A neural network-based supply chain logistics demand prediction method and system

Abstract

The present application relates to the technical fields of supply chain logistics and big data analysis, and discloses a supply chain logistics demand prediction method and system based on a neural network, which comprises the following steps: obtaining supply chain node data, constructing a weighted directed graph to generate a network topology map; extracting upstream and downstream influence features to obtain a node dynamic interaction vector; calculating the influence weight of interaction frequency on demand fluctuation, marking a key path and determining a traffic pressure distribution value; fusing multi-dimensional data to generate a comprehensive feature description vector; correcting a demand time series to correct prediction bias; extracting seasonal rules and node correlations to generate demand prediction adjustment parameters; and using the adjustment parameters to correct new data, time series extrapolation to obtain high-precision demand prediction results. The present application realizes accurate prediction of the whole process of supply chain logistics demand and effectively solves the problem of low prediction accuracy in the prior art.

Description

Technical Field

[0001] This invention relates to the fields of supply chain logistics and big data analysis technology, and in particular to a supply chain logistics demand forecasting method and system based on neural networks. Background Technology

[0002] Currently, supply chain logistics, as a core hub connecting production and consumption, directly impacts stable market supply and core competitiveness of enterprises through its efficient operation. Accurate forecasting of logistics demand is crucial for optimizing resource allocation and improving supply chain responsiveness. With the increasing complexity of global trade patterns, the growing volatility of market demand, and the diversified expansion of supply chain network nodes, achieving high-precision logistics demand forecasting in a complex environment of multiple coupled factors, and ensuring comprehensive capture of the dynamic relationships between upstream and downstream nodes and accurate quantification of traffic contribution relationships, has become a core challenge in this field.

[0003] In existing technologies, traditional logistics demand forecasting solutions primarily rely on single-dimensional data collection or basic statistical model calculations. These solutions typically use fixed-parameter models or empirical threshold methods to extrapolate demand by collecting basic data such as historical flow and simple interaction frequency at nodes. However, since the flow transfer between supply chain nodes is not a simple accumulation of quantities, but rather a dynamic coupling effect influenced by multiple factors such as supply capacity, demand fluctuations, and transmission delays, traditional solutions lack the ability of big data analytics to deeply mine multi-dimensional data, making it difficult to capture the hidden deep correlations between nodes. When faced with scenarios such as upstream supply fluctuations, sudden changes in downstream demand, or insufficient data samples in remote areas, traditional models cannot adaptively adjust their forecasting logic, often leading to a disconnect between forecast results and actual demand. This is especially problematic in scenarios with periodic fluctuations and abnormal events, where large forecast biases and poor adaptability are common, failing to provide a reliable basis for supply chain resource scheduling.

[0004] Existing technologies suffer from low accuracy in logistics demand forecasting. Summary of the Invention

[0005] This invention provides a supply chain logistics demand forecasting method and system based on neural networks, aiming to solve the technical problem of low forecasting accuracy in the prior art.

[0006] In a first aspect, the present invention provides a supply chain logistics demand forecasting method based on neural networks, comprising: Obtain historical traffic records of supply chain nodes, perform topology partitioning on the historical traffic records, and generate a network topology mapping; The abnormal occurrence time corresponding to the abnormal event is determined from the network topology mapping. The transmission delay difference between the logistics node and the upstream neighbor node corresponding to the abnormal occurrence time and the node influence characteristics are calculated. Then, the temporal feature learning is performed in combination with the preset gated cyclic unit model to obtain the dynamic interaction vector of each node. The activity characteristics of nodes are analyzed from the dynamic interaction vector to obtain the upstream interaction frequency and downstream demand fluctuation. The influence weight of the upstream interaction frequency on the downstream demand fluctuation is calculated. Based on the influence weight, the historical maximum carrying capacity is extracted to determine the traffic pressure distribution value. By integrating the flow pressure distribution value, the historical flow records, and the pre-acquired long-term trend growth ratio, a trend-corrected baseline flow is extracted, and a comprehensive feature description vector for each node is generated based on the trend-corrected baseline flow. The comprehensive feature description vector is used to identify short-term abnormal peaks and mark the time points of prediction deviation. If the deviation originates from upstream flow pressure, the weighting ratio is adjusted to obtain the corrected demand time series. The flow change pattern and time series correlation are extracted from the corrected demand time series. The smoothed correlation sequence is obtained by time series smoothing. The smoothed correlation sequence is dynamically adjusted in frequency range to generate demand forecast adjustment parameters. The upstream interaction frequency and the downstream demand fluctuation are weighted and corrected using the demand forecasting adjustment parameters to obtain the corrected interaction sequence and demand sequence. Cross-correlation analysis is then performed to obtain the flow transmission delay time. Based on the flow transmission delay time, periodic fluctuation characteristics are extracted and the demand forecasting result is determined.

[0007] Secondly, the present invention provides a supply chain logistics demand forecasting system based on neural networks, comprising: The topology mapping module is used to obtain historical traffic records of supply chain nodes, perform topology partitioning on the historical traffic records, and generate network topology mapping. The vector interaction module is used to determine the time of occurrence of an abnormal event corresponding to the abnormal event from the network topology mapping, calculate the transmission delay difference between the logistics node and the upstream neighbor node corresponding to the time of occurrence of the abnormal event and the node influence characteristics, and combine the preset gated cyclic unit model to perform time series feature learning to obtain the dynamic interaction vector of each node. The traffic calculation module is used to analyze the node activity characteristics from the dynamic interaction vector, obtain the upstream interaction frequency and downstream demand fluctuation, calculate the influence weight of the upstream interaction frequency on the downstream demand fluctuation, extract the historical maximum carrying capacity based on the influence weight, and determine the traffic pressure distribution value. The feature generation module is used to fuse the flow pressure distribution value, the historical flow records and the pre-acquired long-term trend growth ratio, extract the trend-corrected benchmark flow, and generate a comprehensive feature description vector for each node based on the trend-corrected benchmark flow. The sequence correction module is used to identify short-term abnormal peaks and mark the prediction deviation time points through the comprehensive feature description vector. If the deviation originates from upstream flow pressure, the weighting ratio is adjusted to obtain the corrected demand time series. The parameter generation module is used to extract the flow change pattern and time series correlation from the corrected demand time series, obtain a smoothed correlation sequence after time series smoothing, and perform dynamic frequency range adjustment on the smoothed correlation sequence to generate demand forecast adjustment parameters. The prediction output module is used to use the demand prediction adjustment parameters to perform weighted correction on the upstream interaction frequency and the downstream demand fluctuation to obtain the corrected interaction sequence and demand sequence, and to perform cross-correlation analysis to obtain the flow transmission delay time. Based on the flow transmission delay time, the module extracts periodic fluctuation characteristics and determines the demand prediction result.

[0008] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention integrates big data analysis and neural network technology to construct a weighted directed graph and network topology mapping, accurately capturing the spatial correlation and business characteristics between supply chain nodes, breaking through the limitations of traditional models in capturing the dynamic correlation of nodes. By using neighborhood aggregation, time series feature learning and other means, it realizes the quantitative analysis of the upstream and downstream traffic contribution relationship, significantly improving the accuracy of logistics demand prediction in complex network environments.

[0009] (2) This invention refines the entire process of supply chain logistics demand forecasting through full-link optimization steps such as traffic anomaly detection, demand sequence correction, and dynamic parameter adjustment. It can adaptively handle problems such as market fluctuations and data gaps. Especially for scenarios with insufficient data samples in remote areas, it effectively makes up for data shortcomings and improves the adaptability and reliability of forecasting through multi-dimensional feature fusion and periodic extrapolation.

[0010] (3) This invention extracts core features such as seasonality and node temporal correlation, and dynamically adjusts prediction parameters to reduce the impact of abnormal fluctuations and traffic transmission delays on prediction results. The resulting high-precision prediction results can provide a scientific basis for supply chain resource allocation, path optimization, and inventory management, helping to improve the overall stability and operational efficiency of the supply chain and reduce unnecessary resource waste. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of a supply chain logistics demand forecasting method based on neural networks provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of a supply chain logistics demand forecasting system based on neural networks, provided in the second embodiment of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] Reference Figure 1 The first embodiment of the present invention provides a supply chain logistics demand forecasting method based on neural networks, comprising the following steps: S1, Obtain historical traffic records of supply chain nodes, and perform topological partitioning on the historical traffic records to generate a network topology mapping; S2, determine the time of occurrence of the abnormal event corresponding to the abnormal event from the network topology mapping, calculate the transmission delay difference between the logistics node and the upstream neighbor node corresponding to the time of occurrence of the abnormal event and the node influence characteristics, and combine the preset gated cyclic unit model to learn the time series features to obtain the dynamic interaction vector of each node. S3. Analyze the node activity characteristics from the dynamic interaction vector to obtain the upstream interaction frequency and downstream demand fluctuation, calculate the influence weight of the upstream interaction frequency on the downstream demand fluctuation, extract the historical maximum carrying capacity based on the influence weight, and determine the traffic pressure distribution value. S4, integrate the flow pressure distribution value, the historical flow record and the pre-acquired long-term trend growth ratio, extract the trend correction benchmark flow, and generate a comprehensive feature description vector for each node based on the trend correction benchmark flow; S5. Identify short-term abnormal peaks and mark the prediction deviation time points through the comprehensive feature description vector. If the deviation originates from upstream flow pressure, adjust the weighting ratio to obtain the corrected demand time series. S6, extract the flow change pattern and time series correlation from the corrected demand time series, obtain a smoothed correlation sequence after time series smoothing, and perform dynamic frequency range adjustment on the smoothed correlation sequence to generate demand forecast adjustment parameters. S7. Using the demand forecasting adjustment parameters, the upstream interaction frequency and the downstream demand fluctuation are weighted and corrected to obtain the corrected interaction sequence and demand sequence. Cross-correlation analysis is performed to obtain the flow transmission delay time. Based on the flow transmission delay time, periodic fluctuation characteristics are extracted and the demand forecasting result is determined.

