scenarios, the model's dynamically calculated buffer window better reflects the frequency of disturbances and actual node behavior. For example, in multiple sudden contract modification scenarios, the traditional statically reserved 2-hour buffer window was no longer sufficient to guarantee timeliness. However, the model dynamically calculated a buffer window between 2.7 and 3.5 hours based on ΔB, enabling early allocation and resource assurance, improving on-time delivery rates by 12.6%.

[0045] In this embodiment, step S6 includes the following steps: Based on the optimal buffer time window, the contract modification scheduling simulation of the supply chain multi-round game optimization model is carried out to generate scheduling simulation data; Perform multiple rounds of minimum cost decision-making on scheduling simulation data to generate minimum cost decision samples; Perform intelligent scheduling optimization based on minimum cost decision samples and build an order mutation decision optimization engine.

[0046] In this embodiment, the supply chain scheduling strategy is incorporated into the "contract mechanism" framework. Based on the concept of the optimal buffer time window, the inter-node scheduling rules are readjusted to create a simulation environment that is closer to actual operations, thereby generating a large number of data samples that are consistent with business practices for subsequent optimization training. The specific process is based on a multi-round game optimization model, incorporating dynamic contract modification constraints to construct a scheduling simulation environment. The "optimal buffer time window" refers to a safe time interval (buffer window) reserved in the order contract to prevent delivery delays or failures. The determination of this time window requires a minimum risk calculation based on historical delay distribution, node processing capacity, and fluctuation intensity. It is usually set between 1.5 and 2 times the average node delay value. For example, for a regional distribution center, if the average delay from order receipt to shipment over the past three months is 1.6 hours, a buffer window setting of 2.5 hours is considered a high level of assurance. During the simulation, the system simulated adjustments to the delivery time, resource scheduling, and compensation mechanisms involved in the original supply chain contract. These adjustments included relaxing penalty clauses for late delivery, fine-tuning shipping windows, and implementing a two-choice mechanism for allocation routes. These rule changes were then loaded as perturbation inputs into a multi-round game optimization model. Subsequently, a full-path scheduling simulation was conducted based on the modified strategy scenario, encompassing order intake, resource allocation, route selection, contract response, and system feedback. In a live deployment experiment, a scheduling contract modification simulation model was constructed for 50 logistics nodes in South China, with three buffer window scenarios (1.2 hours, 2.0 hours, and 3.5 hours) set, and a total of 12,000 orders were simulated. The final output data included each order's actual scheduling path, cost changes, completion status, and scheduling response strategy, laying the foundation for subsequent training to generate minimum-cost samples. After obtaining scheduling simulation data from contracts modified based on the buffer mechanism, the next step is to perform multi-round optimization on this data to extract the decision paths that achieve the best cost performance across multiple scheduling iterations. This process then constructs a high-quality training set for the intelligent optimization model to learn the boundaries and evolutionary trajectory of scheduling decisions. This step employs a method based on multi-round path search and cost inversion. First, the strategy sequences generated in each simulated scheduling (such as resource allocation, time window selection, and path switching) are mapped to construct a multi-round scheduling state tree centered on the order. Next, a comprehensive cost evaluation function is defined, taking into account multiple factors such as transportation costs, timeout penalties, resource vacancies, and delay compensation, and scoring each scheduling path. To find the optimal decision path, the system uses a heuristic search algorithm (such as A* search combined with dynamic regularization) to discover the node path with the lowest cost in the state tree. Furthermore, to avoid local optimality, a random perturbation mechanism is introduced to simulate sudden order changes (such as changes in the priority of some orders, sudden order insertions, and sudden stock-outs). Path corrections are then re-executed to ensure that the output minimum-cost path is more robust.

