Logistics data analysis decision system and method based on cloud platform

Abstract

The invention relates to the technical field of logistics data analysis, and discloses a logistics data analysis decision-making system and method based on a cloud platform, and the method comprises the steps: obtaining a multi-source data set, carrying out the time-space standardization processing, and calculating a credibility score; constructing a logistics network graph, predicting the state of each node in the logistics network graph, and identifying bottleneck nodes and key paths; order time sequence features and external factor features are extracted, and the logistics demand quantity of each region is predicted; constructing a distribution task set, predicting road section passing time, and optimizing a distribution path and task distribution; carrying out supply and demand matching analysis, and carrying out resource scheduling optimization; identifying a key scene, simulating an execution process and calculating expected performance and risk indexes of the decision scheme; according to the method, intelligence, automation and optimization of logistics decision making are realized, logistics operation efficiency is improved, operation cost is reduced, service quality is improved, and the method has important theoretical value and practical significance.

Description

Technical Field

[0001] This invention relates to the field of logistics data analysis technology, and more specifically, to a cloud-based logistics data analysis and decision-making system and method. Background Technology

[0002] With the rapid development of cross-border e-commerce, logistics companies need to handle massive orders from multiple countries and regions, coordinate various transportation modes such as sea, air, and land transport, and deal with complex factors such as customs clearance and traffic regulations in different countries. In actual operation, logistics demand is affected by a combination of factors such as promotional activities, holidays, and weather conditions, exhibiting characteristics of sudden fluctuations and uneven spatial and temporal distribution. Meanwhile, the logistics network consists of multiple entities such as warehousing centers, distribution stations, and transportation routes, with complex relationships and dynamic constraints among these entities.

[0003] Traditional logistics data analysis methods are primarily based on single data sources and static models, making them ill-suited to dynamic market environments and complex business scenarios. Existing solutions for multi-source heterogeneous data fusion mostly employ simple data integration methods, lacking dynamic assessment of the reliability of different data sources across time and space, and intelligent fusion mechanisms, resulting in low-quality fused data. In logistics network status assessment, existing methods mainly rely on statistical analysis to evaluate the status of individual nodes or paths, lacking in-depth modeling of the overall network topology and local-global relationships, making it difficult to identify potential bottlenecks and cascading risks. In demand forecasting, existing methods primarily rely on historical time series data, failing to effectively quantify and model the coupled effects of multiple factors, leading to insufficient forecast accuracy when facing sudden demand fluctuations. Regarding route planning, existing methods… Most existing methods rely on static assumptions, assuming all constraints are known and fixed during planning. This makes it difficult to cope with dynamic constraints such as real-time changes in traffic conditions, customs clearance times, and multi-segment transportation connections, leading to a decrease in the actual executability and optimality of the planning scheme. In terms of resource scheduling, existing methods mainly adopt a passive response mechanism, that is, resource allocation is carried out only after demand is generated. They lack the ability to proactively pre-schedule based on demand forecasting and network status assessment, and often react slowly when faced with sudden increases in demand. In terms of decision support, existing systems usually only provide a single decision option, lacking the ability to quantitatively analyze uncertainties and compare and evaluate multiple options, making it difficult to provide managers with comprehensive decision-making basis.

[0004] Therefore, there is a need for a logistics data analysis and decision-making system and method that can integrate multi-source heterogeneous data, dynamically assess network status, accurately predict demand, optimize routes in real time, robustly schedule resources, and intelligently support decision-making, in order to solve complex decision-making problems in cross-border logistics. Summary of the Invention

[0005] This invention provides a cloud-based logistics data analysis and decision-making system and method, which solves the technical problems in related technologies such as low quality of multi-source data fusion, incomplete network status assessment, insufficient accuracy of demand forecasting, difficulty in coping with dynamic constraints in route planning, lag in resource scheduling response, and limited decision support.