[0014] In step S1, historical traffic records of supply chain nodes are obtained, and the historical traffic records are topologically partitioned to generate a network topology mapping, including: S11 maps the historical flow records of supply chain nodes to node spatial coordinates to form a logistics node dataset. The historical flow records include cargo flow direction, delivery delay time and demand fluctuation data. S12, Based on the logistics node dataset, construct a weighted directed graph with nodes as vertices, cargo flow direction as directed edges, and transmission delay time as edge weights; S13, perform spatial clustering on the weighted directed graph to obtain a preliminary topological region, adjust the edge weights based on the demand fluctuation data, and adjust the preliminary topological region according to the adjusted edge weights to generate a revised topological region. S14, extract the spatial neighborhood features of the corrected topology region, map the spatial neighborhood features back to the weighted directed graph, and generate a network topology mapping.

[0015] In step S11, the historical flow records of the supply chain nodes are mapped to the node spatial coordinates to form a logistics node dataset. The historical flow records include cargo flow direction, delivery delay time and demand fluctuation data.

[0016] It should be noted that the historical traffic records for the nodes are collected over the past 12 months, covering complete seasonal cycles, promotional cycles, and regular business cycles. Specifically, this includes daily data on each node's inbound and outbound volumes, and remaining inventory, ensuring the data reflects long-term patterns in node traffic. Upstream and downstream interaction frequency includes the number of daily order interactions between nodes, weekly collaborative scheduling frequency, and monthly joint transportation frequency. Order interaction frequency must be accurate to the initiation and completion times of each order, and collaborative scheduling frequency must record the type and duration of scheduling tasks. Demand fluctuation data includes changes in node demand during promotional periods, holidays, and weekdays. For example, a node might experience a 30% increase in orders during major e-commerce promotions compared to normal times, or a 15% increase in demand during holidays. Inter-node transmission delay time includes the average, longest, and shortest transportation time from the shipping node to the receiving node. Transportation time is calculated using the shipment and receipt times recorded by the logistics tracking system. The spatial coordinates of the nodes are represented by latitude and longitude, and are obtained by converting the actual geographical location information of the nodes. For example, the latitude and longitude of warehouse center A are 116.4°E and 39.9°N, and the latitude and longitude of distribution station B are 116.5°E and 39.8°N. The collected business data is associated with latitude and longitude coordinates one by one, and node data with a missing rate of more than 5% is removed to form a logistics node dataset containing both location and business attributes, ensuring the completeness and usability of the dataset.

[0017] In step S12, a weighted directed graph is constructed based on the logistics node dataset, with nodes as vertices, cargo flow direction as directed edges, and transmission delay time as edge weights.

[0018] It should be noted that the vertices are defined as independent logistics nodes in the supply chain, including warehousing centers, distribution stations, production plants, and sales terminals. Each node is assigned a unique identifier, for example, "CZ-001" represents warehousing center number 001, and "PS-002" represents distribution station number 002. The directed edges are determined based on the actual flow of goods between nodes. If warehousing center CZ-001 consistently sends goods to distribution station PS-002, a directed edge is established from CZ-001 to PS-002; if there is a return flow from PS-002 to CZ-001, a reverse directed edge is established from PS-002 to CZ-001. The weights are set as the average of the transmission delay times between nodes. The calculation method is the sum of the transportation times of all transportation records along the path over the past 6 months divided by the total number of transportation records. For example, the CZ-001 to PS-002 route has 180 transportation records over the past 6 months, with a total transportation time of 360 hours, so the average transportation time is 2 hours, and the weight of the directed edge is set to 2. The reverse path has 30 transportation records over the past 6 months, with a total transportation time of 90 hours, so the average transportation time is 3 hours, and the weight of the reverse directed edge is set to 3. The weighted directed graph constructed in this way can accurately reflect the time cost differences of logistics paths, providing a quantitative basis for subsequent regional division and path optimization.

[0019] In step S13, spatial clustering is performed on the weighted directed graph to obtain a preliminary topological region. The edge weights are then corrected based on the demand fluctuation data. The preliminary topological region is then corrected according to the corrected edge weights to generate a corrected topological region.

[0020] It should be noted that the spatial clustering is implemented using the K-means algorithm. The number of clusters is determined based on the total number of nodes and geographical distribution density. If there are 50 nodes in the supply chain network and their geographical distribution is relatively concentrated, the number of clusters is set to 10 to ensure that each cluster region contains about 5 nodes, avoiding the impact of too many or too few nodes in a region on the analysis results. During the clustering process, the geographical distance between nodes and the weight of directed edges are used as the comprehensive clustering basis. The geographical distance is calculated as the straight-line distance through latitude and longitude conversion. The comprehensive similarity is calculated as a weighted sum of geographical distance similarity and weight similarity. The geographical distance similarity is the ratio of the preset maximum geographical distance to the actual geographical distance, and the weight similarity is the ratio of the preset maximum weight to the actual weight, with each weight accounting for 50%. Nodes with a comprehensive similarity higher than 80% are divided into the same preliminary region. For example, if the geographical distance similarity between node A and node B is 85%, the weight similarity is 82%, and the comprehensive similarity is 83.5%, then they are divided into the same region; if the comprehensive similarity between node A and node C is 75%, then they are divided into different regions. The method for correcting edge weights based on demand fluctuation data is as follows: First, calculate the demand fluctuation coefficient for each node, which is the ratio of the node's maximum demand to its average demand over the past 12 months. If the demand fluctuation coefficient exceeds 2, it indicates that the node's demand fluctuates drastically, and the weights of all directed edges related to that node are increased by 30% to highlight the importance of that path. If the demand fluctuation coefficient is below 0.8, it indicates stable demand, and the weights of related directed edges are reduced by 10% to avoid over-allocation of resources. For example, node D has a demand fluctuation coefficient of 2.3, and the original weight of the directed edges related to that node was 4; after correction, the weight is 5.2. Node E has a demand fluctuation coefficient of 0.7, and the original weight of the related directed edges was 3; after correction, the weight is 2.7. This correction method makes the topology region more closely reflect actual business demand changes, improving the rationality of resource allocation.

[0021] In step S14, the spatial neighborhood features of the corrected topology region are extracted, and the spatial neighborhood features are mapped back to the weighted directed graph to generate a network topology mapping.

[0022] It should be noted that the spatial neighborhood features include the density of nodes within the region, the average weight of directed edges between nodes, the average interaction frequency of nodes within the region, and the geographical area of ​​the region. Node density is calculated as the ratio of the number of nodes within the region to the geographical area of ​​the region. The geographical area is calculated as the polygon area obtained by using the latitude and longitude coordinates of the boundary nodes of the region. These spatial neighborhood features are associated with the edge weights and vertex attributes of the weighted directed graph. For example, if the node density within a region is higher than a preset density threshold (0.5 nodes / km²), the weight of all directed edges within that region is increased by an additional 10%, marking it as a high-priority region; if the average interaction frequency within a region is lower than a preset interaction threshold (5 times / day), the weight of the relevant edges is reduced by 5%. These adjusted features are then reintegrated into the weighted directed graph structure, updating the attribute labels of nodes and the weight values ​​of edges in the graph, generating a network topology mapping that combines spatial association and business characteristics, providing accurate basic data for subsequent feature extraction and model computation.

[0023] In step S2, the time of occurrence of the abnormal event corresponding to the abnormal event is determined from the network topology mapping. The transmission delay difference between the logistics node and the upstream neighboring node corresponding to the time of occurrence of the abnormal event and the node influence characteristics are calculated. Then, time-series feature learning is performed in combination with a preset gated cyclic unit model to obtain the dynamic interaction vector of each node, including: S21, Extract the historical inventory turnover rate of logistics nodes from the network topology mapping; S22, when the historical inventory turnover rate is detected to exceed the preset stable turnover rate, it is determined to be an abnormal event, and the time of the abnormality is determined; S23, calculate the transmission delay difference between the logistics node and the upstream neighboring node at the time of the anomaly, and use the neighborhood aggregation algorithm to extract the transmission delay difference to obtain the influence characteristics of the upstream on the downstream node. S24, the influence features are fused into a feature vector, which is then input into the gated recurrent unit model for temporal feature learning, and the dynamic interaction vector of each node is output.

[0024] In step S21, the historical inventory turnover rate of logistics nodes is extracted from the network topology mapping.