[0047] In the experimental design, multiple rounds of optimal path searches were conducted on the 12,000 order scheduling data output by the simulation. An average of 7.4 different scheduling schemes were evaluated for each order, and finally 9,300 paths that maintained the lowest total cost under all disturbance constraints were extracted as "minimum cost decision samples." The sample data includes strategy combinations, node selection, adjustment point locations, cost structures, and scheduling result labels, which are used for supervised learning by the next stage model. The last step is to use the minimum cost decision samples generated above as a training basis to build an intelligent scheduling optimization engine for sudden order scenarios, achieving rapid response and optimal scheduling generation in highly complex dynamic environments. This engine is the core execution module of the entire big data-driven supply chain intelligent scheduling system, with three major features: real-time calculation, strategy evolution, and rapid decision-making.

[0048] The engine building process consists of three main steps: During the training phase, a multi-layer neural network model (such as a Transformer-based temporal policy network) is used to train the model, using minimum-cost decision samples as the training set. The core learning task is to predict the optimal dispatch path given order status, disturbance context, and resource constraints. The model inputs include order attributes (type, deadline, weight), node status (inventory, transportation capacity), and disturbance characteristics (weather, delays, congestion). The output is a dispatch strategy combination and the expected path.

[0049] Mutation Identification Module: This module integrates an order mutation detection module to identify abnormal order behavior (such as priority escalation, large-scale order insertion, and target node blocking) in real time and trigger a replanning request for the current dispatch path. This module identifies mutations through continuous feature flow analysis and pattern matching, with a response time of less than 3 seconds.

[0050] Scheduling Execution Engine: During runtime, the model combines the current node status, logistics network status, and disturbance prediction results, and calls the inference engine to generate a scheduling plan in real time. The engine can achieve multi-strategy coordinated response by linking with the game optimization model. For example, in the event of a sudden typhoon, the system can call on the "robust + service priority" joint strategy to quickly identify available paths and implement the shortest guaranteed delivery plan. In the scenario experiment, the intelligent scheduling optimization engine was applied to the key commercial flow network in the Beijing-Tianjin-Hebei region. The average order scheduling generation time was controlled at 4.8 seconds, and the order mutation response time was shortened by 37.4%. In the scenario with a high burst order rate (30%), the overall system cost was reduced by 15.7% compared to manual rule scheduling, demonstrating strong intelligent adaptability and optimization effects.

[0051] In this embodiment, a supply chain intelligent scheduling decision optimization system based on big data is provided, which is used to execute the supply chain intelligent scheduling decision optimization method based on big data, including: The disturbance response mining module is used to collect heterogeneous data streams from the supply chain, extract abnormal disturbance events, conduct nonlinear disturbance response mining, and construct a supply chain disturbance response field; The state transition evolution module is used to identify the dynamic state parameters of all nodes in the supply chain, analyze the correlation characteristics between parameters, and perform multi-node logistics state transition evolution to generate logistics state transition characteristics for each node; The delay mapping module is used to perform multi-node delay evaluation based on heterogeneous data streams in the supply chain, and perform global distribution mapping to build a delay feature map of upstream and downstream logistics status; The game optimization module is used to conduct real-time supply chain information analysis and multi-round game optimization based on the upstream and downstream logistics status delay characteristic map and supply chain disturbance response field to build a supply chain multi-round game optimization model; The buffer time window calculation module is used to identify real-time contract terms information, conduct contract intention change mining, and then calculate the optimal buffer time window to obtain the optimal buffer time window; The multi-round decision module is used to simulate contract modification scheduling in the supply chain multi-round game optimization model based on the optimal buffer time window, perform multi-round decision-making with minimum cost, and build an order mutation decision optimization engine.

[0052] The present invention is therefore intended to be illustrative and non-restrictive in all respects, with the scope of the invention being defined by the appended claims rather than the foregoing description, and all changes that come within the meaning and range of equivalents of the application documents are intended to be embraced therein.

[0053] The foregoing description is intended only to provide specific embodiments of the present invention, which are intended to enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is to be construed in the widest manner consistent with the principles and novel features disclosed herein.