[0006] This invention provides a cloud-based logistics data analysis and decision-making method, including: We acquire multi-source datasets, perform spatiotemporal standardization on data at different time granularities, calculate confidence scores, and use an adaptive weight fusion method to obtain the fused logistics dataset. Based on the fused logistics dataset, a logistics network graph is constructed, the status of each node in the logistics network graph is predicted, bottleneck nodes and critical paths are identified, and a logistics network status assessment report is obtained. Based on the logistics network status assessment report and the fused logistics dataset, order time-series features are extracted and an external factor feature matrix is ​​constructed to predict logistics demand and obtain logistics demand prediction results. Based on the logistics demand forecasting results, a set of delivery tasks is constructed, the travel time of road segments is predicted, and the delivery routes and task allocation are optimized to obtain a dynamic delivery route scheme. Based on the logistics demand forecast results and dynamic delivery route plans, supply and demand matching analysis is performed, and resource scheduling is optimized according to the supply and demand matching analysis results to obtain a resource pre-scheduling plan. Based on dynamic delivery route schemes and resource pre-scheduling schemes, key scenarios are identified, the execution process is simulated, and the expected performance and risk indicators of the decision-making schemes are calculated to generate intelligent decision support reports.

[0007] In a preferred embodiment, the calculation of the credibility score specifically includes: Acquire data from the same spatiotemporal location in a multi-source dataset, calculate the Pearson correlation coefficient between different data sources in the same spatiotemporal location, and evaluate the degree of consistency between data sources based on the correlation coefficient. A basic credibility score is established based on historical accuracy statistics from various data sources; Timeliness weights are calculated based on data update frequency and collection time intervals. The consistency level, basic credibility score, and timeliness weight are weighted and combined to obtain a comprehensive credibility score.

[0008] In a preferred embodiment, predicting the state of each node in the logistics network graph specifically includes: Obtain the adjacency matrix of the logistics network graph, and normalize the adjacency matrix to obtain the normalized adjacency matrix. Based on the fused logistics dataset, features of warehouse center nodes, delivery station nodes, and customer nodes are extracted, and node feature vectors are constructed. Two graph convolutional layers are used to process the node feature vectors. Each layer aggregates the neighborhood node information and updates the node representation through a normalized adjacency matrix. Based on the real state labels of a small number of labeled nodes, a semi-supervised learning method is used to train the updated node representations to predict the state of all nodes.

[0009] In a preferred embodiment, the prediction of logistics demand in each region specifically includes: Based on the logistics network status assessment report and the fused logistics dataset, an external factor feature matrix is ​​extracted. Element-wise multiplication is performed on any two features in the external factor feature matrix to obtain second-order interactive features. Third-order interaction features are obtained by further cross-referencing the second-order interaction features with the original features. The attention pooling mechanism is used to weight and aggregate the second-order and third-order interaction features to obtain the fused features; Based on fusion features, a sequence-to-sequence model is used to predict the logistics demand in various regions over future time periods.

[0010] In a preferred embodiment, the prediction of logistics demand in each region further includes: The fused feature sequence is obtained, and a bidirectional long short-term memory network encoder is used to encode the fused feature sequence. The forward and backward dependencies of the temporal features are extracted to obtain the encoder output. The Long Short-Term Memory Network Decoder employs an attention mechanism, which focuses on the relevant temporal information of the encoder output in each decoding step. The decoder output is mapped to the predicted logistics demand for each region through a fully connected layer.

[0011] In a preferred embodiment, the optimization of delivery routes and task allocation specifically includes: Based on the logistics demand forecast results, a set of delivery tasks is constructed, and the K-means clustering algorithm is used to divide the delivery tasks into multiple sub-regions according to geographical location and delivery time window constraints; For each sub-region, a vehicle routing planning sub-problem is constructed. The objective function of the sub-problem includes transportation cost, time cost, and service quality penalty. The Lagrange relaxation method is used to handle the coupling constraints between subproblems, decomposing the original problem into independent subproblems; Each subproblem is allocated independent computing resources on the cloud platform for parallel optimization and solution.

[0012] In a preferred embodiment, the resource scheduling optimization specifically includes: Based on the historical demand data in the fused logistics dataset, statistical analysis is performed to calculate the mean and standard deviation of demand in each region and to construct a probability distribution model for demand uncertainty. A robust optimization objective function is constructed based on the worst-case demand scenario. The objective function includes a weighted combination of resource scheduling cost and demand satisfaction. Establish resource availability constraints, resource transfer time constraints, and non-negativity constraints; The optimal resource scheduling scheme is obtained by solving the robust optimization model using a linear programming solver.