[0025] It should be noted that the historical inventory turnover rate of the nodes is calculated over the past 6 months, and is calculated as the ratio of monthly outbound volume to monthly average inventory holding. Monthly outbound volume is the total number of units shipped in that month, and monthly average inventory holding is the arithmetic mean of the beginning and ending inventory holdings for that month. For example, if node F's outbound volume in May is 100 units, its beginning inventory holding is 80 units, its ending inventory holding is 120 units, and its monthly average inventory holding is 100 units, then its inventory turnover rate for May is 1. The interaction timestamps include the specific time of each inventory change (accurate to the minute), the initiation time of the order interaction between nodes, and the completion time. For example, if node F receives an order from node G at 14:30 on May 10th and completes the outbound operation at 15:00, reducing its inventory by 50 units, the corresponding interaction timestamps are 14:30 (order initiated) and 15:00 (outbound completed). The historical inventory turnover rate of each node is aligned with the corresponding interaction timestamp in chronological order. The node status time series data is constructed with days as the time unit. Each time unit contains the inventory turnover rate of the day and all relevant interaction timestamps to ensure that the data can reflect the daily operating status and interaction patterns of the nodes. The time span covers the past 6 months to avoid feature bias caused by too short a data period.

[0026] In step S22, when the historical inventory turnover rate is detected to exceed the preset stable turnover rate, it is determined to be an abnormal event, and the time of the abnormality is determined.

[0027] It should be noted that the preset stability threshold is set based on the statistical analysis of historical inventory turnover rate fluctuations across all nodes, calculating the standard deviation of inventory turnover rates for all nodes, and using three times the standard deviation as the preset stability threshold. For example, if the standard deviation of inventory turnover rates for all nodes is 0.5, then the preset stability threshold is set to 1.5. The fluctuation amplitude is calculated as the absolute value of the difference between the inventory turnover rate within a certain time unit and the average inventory turnover rate over the past 30 days. For example, if node F's inventory turnover rate on June 15th is 2.8, and the average inventory turnover rate over the past 30 days is 1.2, the absolute value of the difference is 1.6, exceeding the preset stability threshold of 1.5. Therefore, it is determined that there is an abnormal fluctuation in the node status time series data within that time unit. This time unit (June 15th) is locked as the moment of occurrence of the traffic transmission anomaly, and all interaction timestamps and inventory changes at that moment are recorded to provide accurate time point references for subsequent anomaly analysis. This anomaly detection method is based on the fact that abnormal fluctuations in inventory turnover are usually directly related to abnormal flow transmission. Untimely supply from upstream will cause a sharp increase in inventory turnover, while a sudden drop in downstream demand will cause a sharp decrease in inventory turnover, which can quickly locate potential flow transmission problems.

[0028] In step S23, the transmission delay difference between the logistics node and the upstream neighboring node is calculated at the time the anomaly occurs, and the transmission delay difference is extracted using a neighborhood aggregation algorithm to obtain the influence characteristics of the upstream node on the downstream node.

[0029] It should be noted that the upstream neighboring nodes are defined as nodes that have a direct supply relationship with the affected node and have interacted at least 10 times in the past 3 months. For example, at the time of the anomaly, the upstream neighboring nodes of node F are nodes G and H, both of which have a stable supply relationship with node F, and their interaction frequencies in the past 3 months are 15 and 12 times respectively. The transmission delay difference is calculated by subtracting the average transmission delay time during the normal period from the transmission delay time from the upstream neighboring node to the affected node at the time of the anomaly. The normal period is defined as the non-abnormal period of 30 days before and after the time of the anomaly. For example, if the average transmission delay time from node G to node F during the normal period is 2 hours, and the transmission delay time at the time of the anomaly is 5 hours, then the transmission delay difference is 3 hours; if the average transmission delay time from node H to node F during the normal period is 1.5 hours, and the transmission delay time at the time of the anomaly is 4 hours, then the transmission delay difference is 2.5 hours. The neighborhood aggregation method employs a weighted average aggregation strategy. The weights are the percentage of interaction frequency between upstream neighboring nodes and affected nodes. The calculation method is to divide the interaction frequency of a specific upstream neighboring node with an affected node by the total interaction frequency of all upstream neighboring nodes with affected nodes. For example, the interaction frequency percentage between node G and node F is 15 / (15+12) = 55.6%, and that of node H is 44.4%. The extracted impact features include eight dimensions: transmission delay difference, interaction frequency percentage, historical average inventory turnover rate, demand fluctuation coefficient, average supply quantity, and supply stability. Each dimension's features are weighted and averaged according to their interaction frequency percentage to obtain the upstream impact feature vector for affected nodes. For example, the weighted average of the transmission delay difference is 3 × 55.6% + 2.5 × 44.4% = 2.78 hours. Other dimensions are calculated in the same way, ultimately forming an 8-dimensional impact feature vector that comprehensively reflects the overall impact of upstream nodes on downstream nodes.

[0030] In step S24, the node influence features are fused into a feature vector, input into a preset gated recurrent unit model for temporal feature learning, and output dynamic interaction vectors of each node.

[0031] It should be noted that the construction parameters of the gated recurrent unit model are as follows: input layer dimension is 8 (corresponding to an 8-dimensional influence feature vector), number of hidden layers is 3, number of neurons in each layer is 64, and output layer dimension is 16 (corresponding to a 16-dimensional dynamic interaction vector). The model uses the ReLU activation function, the Adam optimizer, the learning rate is set to 0.001, and the number of training iterations is set to 100. The training dataset uses node feature data and corresponding actual interaction effect data from the past 12 months. The actual interaction effect data includes order completion rate between nodes, logistics and transportation satisfaction, inventory turnover optimization rate, etc. Supervised learning samples are constructed using these data, enabling the model to learn the mapping relationship between feature vectors and actual interaction effects. The influence feature vector of each node is input into the trained gated recurrent unit model. Through the synergistic effect of the forget gate, update gate, and output gate, the model learns the temporal correlation information and dynamic change patterns in the feature vector, and outputs a 16-dimensional dynamic interaction vector that can represent the dynamic correlation strength of nodes. Each dimension of this vector corresponds to the association features of the node under different interaction scenarios. For example, the higher the value of a certain dimension, the stronger the association between the node and the upstream node in the promotion scenario; the lower the value of a certain dimension, the weaker the dependence of the node on the downstream node in the normal scenario, providing core feature support for subsequent critical path identification and traffic pressure calculation.

[0032] In step S3, the node activity characteristics are analyzed from the dynamic interaction vector to obtain the upstream interaction frequency and downstream demand fluctuation. The influence weight of the upstream interaction frequency on the downstream demand fluctuation is calculated. Based on the influence weight, the historical maximum carrying capacity is extracted to determine the flow pressure distribution value, including: S31, Analyze the frequency and strength of the connection between the node and its surrounding nodes from the dynamic interaction vector to obtain the node activity characteristics, and perform statistics on the node activity characteristics to obtain the upstream interaction frequency. S32, map the upstream interaction frequency to the downstream node time axis to obtain the downstream demand fluctuation, and use the correlation coefficient to calculate the influence weight of the upstream interaction frequency on the downstream demand fluctuation. S33, when the influence weight is detected to exceed the preset association strength threshold, the corresponding path is marked as a critical transmission path, and the historical maximum carrying capacity of the critical transmission path is extracted; S34, integrate the historical maximum carrying capacity with the peak amplitude of the downstream demand fluctuation to calculate the upstream flow pressure distribution value to the downstream.

[0033] In step S31, the frequency and strength of the connection between the node and its surrounding nodes are analyzed from the dynamic interaction vector to obtain the node activity characteristics, and the upstream interaction frequency is obtained by statistically analyzing the node activity characteristics.

[0034] It should be noted that the separation of node activity features employs principal component analysis (PCA). This method uses orthogonal transformation to convert the 16-dimensional data of the dynamic interaction vector into several principal components. The top three principal components, which account for 85% of the cumulative variance explanation, are selected as the core dimensions of the node activity features, ensuring that the separated features fully reflect the key information of node activity. The original features corresponding to these three principal components include interaction frequency, interaction duration, and interaction stability. Interaction frequency is the total number of interactions over the past six months, interaction duration is the average duration of each interaction, and interaction stability is the coefficient of variation of interaction frequency (the ratio of the standard deviation of interaction frequency to the mean). Based on the feature data of these three core dimensions, the upstream interaction frequency is statistically obtained, specifically including the average number of daily interactions, weekly peak interactions, monthly total interactions, and interaction time distribution. The statistical period for upstream interaction frequency is consistent with the statistical period for downstream demand fluctuations, both being on the daily level, ensuring the accuracy of subsequent correlation analysis.

[0035] It is worth noting that the dynamic interaction vector is a high-dimensional feature vector output after learning by a gated recurrent unit model, which encodes the temporal pattern and correlation strength of interactions between nodes. In order to extract explicit upstream interaction frequency indicators from it, principal component analysis is first used to reduce the dimensionality of the dynamic interaction vector, obtaining the top three principal components with a cumulative variance explanation of 85%. Through principal component loading matrix analysis, it is identified that these three principal components are highly correlated with the original interaction frequency, interaction duration, and interaction stability of the nodes, respectively. Subsequently, based on the inverse transformation of the principal component score matrix and the loading matrix, the normalized interaction frequency, interaction duration, and interaction stability values ​​are approximately reconstructed. Then, combined with the original historical traffic records of the nodes, such as order timestamps and interaction logs, the upstream interaction frequency indicators such as the daily average number of interactions, weekly peak interaction, and monthly total interaction are calculated using conventional statistical methods. This process utilizes the feature extraction capabilities of neural networks while ensuring the interpretability and traceability of business indicators.