Claims

1. A supply chain intelligent scheduling decision optimization method based on big data, characterized by: The following steps are involved: Step S1: Collect heterogeneous data streams from the supply chain, extract abnormal disturbance events, and conduct nonlinear disturbance response mining to construct a supply chain disturbance response field; Step S2: Identify the dynamic state parameters of all nodes in the supply chain, analyze the correlation characteristics between the parameters, and perform multi-node logistics state transition evolution to generate the logistics state transition characteristics of each node; Step S3: Perform multi-node delay evaluation based on heterogeneous data streams in the supply chain, perform global distribution mapping, and construct a delay feature map of upstream and downstream logistics status; Step S4: Based on the upstream and downstream logistics state delay characteristic map and the supply chain disturbance response field, the logistics state transition characteristics are analyzed in real time and multi-round game optimization is performed to build a supply chain multi-round game optimization model; Step S5: Identify real-time contract terms information, conduct contract intention change mining, and then calculate the optimal buffer time window to obtain the optimal buffer time window; Step S6: Based on the optimal buffer time window, the supply chain multi-round game optimization model is simulated for contract modification scheduling, and multi-round decision-making with minimum cost is performed to build an order mutation decision optimization engine.

2. The supply chain intelligent scheduling decision optimization method based on big data drive according to claim 1 is characterized in that: The specific steps of step S1 are: Collect heterogeneous supply chain data streams, including logistics trajectory data, order event streams, and real-time environmental weather information; Identify abnormal noise in heterogeneous data streams of the supply chain and perform adaptive filtering and noise reduction to obtain filtered and optimized heterogeneous data streams; Based on filtering and optimizing heterogeneous data streams, we can identify abnormal logistics transition trajectories, order change nodes, and meteorological mutation points to obtain multi-dimensional disturbance mutation points. Nonlinear disturbance response mining is performed on multi-dimensional disturbance mutation points to construct a supply chain disturbance response field.

3. The supply chain intelligent scheduling decision optimization method based on big data drive according to claim 2 is characterized in that: The specific steps of mining the nonlinear disturbance response of multi-dimensional disturbance mutation points and constructing the supply chain disturbance response field are as follows: Calculating the disturbance amplitude and duration of the multi-dimensional disturbance mutation point; Conduct disturbance propagation trend analysis on multi-dimensional disturbance mutation points in the supply chain to obtain the propagation trend characteristics of each disturbance mutation point; Quantifying the impact of the disturbance based on the disturbance amplitude, duration, and propagation trend characteristics to generate a fluctuation state of the supply chain disturbance point; Performing local disturbance correlation evolution on the multi-dimensional disturbance mutation points of the supply chain to identify the local disturbance correlation area; Based on the fluctuation state of the supply chain disturbance point, the nonlinear disturbance response of the local associated area of ​​the disturbance is mined to generate the nonlinear disturbance response tensor of multiple nodes in the supply chain; The nonlinear disturbance response tensor is mapped across the entire graph to construct a supply chain disturbance response field.

4. The supply chain intelligent scheduling decision optimization method based on big data drive according to claim 1 is characterized in that: The specific steps of step S2 are: Identify dynamic state parameters of all supply chain nodes, including node throughput, waiting time, inventory fullness, and logistics in and out rates; Calculating the intra-cycle variation of the node throughput and performing potential correlation mining on the waiting time to obtain a time series throughput-waiting time correlation feature; Identify the correlation fluctuations between inventory fullness and logistics inflow and outflow rates, and obtain the correlation characteristics between inventory fullness and logistics inflow and outflow rates; The multi-node logistics state transition evolution is performed based on the fullness and in / out rate correlation characteristics and the time series throughput-waiting time correlation characteristics to generate the logistics state transition characteristics of each node.