[0013] In a preferred embodiment, the generation of the intelligent decision support report specifically includes: Based on historical logistics operation data in the fused logistics dataset, K-means clustering algorithm is used for cluster analysis to identify typical logistics operation scenario categories; Based on the cosine similarity calculation between the current state and each typical scenario, the key scenario most similar to the current state is selected. The Monte Carlo method is used to generate random samples of key scenarios, and parallel simulation analysis is performed on a cloud platform. Based on the simulation results, a weighted evaluation method is used to comprehensively evaluate different decision-making schemes, generating a decision support report that includes risk analysis and sensitivity analysis.

[0014] In a preferred embodiment, the step of comprehensively evaluating different decision-making schemes using a weighted evaluation method based on simulation results specifically includes: The weighted comprehensive score of the decision-making scheme in terms of cost, efficiency, and service quality is calculated based on the simulation results of each scenario. Sensitivity analysis was used to identify the key parameters and uncertainties that have the greatest impact on decision-making outcomes. Based on a risk assessment model, the potential risks and benefits of different decision-making options are quantified; Generate comprehensive decision support reports that include decision recommendations, risk warnings, key indicator forecasts, and visualization charts.

[0015] In a preferred embodiment, a cloud-based logistics data analysis and decision-making system is used to execute the aforementioned cloud-based logistics data analysis and decision-making method, including: The multi-source data fusion module is used to acquire multi-source datasets, perform spatiotemporal standardization on data at different time granularities, calculate confidence scores, and use an adaptive weight fusion method to obtain the fused logistics dataset. The network status assessment module constructs a logistics network graph based on the fused logistics dataset, predicts the status of each node in the logistics network graph, identifies bottleneck nodes and critical paths, and obtains a logistics network status assessment report. The demand forecasting module extracts order time-series features and constructs an external factor feature matrix based on the logistics network status assessment report and the fused logistics dataset to predict logistics demand and obtain logistics demand forecasting results. The route planning module constructs a set of delivery tasks based on logistics demand forecasting results, predicts road segment travel time, optimizes delivery routes and task allocation, and obtains dynamic delivery route solutions. The resource scheduling module performs supply and demand matching analysis based on logistics demand forecasting results and dynamic delivery route plans, and optimizes resource scheduling based on the supply and demand matching analysis results to obtain a resource pre-scheduling plan. The decision support module identifies key scenarios based on dynamic delivery route plans and resource pre-scheduling plans, simulates the execution process, calculates the expected performance and risk indicators of the decision-making plan, and generates an intelligent decision support report.

[0016] The beneficial effects of this invention are as follows: By using an adaptive multi-source data fusion method based on spatiotemporal weights, which comprehensively considers the credibility of data sources, data timeliness, and spatial correlation, and dynamically adjusts the fusion weights, the heterogeneity and inconsistency of logistics data are effectively solved, improving the quality and reliability of the fused data and providing a high-quality data foundation for subsequent analysis and decision-making. Compared with traditional simple weighted fusion, it can more accurately reflect the quality differences and spatiotemporal characteristics of different data sources, avoiding the adverse effects of low-quality data on the fusion results. By employing a semi-supervised graph neural network-based logistics network state assessment method, this approach fully utilizes network topology and limited labeled data to accurately predict the state of all nodes. It effectively identifies bottleneck nodes and critical paths within the network, providing comprehensive network state information for resource allocation and route planning. By aggregating neighborhood information through graph convolution operations, it captures the relationships between nodes and the global characteristics of the network. Compared to traditional methods based on single-node statistical analysis, the assessment results are more accurate and comprehensive, especially achieving good results even with scarce labeled data. Furthermore, demand forecasting is conducted through a multi-factor coupled deep interactive network. This not only considers the temporal characteristics of historical orders but also incorporates external factors such as promotional activities, holidays, and weather. Multi-layer feature cross-capture captures the nonlinear interactions between factors, improving the accuracy and robustness of demand forecasting and providing reliable demand prediction for resource pre-scheduling and route planning. By employing a cloud-based parallel computing-based dynamic path planning method, this approach decomposes large-scale path planning problems into multiple sub-problems for parallel solution using task decomposition and distributed optimization techniques. An iterative coordination mechanism ensures global consistency, shortening computation time and meeting real-time decision-making requirements. Furthermore, this method fully considers dynamic constraints such as real-time traffic conditions, customs clearance times, and multi-segment transportation connections, resulting in path solutions with stronger practical feasibility and optimality. Through robust resource scheduling optimization, based on uncertainties in demand forecasting, proactive resource pre-scheduling is performed, and dynamic adjustments are made through real-time feedback. This effectively addresses sudden fluctuations in demand, improving the robustness and flexibility of resource allocation and reducing the risks of resource shortages and idle resources. This invention employs an intelligent decision support method based on key scenario identification. It uses cluster analysis to identify typical scenarios, Monte Carlo sampling and parallel simulation techniques to evaluate the performance of decision-making schemes under different scenarios, and sensitivity analysis to identify key influencing factors. This provides managers with multi-dimensional and visualized decision support information, including performance expectations, risk assessments, and optimization suggestions, thus improving the scientific rigor and comprehensiveness of decision-making. Furthermore, this invention can quantify the impact of uncertainties on decision outcomes, helping managers weigh multiple objectives, formulate strategies to address different scenarios, and reduce operational risks. Attached Figure Description