[0036] In step S32, the upstream interaction frequency is mapped to the downstream node time axis to obtain the downstream demand fluctuation, and the influence weight of the upstream interaction frequency on the downstream demand fluctuation is calculated using the correlation coefficient.

[0037] It should be noted that the timeline mapping method involves mapping upstream interaction frequencies to downstream node demand data one-to-one using timestamps. For example, the average daily interaction frequency of upstream nodes corresponds to the daily demand of downstream nodes, ensuring that upstream interaction frequencies and downstream demand data form paired samples within each time unit. The time span is the past 6 months, resulting in 180 paired samples. The correlation coefficient is calculated by first calculating the covariance between upstream interaction frequency and downstream demand fluctuations, then dividing by the product of their standard deviations. The covariance reflects the consistency of the overall trend of the two sets of data, while the standard deviation reflects the dispersion of each set of data. The correlation coefficient ranges from -1 to 1. The closer the value is to 1, the stronger the positive correlation between the two sets of data, i.e., the higher the upstream interaction frequency, the greater the downstream demand; the closer the value is to -1, the stronger the negative correlation; and the closer the value is to 0, the weaker the correlation. For example, if the covariance between the average daily number of upstream interactions and the daily demand of downstream is 120, the standard deviation of upstream interaction frequency is 15, and the standard deviation of downstream demand data is 8, then the correlation coefficient is 120 divided by (15 multiplied by 8), which is 1, indicating that the two are perfectly positively correlated and the influence weight is 1. If the correlation coefficient is 0.8, then the influence weight is 0.8, which directly reflects the strength of the influence of upstream interaction frequency on downstream demand fluctuations.

[0038] In step S33, when the influence weight is detected to exceed the preset association strength threshold, the corresponding path is marked as a critical transmission path, and the historical maximum carrying capacity of the critical transmission path is extracted.

[0039] It should be noted that the preset correlation strength threshold is set based on the statistical distribution of the correlation coefficients between upstream interaction frequency and downstream demand fluctuations among all nodes, with the 75th percentile used as the preset correlation strength threshold. Analyzing historical data from the past 12 months, the correlation coefficients between all nodes range from 0.2 to 1.0, with the 75th percentile at 0.6. Therefore, the preset correlation strength threshold is set to 0.6. If the influence weight of a path exceeds 0.6, it indicates that the upstream interaction frequency of that path has a significant impact on downstream demand fluctuations, and the path is marked as a critical transmission path. For example, the influence weight of the path "Node G - Node F" is 0.75, exceeding the threshold of 0.6, and is marked as a critical transmission path; the influence weight of the path "Node H - Node I" is 0.5, not reaching the threshold, and is not marked as a critical path. The extraction of the historical maximum carrying capacity involves querying the historical operating data of the critical transmission path and extracting the maximum freight volume of that path over the past 12 months, including the maximum daily, weekly, and monthly transport volumes. The historical maximum carrying capacity is determined by taking the maximum value among the extracted maximum transport volumes. For example, if the maximum daily transport volume of a critical route is 500 tons, the maximum weekly transport volume is 3,000 tons, and the maximum monthly transport volume is 12,000 tons in the past 12 months, then the historical maximum carrying capacity of the route is determined to be 12,000 tons, reflecting the maximum transport capacity limit of the route.

[0040] In step S34, the upstream flow pressure distribution value is calculated by combining the historical maximum carrying capacity with the peak amplitude of the downstream demand fluctuation.

[0041] It should be noted that the peak amplitude of the downstream demand fluctuation is calculated as the difference between the maximum and average demand of the downstream node over the past 6 months. For example, if the average daily demand of downstream node F over the past 6 months is 300 tons and the maximum daily demand is 600 tons, then the peak amplitude is 300 tons. If the average monthly demand is 9,000 tons and the maximum monthly demand is 15,000 tons, then the peak amplitude is 6,000 tons. The statistical period for the peak amplitude is consistent with the statistical period for the historical maximum capacity to ensure a unified calculation method. The flow pressure distribution value is calculated by dividing the downstream demand peak amplitude by the historical maximum capacity of the critical path, and then multiplying by 100 to obtain a percentage pressure value. For example, if the peak amplitude is 300 tons and the historical maximum capacity is 500 tons, then the flow pressure distribution value is (300 divided by 500) multiplied by 100, resulting in 60%. If the peak amplitude is 6,000 tons and the historical maximum capacity is 12,000 tons, then the flow pressure distribution value is 50%. The pressure value represents the level of pressure on the critical path during peak downstream demand. A higher value indicates greater pressure on the path and a greater likelihood of transportation bottlenecks. When the pressure value exceeds 80%, it indicates that the path may experience insufficient capacity during peak demand periods, requiring advance adjustments to resource allocation.

[0042] In step S4, the flow pressure distribution value, the historical flow record, and the pre-acquired long-term trend growth ratio are fused to extract the trend-corrected baseline flow, and a comprehensive feature description vector for each node is generated based on the trend-corrected baseline flow.

[0043] S41, The flow pressure distribution value is embedded into the historical flow record through time alignment mapping to generate a pressure-weighted historical sequence; S42, Load the long-term trend growth ratio onto the pressure-weighted historical sequence, and obtain the trend-corrected baseline flow rate by linear trend extrapolation; S43, Perform frequency domain decomposition on the trend-corrected baseline flow to extract seasonal factor values; S44. Construct a weighted fusion matrix based on the factor values, perform convolution operation between the weighted fusion matrix and the flow pressure distribution values ​​to obtain pressure features, and generate comprehensive feature description vectors for each node through vectorization compression.

[0044] In step S41, the flow pressure distribution value is embedded into the historical flow record through time alignment mapping to generate a pressure-weighted historical sequence.

[0045] It should be noted that the time alignment mapping method involves mapping the flow pressure distribution value to the historical flow record of each node one-to-one according to the timestamp. Each historical flow record for each time unit (day) corresponds to a flow pressure distribution value, ensuring consistency across the time dimension. The time span is the past 6 months, resulting in 180 sets of corresponding data. The historical flow record of each node represents the actual daily inflow and outflow volume. For example, the historical flow of node F on June 1st was 400 tons, corresponding to a flow pressure distribution value of 60%. The pressure-weighted historical sequence is generated by multiplying the historical flow value of each time unit by the corresponding flow pressure distribution value, and then dividing by 100 for normalization. The calculation method is: historical flow value multiplied by flow pressure distribution value and then divided by 100. For example, the weighted value for June 1st is 400 multiplied by 60 and then divided by 100, resulting in 240 tons. This weighting is based on the fact that the flow pressure distribution value can reflect the pressure level of the flow within that time unit. The weighted data can highlight the flow characteristics during high-pressure periods, providing more accurate basic data for subsequent trend analysis and periodic extraction.

[0046] In step S42, the long-term trend growth ratio is loaded onto the pressure-weighted historical sequence, and the trend-corrected baseline flow rate is obtained by linear trend extrapolation.

[0047] It should be noted that the long-term trend growth rate is calculated as the average annual growth rate of the historical pressure-weighted sequence over the past 12 months. The calculation logic is (year-end sequence mean minus year-end sequence mean) divided by the year-end sequence mean. For example, if the monthly average of the pressure-weighted sequence at the beginning of the year is 200 tons and the monthly average at the end of the year is 240 tons, then the average annual growth rate is (240 minus 200) divided by 200, resulting in 20%. The loading method involves multiplying each value in the pressure-weighted historical sequence by (1 plus the long-term trend growth rate) to obtain the trend prediction flow sequence. For example, if the weighted value for a certain time unit is 240 tons and the long-term trend growth rate is 20%, then the trend prediction value is 240 multiplied by (1 plus 20%), resulting in 288 tons. The linear trend extrapolation method is implemented by determining a linear trend line based on the first and last two time unit values ​​of the trend prediction flow sequence. The first and last values ​​are the trend prediction values ​​of the first and last time units of the sequence, respectively. The flow values ​​for the next three months are predicted according to the trend line equation. The predicted values ​​are then merged with the original trend prediction flow sequence to obtain the trend-corrected baseline flow. This baseline flow includes both historical trends and forward-looking predictions, aligning with the long-term development patterns of the business.

[0048] In step S43, frequency domain decomposition is performed on the trend-corrected baseline flow to extract seasonal factor values.

[0049] It should be noted that the purpose of the frequency domain decomposition is to convert the flow sequence in the time domain into frequency components in the frequency domain. Each frequency component corresponds to a different fluctuation period. The decomposition process involves mathematically transforming a continuous trend-corrected baseline flow sequence into a superposition of multiple sinusoidal components. The seasonality factor values ​​are extracted from the amplitude of each major frequency component. The larger the amplitude, the stronger the fluctuation intensity of the corresponding period. For example, the amplitude of the weekly periodic component is 50, the amplitude of the monthly periodic component is 45, and the amplitude of the quarterly periodic component is 40. These values ​​are the seasonality factor values, which can quantify the fluctuation intensity of different periods.

[0050] In step S44, a weighted fusion matrix is ​​constructed based on the factor values. The weighted fusion matrix is ​​then convolved with the flow pressure distribution values ​​to obtain pressure features, which are then vectorized and compressed to generate comprehensive feature description vectors for each node.