5. The supply chain intelligent scheduling decision optimization method based on big data drive according to claim 1 is characterized in that: The specific steps of step S3 are: Extract timestamps of different supply chain nodes based on filtering optimization of heterogeneous data streams; Calculating the timestamp deviation to obtain time deviation characteristics of different nodes; identifying information delays, logistics delays, and decision delays at multiple nodes based on the dynamic state parameters; Performing a time series non-aligned delay evaluation on the information delay, logistics delay, and decision delay based on the time deviation feature, thereby obtaining non-aligned delay evaluation information between each node; Based on the non-aligned delay evaluation information, the upstream and downstream node delays of the logistics state transfer characteristics of each node are calculated, and global distribution mapping is performed to construct an upstream and downstream logistics state delay feature map.

6. The supply chain intelligent scheduling decision optimization method based on big data drive according to claim 1 is characterized in that: The specific steps of step S4 are: Define cost-priority scheduling decision-makers, service-priority scheduling decision-makers, robustness enhancement decision-makers, and speed emergency response decision-makers to build a scheduling decision strategy pool; Based on the upstream and downstream logistics status delay characteristic map and supply chain disturbance response field, real-time supply chain information analysis is carried out to build a simulation game model; The scheduling decision strategy pool is input into the simulation game model for multi-strategy agent game training, thereby extracting the profit dynamics, adaptive trajectories and strategic synergy factors of different strategies; Based on the profit dynamics, adaptive trajectory and strategic synergy factors, a comprehensive strategy evolution analysis and fitting effect are performed to obtain the scheduling game effect evolution curve; According to the evolution curve of the scheduling game effect, the simulation game model is optimized through multiple rounds of game, and a supply chain multi-round game optimization model is constructed.

7. The supply chain intelligent scheduling decision optimization method based on big data drive according to claim 1 is characterized in that: The specific steps of step S5 are: Identify real-time contract terms information; identify user modifications to the real-time contract terms information, and perform deep semantic analysis to identify semantic features of contract modifications; Mining intention changes based on the semantic features of contract modifications, analyzing supply chain behavior changes, and marking behavior change points; Quantify the upstream and downstream related behaviors of the behavior transition points to obtain the fluctuation value of the upstream and downstream node behavior impact; The optimal buffer time window is calculated based on the fluctuation value of the upstream and downstream node behaviors to obtain the optimal buffer time window.

8. The supply chain intelligent scheduling decision optimization method based on big data drive according to claim 1 is characterized in that: The specific steps of step S6 are: Based on the optimal buffer time window, the contract modification scheduling simulation of the supply chain multi-round game optimization model is carried out to generate scheduling simulation data; Perform multiple rounds of minimum cost decision-making on scheduling simulation data to generate minimum cost decision samples; Perform intelligent scheduling optimization based on minimum cost decision samples and build an order mutation decision optimization engine.

9. A supply chain intelligent scheduling decision optimization system driven by big data, characterized by: The method for implementing the big data-driven supply chain intelligent scheduling decision optimization method according to claim 1 comprises: The disturbance response mining module is used to collect heterogeneous data streams from the supply chain, extract abnormal disturbance events, conduct nonlinear disturbance response mining, and construct a supply chain disturbance response field; The state transition evolution module is used to identify the dynamic state parameters of all nodes in the supply chain, analyze the correlation characteristics between parameters, and perform multi-node logistics state transition evolution to generate logistics state transition characteristics for each node; The delay mapping module is used to perform multi-node delay evaluation based on heterogeneous data streams in the supply chain, and perform global distribution mapping to build a delay feature map of upstream and downstream logistics status; The game optimization module is used to conduct real-time supply chain information analysis and multi-round game optimization based on the upstream and downstream logistics status delay characteristic map and supply chain disturbance response field to build a supply chain multi-round game optimization model; The buffer time window calculation module is used to identify real-time contract terms information, conduct contract intention change mining, and then calculate the optimal buffer time window to obtain the optimal buffer time window; The multi-round decision module is used to simulate contract modification scheduling in the supply chain multi-round game optimization model based on the optimal buffer time window, perform multi-round decision-making with minimum cost, and build an order mutation decision optimization engine.

Images