[0017] Figure 1 This is a flowchart of the cloud platform-based logistics data analysis and decision-making method in this invention; Figure 2 This is a module diagram of the cloud-based logistics data analysis and decision-making system in this invention. Detailed Implementation

[0018] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0019] At least one embodiment of the present invention discloses a logistics data analysis and decision-making method based on a cloud platform, such as... Figure 1 As shown, it includes: Step 1: Obtain multi-source datasets, perform spatiotemporal standardization on data at different time granularities, calculate confidence scores, and use an adaptive weight fusion method to obtain the fused logistics dataset; The purpose of this step is to intelligently integrate logistics data from different channels to solve the problems of data heterogeneity and reliability differences.

[0020] Based on the data interface of the logistics business system, a mechanism combining timed polling and event triggering is used to acquire multi-source datasets, including order datasets, warehousing datasets, transportation datasets, customs datasets, weather datasets, and traffic datasets. The order dataset includes fields such as order number, product information, delivery address, order time, and expected delivery time. The warehousing dataset includes fields such as warehouse number, inventory quantity, inbound time, and outbound time. The transportation dataset includes fields such as vehicle number, current location, speed, and load capacity. The customs dataset includes fields such as customs declaration number, clearance status, and clearance time. The weather dataset includes fields such as region code, temperature, rainfall, wind speed, and data collection time. The traffic dataset includes fields such as road segment number, traffic volume, average speed, congestion index, and data collection time.

[0021] Based on the timestamp fields of the aforementioned datasets, a sliding time window alignment method is used to uniformly align data with different time granularities to a standard time granularity. Specifically, the standard time granularity is set to 5 minutes. For data with a time granularity greater than 5 minutes (such as hourly weather data), a linear interpolation method is used to generate data points with a 5-minute granularity. For data with a time granularity less than 5 minutes (such as second-level GPS positioning data), the average value within the time window is used to aggregate data points with a 5-minute granularity, resulting in the time-aligned dataset.

[0022] Based on the address information field in the time-aligned dataset, a geocoding service interface is used to convert the text-based address into latitude and longitude coordinates. For the delivery address in the order data, the geocoding API is called to obtain its latitude and longitude coordinates. For the warehouse address in the warehousing data, a pre-established warehouse location database is queried to obtain standard coordinates. For the vehicle location in the transportation data, the latitude and longitude information reported by the GPS device is directly extracted.

[0023] Based on the acquired latitude and longitude coordinates, a coordinate system transformation method is used to convert different coordinate systems into the WGS84 international standard coordinate system. The transformation process is implemented through a coordinate correction algorithm, which performs coordinate translation and rotation transformations according to the offset parameters between different coordinate systems and WGS84, resulting in a spatially standardized dataset.