[0051] It should be noted that the dimension of the weighted fusion matrix is ​​determined by the number of major seasonal factors. If three major factors (weekly, monthly, and quarterly) are extracted, a 3×3 weighted fusion matrix is ​​constructed. The matrix element values ​​are the normalized results of the corresponding seasonal factor values. The normalization is calculated by dividing each factor value by the sum of all factor values. For example, if the three factor values ​​are 50, 45, and 40, and the sum is 135, the normalized values ​​are 50 / 135≈0.37, 45 / 135≈0.33, and 40 / 135≈0.30, respectively. The elements on the main diagonal of the matrix are set to these normalized values, and the remaining elements are 0, ensuring that the matrix highlights the influence weight of the major factors. The convolution operation process involves using a weighted fusion matrix as the convolution kernel to perform a sliding window calculation on the one-dimensional sequence of flow pressure distribution values. The window size is consistent with the convolution kernel dimension (3 time units). The calculation result for each window is the sum of the products of the flow pressure distribution values ​​within the window and the corresponding elements of the convolution kernel for the 3 time units within the window. For example, if the values ​​within the window are 60%, 55%, and 58%, and the convolution kernel elements are 0.37, 0.33, and 0.30, then the calculation result is 60×0.37+55×0.33+58×0.30=57.6, which is 57.6% in percentage form. The vectorization compression uses principal component analysis to compress the high-dimensional convolution operation result (180 values) into a 12-dimensional vector. The variance explanation ratio of principal component analysis is set to 90% to ensure that the compressed vector retains more than 90% of the core features. The final 12-dimensional vector is the comprehensive feature description vector of each node, fully integrating pressure features, trend features, and seasonal features.

[0052] In step S5, short-term abnormal peaks are identified and prediction deviation time points are marked through the comprehensive feature description vector. If the deviation originates from upstream flow pressure, the weighting ratio is adjusted to obtain the corrected demand time series.

[0053] S51, Based on the comprehensive feature description vector, extract the upstream flow pressure distribution value, use sliding window variance detection to identify short-term abnormal peaks, obtain flow data fluctuations, and mark the prediction deviation time points according to the flow data fluctuations; S52, calculate the temporal overlap rate and amplitude synchronization rate between the abnormal peak value at the predicted deviation time point and the upstream flow pressure distribution value; S53, if both the time domain overlap rate and the amplitude synchronization rate exceed the preset correlation threshold, the deviation is determined to originate from upstream flow pressure, and upstream influencing factors are extracted to generate a modified weighted coefficient matrix. S54, reconstruct the upstream flow pressure distribution value according to the modified weighted coefficient matrix, adjust the upstream influence weighting ratio, and obtain the corrected demand time series.

[0054] In step S51, based on the comprehensive feature description vector, the upstream flow pressure distribution value is extracted, and short-term abnormal peaks are identified by sliding window variance detection to obtain flow data fluctuations. The prediction deviation time points are marked according to the flow data fluctuations.

[0055] It should be noted that the sliding window is set to a window size of 7 days and a sliding step of 1 day. The window size is determined based on the fact that the short-term fluctuation cycle of logistics demand usually does not exceed 7 days, and this window can effectively capture short-term anomalies. The sliding window process is applied to the demand sequence corresponding to the comprehensive feature description vector, with each window containing demand data for 7 consecutive days. The variance is calculated as the sum of the squared differences between each demand data point within the window and the average demand data within the window, divided by the number of data points within the window (7). The criteria for determining an abnormal window is that the variance value exceeds 3 times the average variance. The average variance is the arithmetic mean of the variance values ​​of all windows. For example, if the average variance of all windows is 5000, 3 times the average variance is 15000. If the variance of a certain window is 18000, it is determined to be an abnormal window. The short-term abnormal peak value is the maximum demand data within the abnormal window. For example, if the maximum demand data of the above abnormal window is 500 tons, it is an abnormal peak value. The prediction deviation time point is the time unit (day) corresponding to the abnormal peak. If multiple demand data exceed the normal range (95th percentile of demand data in the past 6 months) within the abnormal window, the time units corresponding to these data are marked as prediction deviation time points. The 95th percentile of the normal range is calculated by arranging the demand data in the past 6 months in ascending order and taking the data at the 95th position as the threshold.

[0056] In step S52, the temporal overlap rate and amplitude synchronization rate between the abnormal peak value at the predicted deviation time point and the upstream flow pressure distribution value are calculated.

[0057] It should be noted that the mapping method involves perfectly aligning the demand data corresponding to the abnormal peak value with the upstream flow pressure distribution value by timestamp, ensuring that each prediction deviation time point corresponds to the pressure data of the upstream node. The time domain overlap rate is calculated by dividing the number of time units where the prediction deviation time point overlaps with the upstream pressure fluctuation period by the total number of time units of the prediction deviation time point. The upstream pressure fluctuation period is defined as a continuous time unit where the upstream flow pressure distribution value exceeds 120% of the average pressure value. For example, if the prediction deviation time point has a total of 5 days, and 3 days overlap with the upstream pressure fluctuation period, then the time domain overlap rate is 3 divided by 5, resulting in 60%. The amplitude synchronization rate is calculated by dividing the number of time units where the change trend of the abnormal peak value is consistent with the change trend of the upstream flow pressure distribution value by the number of overlapping time units. Consistent change trends mean that both rise or fall simultaneously. For example, if both the abnormal peak value and the upstream pressure show an upward trend within the overlapping 3 days, then the amplitude synchronization rate is 3 divided by 3, resulting in 100%; if 2 days are consistent and 1 day is inconsistent, then the amplitude synchronization rate is 2 divided by 3, resulting in 66.7%.

[0058] In step S53, if both the temporal overlap rate and the amplitude synchronization rate exceed the preset correlation threshold, the deviation is determined to originate from upstream flow pressure, and upstream influencing factors are extracted to generate a modified weighted coefficient matrix.

[0059] It should be noted that the preset correlation thresholds are set based on historical deviation cases over the past 12 months. Deviation events originating from upstream flow pressure all had a temporal overlap rate exceeding 50% and an amplitude synchronization rate exceeding 70%. Therefore, the temporal overlap rate threshold is set to 50%, and the amplitude synchronization rate threshold is set to 70%. If the temporal overlap rate at a certain predicted deviation time point is 60% (exceeding 50%) and the amplitude synchronization rate is 80% (exceeding 70%), the deviation is determined to originate from upstream flow pressure. If either indicator fails to meet the criteria, such as a temporal overlap rate of 45% or an amplitude synchronization rate of 75%, the deviation is determined to originate from other factors. The upstream influencing factors are extracted as six core dimensions related to the deviation, including the rate of change in transmission delay of upstream nodes, the rate of change in interaction frequency, the rate of change in inventory turnover rate, the supply volume fluctuation coefficient, transportation stability, and demand response speed. The modified weighted coefficient matrix is ​​constructed as a 6×6 matrix. The main diagonal elements are the normalized results of the correlation coefficients between each influencing factor and the deviation of downstream demand. The correlation coefficient is calculated by dividing the covariance of the factor data and the deviation data by the product of their standard deviations. The normalization method is to divide the absolute value of each correlation coefficient by the sum of the absolute values ​​of all correlation coefficients to ensure that the sum of the weights is 1. For example, if the absolute values ​​of the correlation coefficients of the six factors are 0.8, 0.7, 0.6, 0.5, 0.4, and 0.3, respectively, and the sum is 3.3, the normalized weights are 0.24, 0.21, 0.18, 0.15, 0.12, and 0.09, respectively. The remaining elements of the matrix are 0, ensuring that the matrix can accurately reflect the influence weight of each factor.

[0060] In step S54, the upstream flow pressure distribution value is reconstructed according to the modified weighted coefficient matrix, the upstream influence weighting ratio is adjusted, and the corrected demand time series is obtained.

[0061] It should be noted that the reconstruction method of the upstream flow pressure distribution value is to multiply the original pressure value of each upstream node by the corresponding weight in the modified weighting coefficient matrix, and then sum them to obtain the reconstructed pressure value. The adjustment logic of the upstream influence weighting ratio is as follows: if the reconstructed pressure value is 20% or more higher than the original pressure value, the influence weighting ratio of the corresponding upstream node is increased by 15%; if the reconstructed pressure value is 20% or more lower than the original pressure value, the weighting ratio is decreased by 10%; if the change is within ±20%, the weighting ratio remains unchanged. For example, if the original pressure value is 50% and the reconstructed value is 62% (an increase of 24%), the weighting ratio of the corresponding upstream node increases from 0.3 to 0.345 (0.3 × 1.15). The corrected demand time series is generated by multiplying the demand data at the predicted deviation time points in the original demand time series by (1 plus the difference between the adjusted weighting ratio and the original weighting ratio). For example, if the original demand data is 500 tons and the difference in the adjusted weighting ratio is 0.045, then the corrected demand data is 500 × (1 + 0.045) = 522.5 tons. The demand data at the non-deviation time points remain unchanged, ultimately forming a corrected demand time series that reflects actual business fluctuations.

[0062] It is worth noting that the original upstream flow pressure distribution value was a single scalar value representing the comprehensive pressure level of a specific upstream node on downstream nodes. To achieve refined reconstruction based on influencing factors, this scalar pressure value is first expanded into a six-dimensional feature vector. The six upstream influencing factors extracted in step S53 are calculated using the actual business data of the upstream node at the current moment. The current values ​​of delay change rate, interaction frequency change rate, inventory turnover rate change rate, supply volume fluctuation coefficient, transportation stability, and demand response speed are then passed. The value of each factor is multiplied by the original value of the upstream node. The initial pressure scalar value is used to obtain a six-dimensional weighted eigenvector. Subsequently, this six-dimensional weighted eigenvector is multiplied element-wise by the modified weighting coefficient matrix (6×6 diagonal matrix). That is, each dimension of the six-dimensional weighted eigenvector is multiplied by the normalized weight of the corresponding dimension on the main diagonal of the modified weighting coefficient matrix, and the six products are summed. The sum is the reconstructed upstream flow pressure distribution value. This reconstructed value retains the original pressure amplitude baseline and incorporates the contribution weight of each influencing factor under the current deviation scenario, which can more accurately reflect the actual pressure transmission intensity of the upstream node.

[0063] In step S6, the flow change pattern and time series correlation are extracted from the corrected demand time series, and a smoothed correlation sequence is obtained after time series smoothing. The smoothed correlation sequence is then dynamically adjusted in frequency range to generate demand forecast adjustment parameters.

[0064] S61, The corrected demand time series is processed using a local weighted regression scatter smoothing algorithm to extract the flow change pattern; S62, extract the neighboring node traffic logs from the historical traffic records, combine the neighboring node traffic logs with the traffic change pattern to determine the temporal correlation between nodes, and perform Gaussian filtering on the temporal correlation to obtain a smooth correlation sequence; S63, calculate the rate of change of the first derivative of the smoothed correlation sequence, and generate a dynamically adjusted frequency range according to a preset rate of change threshold; S64, calculate the weighted average deviation compensation value based on the dynamically adjusted frequency range, and generate demand forecast adjustment parameters.

[0065] In step S61, the corrected demand time series is processed using a local weighted regression scatter smoothing algorithm to extract the flow rate variation pattern.

[0066] It should be noted that the core parameter of the local weighted regression scatter smoothing algorithm is a bandwidth of 0.8. This bandwidth parameter was set based on multiple experimental verifications. This parameter can effectively smooth short-term noise interference while preserving the trend characteristics of the demand sequence, avoiding feature loss due to over-smoothing. The algorithm's calculation logic is as follows: for each data point in the corrected demand time series, a certain range of neighboring data points are selected as the center. Weights are assigned to the neighboring data points based on their distance from the center data point. A weighted regression is then used to fit and obtain the smoothed value of the center data point. For example, if the center data point is 320 tons on day 10, and the neighboring data points are demand data from days 5 to 15, a smoothed value of 315 tons is obtained after weighting based on distance. The extraction of seasonal patterns involves analyzing the smoothed demand sequence using an autocorrelation function. The autocorrelation function is calculated as the correlation coefficient between the current demand sequence and its own sequence delayed by k time units, where k ranges from 1 to 30 days. The peak value of the autocorrelation coefficient corresponding to the peak k value represents the main seasonal cycle. For example, when k=7, the autocorrelation coefficient reaches a peak of 0.85, indicating a weekly cycle; when k=30, the peak autocorrelation coefficient is 0.72, indicating a monthly cycle. Traffic flow patterns within each main cycle are extracted. For instance, within a weekly cycle, traffic on Mondays and Fridays is 15%-20% higher than other days, with Sunday having the lowest traffic. Within a monthly cycle, traffic in the middle and latter part of the month is about 25% higher than in the first ten days, forming a structured database of traffic flow patterns.

[0067] In step S62, neighboring node traffic logs are extracted from the historical traffic records. The temporal correlation between nodes is determined by combining the neighboring node traffic logs with the traffic change pattern. Gaussian filtering is then applied to the temporal correlation to obtain a smooth correlation sequence.

[0068] It should be noted that the neighboring nodes are defined as nodes geographically within 50 kilometers of the target node and that have had stable logistics interactions (interaction frequency of no less than 20 times) within the past 6 months. For example, the neighboring nodes of target node F are delivery stations G, H, and I. The collection of the traffic logs of the neighboring nodes includes daily traffic data, order interaction records, and cargo transit details for the past 6 months, ensuring the integrity of the log data. The temporal correlation between the nodes is determined by calculating the cross-correlation coefficient of the traffic sequence between the target node and each neighboring node. A cross-correlation coefficient exceeding 0.6 is considered to indicate a dependency relationship, and the correlation type is determined as leading correlation, synchronous correlation, or lagging correlation. For example, the cross-correlation coefficient between target node F and neighboring node G is 0.75, and the traffic change of G leads F by 2 days, which is determined to be a leading correlation. The Gaussian filter is used to smooth time-series correlated sequences. The standard deviation of the filter is set to 1.5. The filtering logic is to assign weights to neighboring data points according to a Gaussian distribution, with each data point as the center, and calculate the weighted average as the smoothed value. The purpose is to remove short-term noise interference and highlight long-term dependence trends. For example, abnormal fluctuation data in time-series correlated sequences are corrected to values ​​that conform to long-term trends after filtering, resulting in a smoothed correlated sequence.

[0069] In step S63, the rate of change of the first derivative of the smoothed correlation sequence is calculated, and a dynamic adjustment frequency range is generated based on a preset rate of change threshold.

[0070] It should be noted that the first derivative is calculated as the difference between the correlation values ​​of two adjacent time units in the smoothed correlation sequence, reflecting the rate of change of the correlation sequence. For example, if the correlation value on day 10 is 0.75 and on day 11 it is 0.82, the first derivative is 0.07; if it is 0.82 on day 11 and on day 12 it is 0.78, the first derivative is -0.04. The rate of change of the first derivative is calculated as the absolute value of the difference between the current first derivative and the previous first derivative, divided by the absolute value of the previous first derivative. For example, if the previous first derivative was 0.07 and the current one is -0.04, the absolute value of the difference is 0.11, and the rate of change is 0.11 divided by 0.07, resulting in 157.1%. The preset rate of change threshold is set to 50%. This threshold is based on the normal rate of change of correlation sequences in statistical historical data. A normal rate of change of more than 90% is below 50%, and a rate of change exceeding 50% indicates a drastic change in the correlation relationship. The method for dividing the dynamic adjustment frequency interval is as follows: the time period between two adjacent time points where the rate of change exceeds 50% is divided into a dynamic interval. The frequency adjustment period of each interval is determined according to the average rate of change of the associated sequence within the interval: when the average rate of change is higher than 100%, the adjustment period is 1 day; when the average rate of change is between 50% and 100%, the adjustment period is 3 days; when the average rate of change is lower than 50%, the adjustment period is 7 days, ensuring that the frequency adjustment can adapt to the fluctuation intensity of the associated relationship.

[0071] In step S64, a weighted average deviation compensation value is calculated based on the dynamic adjustment frequency range to generate demand forecast adjustment parameters.

[0072] It should be noted that the weighted average deviation compensation value is calculated by statistically analyzing the predicted deviation data within each dynamic adjustment frequency interval. The deviation data is calculated as the difference between the actual demand data and the original predicted data within the interval. Weights are assigned based on the adjustment period length of each interval; the longer the adjustment period, the greater the weight. The weight is calculated by dividing the interval's adjustment period length by the sum of the adjustment period lengths of all intervals. For example, if the adjustment periods for the three dynamic intervals are 1 day, 3 days, and 7 days, with a total length of 11 days, the weights are approximately 1 / 11 ≈ 0.09, 3 / 11 ≈ 0.27, and 7 / 11 ≈ 0.64, respectively. The weighted average deviation compensation value is calculated by multiplying the average deviation value of each interval by its corresponding weight and then summing the results. The average deviation value is the arithmetic mean of all deviation data within the interval. For example, if the average deviation values ​​for the three intervals are 25 tons, 18 tons, and 12 tons, the weighted average deviation compensation value is approximately 25 × 0.09 + 18 × 0.27 + 12 × 0.64 ≈ 14.8 tons. The demand forecast adjustment parameters consist of a weighted average deviation compensation value, a dynamic adjustment frequency range, and a seasonal adjustment coefficient. The seasonal adjustment coefficient is determined based on the seasonal patterns extracted in step S61. Each major season (spring, summer, autumn, and winter) corresponds to an adjustment coefficient. For example, the coefficient for higher demand in summer is 1.1, and the coefficient for lower demand in winter is 0.9. The three together constitute the complete demand forecast adjustment parameters, providing the core adjustment basis for subsequent forecasts.

[0073] In step S7, the upstream interaction frequency and the downstream demand fluctuation are weighted and corrected using the demand forecasting adjustment parameters to obtain the corrected interaction sequence and demand sequence. Cross-correlation analysis is then performed to obtain the flow transmission delay time. Based on the flow transmission delay time, periodic fluctuation characteristics are extracted and the demand forecasting result is determined.

[0074] S71, the upstream interaction frequency and the downstream demand fluctuation are weighted and corrected using the demand forecasting adjustment parameters to obtain the corrected interaction sequence and demand sequence. S72, perform cross-correlation analysis on the interaction sequence and the demand sequence to determine the traffic transmission delay time; S73, Align the interaction sequence and the demand sequence according to the traffic transmission delay time to obtain alignment data, and extract periodic fluctuation characteristics from the alignment data; S74, Based on the periodic fluctuation characteristics, perform time-series extrapolation to determine the demand forecast result.