[0024] Based on the spatially standardized dataset, cross-validation is used to evaluate the credibility of each data source. For information about the same logistics entity (such as the same vehicle or the same warehouse) in the same spatiotemporal location, there may be data from multiple data sources. The credibility of the data sources is evaluated by calculating the degree of consistency among these data.

[0025] Specifically, for the first and second data sources, a set of logistics entities they both cover is selected, and the correlation coefficient between the data values ​​provided by the two data sources on these entities is calculated. Logistics entities commonly covered by the two data sources are identified, such as the same vehicle or warehouse, ensuring that data from the same entities is being compared. For each common entity k, data values ​​provided by the first and second data sources are extracted to form data pairs. The arithmetic mean of the data values ​​provided by the first and second data sources on all common entities is calculated, serving as the benchmark for subsequent deviation calculations. For each common entity k, the deviation between the data value from the first data source and its mean, and the deviation between the data value from the second data source and its mean, are calculated, and the two deviations are multiplied. The products of deviations for all entities are summed. The sum of squared deviations of the first and second data sources on all common entities is calculated, i.e., the sum of the squares of the differences between each data value and its mean. The sum of the deviation products is divided by the square root of the product of the two deviation sums to obtain the Pearson correlation coefficient between the two data sources, which reflects the degree of linear correlation between the data from the two data sources.

[0026] Based on the correlation coefficient matrix among the data sources, a consistency score calculation method is used to calculate the basic credibility score of each data source. For each data source, its accuracy performance in providing data over a historical period is statistically analyzed, and the basic credibility score is calculated as the basic weight for that data source. For the target primary data source, its correlation coefficient with all other secondary data sources is multiplied by the corresponding basic credibility score of the secondary data source, and all weighted correlation coefficients are summed. The weighted correlation coefficient sum is divided by the sum of the basic credibility scores of all other data sources to obtain a standardized credibility score. The credibility score is fine-tuned based on factors such as the data coverage and update frequency of the data sources to ensure that the score comprehensively reflects the reliability of the data sources. Based on the credibility scores of each data source, a data source credibility vector is constructed, which contains the credibility scores of all data sources and provides a basis for subsequent weight calculations.

[0027] Based on the data source credibility vector and the collection time of each data point, an adaptive time decay function is used to calculate the time weight. Different types of data have different timeliness characteristics. For rapidly changing data (such as real-time traffic data and vehicle location data), the timeliness decays quickly, requiring a large decay coefficient. For slowly changing data (such as infrastructure data and historical statistical data), the timeliness decays slowly, requiring a small decay coefficient.

[0028] In the time weight calculation process, based on the characteristics of logistics data, the data is divided into different types and corresponding decay coefficients are set: traffic data and transportation data are assigned a decay coefficient of 0.5 (unit: 1 / hour) to reflect their rapid changes; weather data is assigned a decay coefficient of 0.1 to consider the relative stability of weather changes; order data and warehousing data are assigned a decay coefficient of 0.05 to reflect their relatively stable characteristics; and customs data is assigned a decay coefficient of 0.2 to balance its update frequency and importance. The time interval between the collection time of each data point and the current time is calculated and quantified in hours. The time weight is calculated using an exponential decay function, which multiplies the data source's credibility score by an exponential function value with a base of the natural constant e and an exponent of a negative decay coefficient multiplied by the time interval. Finally, the calculated time weights are standardized to ensure that the weights of all data at the same moment sum to 1. Through the above calculations, the time weight of each data point is obtained, forming a comprehensive time weight vector.

[0029] Based on the spatial coordinates of each data point in the spatially standardized dataset, an adaptive Gaussian kernel function is used to calculate spatial weights. In data-dense regions, where the spatial correlation between data points is strong, a smaller kernel width should be used to finely characterize local features. In data-sparse regions, a larger kernel width should be used to expand the information coverage.