[0075] In step S71, the upstream interaction frequency and the downstream demand fluctuation are weighted and corrected using the demand forecasting adjustment parameters to obtain the corrected interaction sequence and demand sequence.

[0076] It should be noted that the collection period for the new data is the most recent month. Upstream interaction frequency includes the number of daily interactions, total number of interactions, and distribution of interaction time periods. Downstream demand fluctuations include daily demand volume, demand change rate, and peak demand periods. Data collection must ensure real-time performance and completeness (missing rate less than 3%). The weighted correction coefficient is applied by multiplying the new data by the corresponding correction coefficient in the demand forecast adjustment parameters. For example, if the correction coefficient for upstream interaction frequency is 1.2 and the number of interactions on a certain day is 30, then the corrected number of interactions is 30 × 1.2 = 36. If the correction coefficient for downstream demand data is 1.1 (summer seasonality coefficient), and the demand volume on a certain day is 200 tons, then the corrected demand volume is 200 × 1.1 = 220 tons. The deviation compensation value is superimposed by directly adding the weighted average deviation compensation value (e.g., 14.8 tons) to the corrected downstream demand data. For example, if the corrected demand volume is 220 tons, the superimposed value is 234.8 tons. The modified interaction sequence and demand sequence are generated by arranging all the modified daily data in timestamp order to form a continuous sequence, ensuring that the sequence can reflect the latest business fluctuations and is consistent with the statistical caliber of historical data.

[0077] It is worth noting that the demand forecasting adjustment parameters also include an upstream interaction frequency correction coefficient. This coefficient is not an independently preset value, but is calculated based on the dynamic adjustment frequency interval generated in S64 and the weighted average deviation compensation value, combined with the upstream influencing factors extracted in S53. Specifically, for each dynamic adjustment frequency interval, the average percentage deviation between the actual observed value of the upstream node interaction frequency in that interval and the historical benchmark value is first calculated. Then, this average deviation percentage is weighted and fused with the weighted average deviation compensation value calculated in S64. The fusion weight is determined by the length of the interval adjustment period; the shorter the adjustment period, the greater the weight, thus obtaining the upstream interaction frequency correction coefficient corresponding to that interval. Finally, the correction coefficients of all dynamic intervals are weighted and averaged according to the interval length to obtain the final unified correction coefficient used for the upstream interaction frequency.

[0078] For example, if the average upstream interaction frequency deviates from the benchmark value by +20% within a certain interval, the weighted average deviation compensation value is +5%, and the fused correction coefficient is 1.25. In practical applications, the correction of the upstream interaction frequency adopts a multiplicative correction, that is, the corrected interaction frequency is equal to the original interaction frequency multiplied by the upstream interaction frequency correction coefficient. The correction of downstream demand fluctuations is carried out independently. The demand data is first multiplied by the seasonal adjustment coefficient, and then the weighted average deviation compensation value is added (additive correction). Through the above clear division of labor, the demand forecast adjustment parameters fully cover both upstream and downstream correction operations, ensuring that the weighted correction steps are clear and implementable.

[0079] In step S72, cross-correlation analysis is performed on the interaction sequence and the demand sequence to determine the traffic transmission delay time.

[0080] It should be noted that the purpose of the cross-correlation analysis is to identify the time delay in the impact of changes in upstream interaction frequency on downstream demand. The time lag range for the analysis is set to 0-10 days, which covers a reasonable time interval for logistics delivery, from 1 day for short-distance transportation to 7-10 days for long-distance transportation. The cross-correlation coefficient is calculated by taking each lag of k days (k=0,1,...,10) and calculating the correlation coefficient between the interaction sequence and the demand sequence lags by k days. For example, the correlation coefficient is 0.88 when k=2, 0.65 when k=3, and 0.72 when k=1. The flow transmission delay time is determined as the lag day corresponding to the highest cross-correlation coefficient. For example, the correlation coefficient is highest when k=2 (0.88), so the flow transmission delay time is determined to be 2 days. This indicates that changes in upstream interaction frequency will have the most significant impact on downstream demand after 2 days, and this delay time provides the core basis for subsequent data alignment.

[0081] In step S73, the interaction sequence and the demand sequence are aligned according to the traffic transmission delay time to obtain aligned data, and periodic fluctuation characteristics are extracted from the aligned data.

[0082] It should be noted that the data alignment method involves shifting the corrected upstream interaction sequence forward by a corresponding delay time (e.g., 2 days) to match the upstream interaction data with the downstream demand data in the time dimension. For example, upstream interaction data on June 1st is aligned to downstream demand data on June 3rd, ensuring that the causal relationship between the two is accurately reflected. The extraction of periodic demand fluctuation characteristics is based on the aligned sequence, using autocorrelation function and power spectral density analysis. The autocorrelation function is used to identify the period length, and the power spectral density is used to quantify the fluctuation intensity of each period. Considering the special characteristics of remote areas, the focus is on extracting their unique periodic features. Even with limited data samples in these areas, hidden periodic patterns can still be captured through aligned sequence analysis, providing support for accurate prediction.

[0083] In step S74, time-series extrapolation is performed based on the periodic fluctuation characteristics to determine the demand forecast result.

[0084] It should be noted that the time-series extrapolation uses the ARIMA model with parameters set to p=2, d=1, and q=2. These parameters were determined using the AIC and BIC information criteria, and this parameter combination minimizes the criterion values, ensuring optimal model fit. The model's training data consists of aligned sequence data from the past 12 months. The training process involves the model learning the trend, periodicity, and random fluctuation characteristics of the sequences, enabling it to accurately simulate the changing patterns of historical data. The extrapolation period is the next three months. The extrapolation logic involves the model using the trained parameters, combined with the extracted periodic fluctuation characteristics, to predict the demand for each time unit (day / week) in the future. For example, it predicts that in remote areas, the average demand on Wednesdays will be 4.5 tons over the next three months, while the average demand on other days will be 2.5 tons, with the demand in the last week of each month being 30% higher than other weeks. The forecast results are output in the form of structured data, including daily / weekly forecast demand, forecast confidence interval (95% confidence level), and risk warnings. This provides a high-precision decision-making basis for resource allocation, inventory management, and transportation planning in remote areas, effectively compensating for the forecast bias caused by insufficient data in these areas.

[0085] Reference Figure 2 The second embodiment of the present invention provides a supply chain logistics demand forecasting system based on neural networks, comprising: The topology mapping module is used to obtain historical traffic records of supply chain nodes, perform topology partitioning on the historical traffic records, and generate network topology mapping. The vector interaction module is used to determine the time of occurrence of an abnormal event corresponding to the abnormal event from the network topology mapping, calculate the transmission delay difference between the logistics node and the upstream neighbor node corresponding to the time of occurrence of the abnormal event and the node influence characteristics, and combine the preset gated cyclic unit model to perform time series feature learning to obtain the dynamic interaction vector of each node. The traffic calculation module is used to analyze the node activity characteristics from the dynamic interaction vector, obtain the upstream interaction frequency and downstream demand fluctuation, calculate the influence weight of the upstream interaction frequency on the downstream demand fluctuation, extract the historical maximum carrying capacity based on the influence weight, and determine the traffic pressure distribution value. The feature generation module is used to fuse the flow pressure distribution value, the historical flow records and the pre-acquired long-term trend growth ratio, extract the trend-corrected benchmark flow, and generate a comprehensive feature description vector for each node based on the trend-corrected benchmark flow. The sequence correction module is used to identify short-term abnormal peaks and mark the prediction deviation time points through the comprehensive feature description vector. If the deviation originates from upstream flow pressure, the weighting ratio is adjusted to obtain the corrected demand time series. The parameter generation module is used to extract the flow change pattern and time series correlation from the corrected demand time series, obtain a smoothed correlation sequence after time series smoothing, and perform dynamic frequency range adjustment on the smoothed correlation sequence to generate demand forecast adjustment parameters. The prediction output module is used to use the demand prediction adjustment parameters to perform weighted correction on the upstream interaction frequency and the downstream demand fluctuation to obtain the corrected interaction sequence and demand sequence, and to perform cross-correlation analysis to obtain the flow transmission delay time. Based on the flow transmission delay time, the module extracts periodic fluctuation characteristics and determines the demand prediction result.

[0086] It should be noted that the neural network-based supply chain logistics demand forecasting system provided in this embodiment of the invention is used to execute all the process steps of the neural network-based supply chain logistics demand forecasting method described in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0087] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0088] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A supply chain logistics demand forecasting method based on neural networks, characterized in that, include: Obtain historical traffic records of supply chain nodes, perform topology partitioning on the historical traffic records, and generate a network topology mapping; The abnormal occurrence time corresponding to the abnormal event is determined from the network topology mapping. The transmission delay difference between the logistics node and the upstream neighbor node corresponding to the abnormal occurrence time and the node influence characteristics are calculated. Then, the temporal feature learning is performed in combination with the preset gated cyclic unit model to obtain the dynamic interaction vector of each node. The activity characteristics of nodes are analyzed from the dynamic interaction vector to obtain the upstream interaction frequency and downstream demand fluctuation. The influence weight of the upstream interaction frequency on the downstream demand fluctuation is calculated. Based on the influence weight, the historical maximum carrying capacity is extracted to determine the traffic pressure distribution value. By integrating the flow pressure distribution value, the historical flow records, and the pre-acquired long-term trend growth ratio, a trend-corrected baseline flow is extracted, and a comprehensive feature description vector for each node is generated based on the trend-corrected baseline flow. The comprehensive feature description vector is used to identify short-term abnormal peaks and mark the time points of prediction deviation. If the deviation originates from upstream flow pressure, the weighting ratio is adjusted to obtain the corrected demand time series. The flow change pattern and time series correlation are extracted from the corrected demand time series. The smoothed correlation sequence is obtained by time series smoothing. The smoothed correlation sequence is dynamically adjusted in frequency range to generate demand forecast adjustment parameters. The upstream interaction frequency and the downstream demand fluctuation are weighted and corrected using the demand forecasting adjustment parameters to obtain the corrected interaction sequence and demand sequence. Cross-correlation analysis is then performed to obtain the flow transmission delay time. Based on the flow transmission delay time, periodic fluctuation characteristics are extracted and the demand forecasting result is determined.

2. The method according to claim 1, characterized in that, The step of obtaining historical traffic records of supply chain nodes and performing topology partitioning on the historical traffic records to generate a network topology mapping includes: The historical flow records of supply chain nodes are mapped to node spatial coordinates to form a logistics node dataset. The historical flow records include cargo flow direction, delivery delay time and demand fluctuation data. Based on the aforementioned logistics node dataset, a weighted directed graph is constructed with nodes as vertices, cargo flow direction as directed edges, and transmission delay time as edge weights. Spatial clustering is performed on the weighted directed graph to obtain a preliminary topological region. The edge weights are then adjusted based on the demand fluctuation data. The preliminary topological region is then corrected based on the adjusted edge weights to generate a corrected topological region. Extract the spatial neighborhood features of the corrected topology region, map the spatial neighborhood features back to the weighted directed graph, and generate a network topology mapping.

3. The method according to claim 1, characterized in that, The step of determining the time of occurrence of the abnormal event corresponding to the abnormal event from the network topology mapping, and calculating the transmission delay difference and node impact characteristics between the logistics node and the upstream neighboring node corresponding to the time of occurrence of the abnormal event, includes: Extract the historical inventory turnover rate of logistics nodes from the network topology mapping; When the historical inventory turnover rate is detected to exceed the preset stable turnover rate, it is determined to be an abnormal event, and the time of the abnormality is determined. Calculate the transmission delay difference between the logistics node and the upstream neighboring node at the time the anomaly occurs, and use the neighborhood aggregation algorithm to extract the transmission delay difference to obtain the upstream's influence on the downstream node.

4. The method according to claim 1, characterized in that, The process involves analyzing node activity characteristics from the dynamic interaction vector to obtain upstream interaction frequency and downstream demand fluctuations, calculating the influence weight of the upstream interaction frequency on the downstream demand fluctuations, extracting the historical maximum carrying capacity based on the influence weights, and determining the traffic pressure distribution value, including: The frequency and strength of connections between nodes and surrounding nodes are analyzed from the dynamic interaction vectors to obtain node activity characteristics. The upstream interaction frequency is obtained by statistically analyzing the node activity characteristics. The upstream interaction frequency is mapped to the downstream node time axis to obtain the downstream demand fluctuation, and the influence weight of the upstream interaction frequency on the downstream demand fluctuation is calculated using the correlation coefficient. When the influence weight is detected to exceed the preset association strength threshold, the corresponding path is marked as a critical transmission path, and the historical maximum carrying capacity of the critical transmission path is extracted. By combining the historical maximum carrying capacity with the peak amplitude of the downstream demand fluctuation, the upstream flow pressure distribution value to the downstream is calculated.

5. The method according to claim 1, characterized in that, The process involves fusing the flow pressure distribution value, the historical flow records, and the pre-acquired long-term trend growth ratio to extract a trend-corrected baseline flow. Based on this trend-corrected baseline flow, a comprehensive feature description vector for each node is generated, including: The traffic pressure distribution value is embedded into the historical traffic record by time alignment mapping to generate a pressure-weighted historical sequence. The long-term trend growth ratio is loaded onto the pressure-weighted historical sequence, and the trend-corrected baseline flow rate is obtained by extrapolating the linear trend. Perform frequency domain decomposition on the trend-corrected baseline flow to extract seasonality factor values; A weighted fusion matrix is ​​constructed based on the factor values. The weighted fusion matrix is ​​then convolved with the flow pressure distribution values ​​to obtain pressure features. Finally, the features are vectorized and compressed to generate a comprehensive feature description vector for each node.

6. The method according to claim 1, characterized in that, The process of identifying short-term abnormal peaks and marking prediction deviation time points using the comprehensive feature description vector, and adjusting the weighting ratio to obtain the corrected demand time series if the deviation originates from upstream flow pressure, includes: Based on the comprehensive feature description vector, the upstream flow pressure distribution value is extracted, and short-term abnormal peaks are identified by sliding window variance detection to obtain flow data fluctuations. The prediction deviation time points are marked according to the flow data fluctuations. Calculate the temporal overlap rate and amplitude synchronization rate between the abnormal peak value at the predicted deviation time point and the upstream flow pressure distribution value; If both the temporal overlap rate and the amplitude synchronization rate exceed the preset correlation threshold, the deviation is determined to originate from upstream flow pressure, and upstream influencing factors are extracted to generate a corrected weighted coefficient matrix. The upstream flow pressure distribution value is reconstructed based on the modified weighted coefficient matrix, and the upstream influence weighting ratio is adjusted to obtain the corrected demand time series.

7. The method according to claim 1, characterized in that, The process of extracting flow change patterns and time-series correlations from the corrected demand time series, obtaining a smoothed correlation sequence through time series smoothing, and dynamically adjusting the frequency range of the smoothed correlation sequence to generate demand forecast adjustment parameters includes: The corrected demand time series was processed using a local weighted regression scatter smoothing algorithm to extract the flow change pattern; The traffic logs of neighboring nodes are extracted from the historical traffic records. The temporal correlation between nodes is determined by combining the traffic logs of neighboring nodes with the traffic change pattern. The temporal correlation is then subjected to Gaussian filtering to obtain a smooth correlation sequence. Calculate the rate of change of the first derivative of the smoothed correlation sequence, and generate a dynamically adjusted frequency range based on a preset rate of change threshold; The weighted average deviation compensation value is calculated based on the dynamically adjusted frequency range, and the demand forecast adjustment parameters are generated.

8. The method according to claim 1, characterized in that, The process of using the demand forecasting adjustment parameters to weight and correct the upstream interaction frequency and the downstream demand fluctuations to obtain the corrected interaction sequence and demand sequence, performing cross-correlation analysis to obtain the traffic transmission delay time, extracting periodic fluctuation characteristics based on the traffic transmission delay time, and determining the demand forecasting result includes: The upstream interaction frequency and the downstream demand fluctuation are weighted and corrected using the demand forecasting adjustment parameters to obtain the corrected interaction sequence and demand sequence. Perform cross-correlation analysis on the interaction sequence and the demand sequence to determine the traffic transmission delay time; Alignment data is obtained by aligning the interaction sequence and the demand sequence according to the traffic transmission delay time, and periodic fluctuation characteristics are extracted from the alignment data; Based on the aforementioned periodic fluctuation characteristics, time-series extrapolation is performed to determine the demand forecast results.

9. A supply chain logistics demand forecasting system based on neural networks, characterized in that, include: The topology mapping module is used to obtain historical traffic records of supply chain nodes, perform topology partitioning on the historical traffic records, and generate network topology mapping. The vector interaction module is used to determine the time of occurrence of an abnormal event corresponding to the abnormal event from the network topology mapping, calculate the transmission delay difference between the logistics node and the upstream neighbor node corresponding to the time of occurrence of the abnormal event and the node influence characteristics, and combine the preset gated cyclic unit model to perform time series feature learning to obtain the dynamic interaction vector of each node. The traffic calculation module is used to analyze the node activity characteristics from the dynamic interaction vector, obtain the upstream interaction frequency and downstream demand fluctuation, calculate the influence weight of the upstream interaction frequency on the downstream demand fluctuation, extract the historical maximum carrying capacity based on the influence weight, and determine the traffic pressure distribution value. The feature generation module is used to fuse the flow pressure distribution value, the historical flow records and the pre-acquired long-term trend growth ratio, extract the trend-corrected benchmark flow, and generate a comprehensive feature description vector for each node based on the trend-corrected benchmark flow. The sequence correction module is used to identify short-term abnormal peaks and mark the prediction deviation time points through the comprehensive feature description vector. If the deviation originates from upstream flow pressure, the weighting ratio is adjusted to obtain the corrected demand time series. The parameter generation module is used to extract the flow change pattern and time series correlation from the corrected demand time series, obtain a smoothed correlation sequence after time series smoothing, and perform dynamic frequency range adjustment on the smoothed correlation sequence to generate demand forecast adjustment parameters. The prediction output module is used to use the demand prediction adjustment parameters to perform weighted correction on the upstream interaction frequency and the downstream demand fluctuation to obtain the corrected interaction sequence and demand sequence, and to perform cross-correlation analysis to obtain the flow transmission delay time. Based on the flow transmission delay time, the module extracts periodic fluctuation characteristics and determines the demand prediction result.

Images