[0030] In the spatial weight calculation process, for each spatial region, the data point density within that region is statistically analyzed. Data point density is defined as the number of data points per unit area. Based on the data point density, an adaptive kernel width calculation method is adopted, where the kernel width is inversely proportional to the data point density: a reference kernel width of 5 kilometers and a reference data point density of 10 data points per square kilometer are set, with an adjustment index of 0.5. The actual kernel width is calculated by multiplying the reference kernel width by the negative 0.5 power of the ratio of the current region density to the reference density. Based on the adaptive kernel width, a Gaussian kernel function is used to calculate the spatial weights between data points: the geographical distance between two data points is calculated using the Haversine formula based on latitude and longitude coordinates to calculate the spherical distance, and the Gaussian kernel function is used, i.e., with the natural constant e as the base and the negative square of the distance divided by twice the square of the kernel width as the exponent. Finally, by calculating the spatial weights between all pairs of data points, a complete spatial weight matrix is ​​constructed.

[0031] Based on a comprehensive time weight vector, spatial weight matrix, and data source credibility vector, a weighted fusion method is used to fuse multi-source data to obtain a comprehensive credibility score. For logistics attribute values ​​(such as demand and traffic conditions at the target spatiotemporal location), it is necessary to fuse information from multiple surrounding data points.

[0032] In the multi-source data fusion process, all data points within a certain range (e.g., a radius of 20 kilometers) around the target location and within a certain time window (e.g., 2 hours prior) before the target time are selected to form a candidate data point set. For each data point in the candidate data point set, its comprehensive fusion weight is calculated: the time weight of the data point is multiplied by the spatial weight of the target location, and then multiplied by a normalization factor to ensure that the sum of all weights is 1. The normalization factor is calculated as the reciprocal of the sum of the products of the time weights and spatial weights of all candidate data points. Finally, the calculated comprehensive fusion weight is used to perform a weighted average of the attribute values ​​of all candidate data points to obtain the fused attribute value of the target location.

[0033] Based on the comprehensive fusion weights of each candidate data point, a weighted average method is used to calculate the fusion attribute value of the target location. By performing the above fusion calculation on all target spatiotemporal locations, the fused logistics dataset is obtained.

[0034] Furthermore, a dynamic weight adjustment mechanism can be used to replace the fixed weighted fusion method to further improve the adaptability of the fusion. The purpose of the dynamic weight adjustment mechanism is to dynamically optimize the fusion weights based on real-time changes in data quality. Specifically, a data quality monitoring module is established to monitor the data integrity rate, update frequency, and outlier ratio of each data source in real time. Based on the monitoring results, a sliding window statistical method is used to calculate the quality score of each data source within the most recent time window (e.g., the most recent hour).

[0035] Data Preprocessing Instructions: When calculating the quality score, since data integrity rate (percentage, 0%–100%), update frequency (times / hour), and outlier ratio (percentage, 0%–100%) have different dimensions and numerical ranges, data preprocessing is required to ensure the rationality of the weighted calculation: Percentage data standardization: Convert the data integrity rate and outlier ratio from percentage form to decimal form, i.e., divide by 100, so that their value range is [0,1]; Update frequency normalization: Convert the update frequency to a standardized value using the maximum and minimum normalization methods, where the minimum and maximum update frequency thresholds are the minimum and maximum update frequency thresholds set by the system, respectively; Outlier ratio inverse processing: Since a lower outlier ratio indicates better quality, inverse processing is required.

[0036] The quality score is calculated by weighted summation, taking into account data completeness, update timeliness, and the degree of data anomalies. Based on the real-time quality score, the data source credibility vector is dynamically adjusted, reducing the credibility of data sources with declining quality and increasing the credibility of data sources with improving quality. This achieves dynamic optimization of the fusion weights and improves the adaptability of the fusion result to changes in data quality.

[0037] Step 2: Construct a logistics network graph based on the fused logistics dataset, predict the status of each node in the logistics network graph, identify bottleneck nodes and critical paths, and obtain a logistics network status assessment report. The purpose of this step is to assess the status of each node and path in the logistics network and identify potential bottlenecks and risk points.

[0038] The process involves acquiring the enterprise's logistics network topology information and constructing a logistics network graph based on the fused logistics dataset and the enterprise's logistics network topology. The node set in the logistics network graph includes warehousing center nodes, forward warehouse nodes, distribution station nodes, and major customer nodes, each represented by a unique identifier. The edge set in the logistics network graph represents the transportation routes and logistics relationships between nodes. For node pairs with direct transportation connections, a directed edge is added to the graph. Based on historical transportation records and actual business rules, attribute information is assigned to each edge, including transportation mode (sea / air / land), average transportation time, transportation cost, path length, etc. Through the above construction process, a logistics network graph containing the node set, edge set, and their attribute information is obtained.

[0039] Based on the fused logistics dataset, feature vectors are extracted for each node in the logistics network graph. For warehouse center nodes, extracted features include current total inventory, inventory turnover rate, average daily outbound volume, average daily inbound volume, storage capacity, and current load rate. For delivery station nodes, extracted features include average daily order processing volume, delivery radius, number of delivery vehicles, average delivery time, and on-time delivery rate. For customer nodes, extracted features include historical order frequency, average order amount, geographical remoteness, and delivery difficulty level. Specifically, for each node, all data records related to that node are queried from the fused logistics dataset, and statistical calculations are performed on these records to obtain various feature values. The current total inventory is obtained by summing the inventory quantities of all goods in the warehouse; the inventory turnover rate is obtained by calculating the ratio of the total outbound volume to the average inventory volume over the past 30 days; and the current load rate is obtained by calculating the ratio of the current total inventory to the storage capacity. For nodes with missing features, the average value of nodes of the same type is used to fill in the gaps. Based on the extracted feature values, feature vectors for each node are constructed. The feature dimension for warehouse center nodes is 10, for delivery station nodes it is 8, and for customer nodes it is 6. Zero-padding is applied to the feature vectors of different node types to unify the feature vector dimension of all nodes to 10. The feature vectors of all nodes are then arranged in order of node number to obtain the node feature matrix.

[0040] Based on the logistics network graph and node feature matrix, a graph convolutional neural network is used for semi-supervised learning. In practical applications, only a few key nodes (such as major warehousing centers) are equipped with comprehensive monitoring equipment to obtain the true state labels of the nodes, while most nodes lack state labels. Based on the adjacency matrix of the logistics network graph, a symmetric normalization method is used to calculate the normalized adjacency matrix. An identity matrix is ​​added to the original adjacency matrix, and self-connections are added to each node to ensure that the node can retain its own feature information. The degree of each node (i.e., the number of edges connected to the node, including self-connections) is counted to construct a degree matrix. The degree matrix is ​​a diagonal matrix, and the diagonal elements are the degrees of the corresponding nodes. Then, the negative 1 / 2 power of each diagonal element of the degree matrix is ​​calculated to obtain the normalization coefficient matrix. Finally, the enhanced adjacency matrix is ​​left-multiplied and right-multiplied using the normalization coefficient matrix to obtain the symmetric normalized adjacency matrix. This normalization method can effectively avoid the numerical instability problem in the graph convolution process. Based on the normalized adjacency matrix and the node feature matrix, a two-layer graph convolutional neural network is constructed. The first-layer graph convolution multiplies the node feature matrix with its weight matrix to perform feature transformation. The dimension of the weight matrix is ​​the node feature dimension multiplied by the hidden layer dimension (set to 64). A normalized adjacency matrix is ​​then used to left-multiply the transformed features, aggregating neighborhood information so that each node can incorporate the feature information of its neighbors. Finally, the ReLU activation function is applied to the aggregated features, introducing a non-linear transformation to enhance the model's expressive power. The first-layer graph convolution aggregates the first-order neighbor information of each node to obtain a node representation that incorporates local neighborhood features. The second-layer graph convolution multiplies the output of the first-layer graph convolution with its weight matrix to perform feature transformation. The dimension of the weight matrix is ​​the hidden layer dimension multiplied by the number of node state categories (set to 4, including normal, light congestion, moderate congestion, and heavy congestion). A normalized adjacency matrix is ​​then used to left-multiply the transformed features to further aggregate second-order neighbor information, resulting in a node representation that incorporates features from a wider range of neighborhoods. Finally, a Softmax function is applied to the aggregated features to transform the feature vector of each node into a probability distribution of state categories, ensuring that the sum of the probabilities of all categories is 1. The Softmax output of the second-layer graph convolution is used to obtain the probability distribution of each node belonging to each state category. The category with the highest probability is selected as the predicted state of the node, yielding the node state estimation result. Based on the true states and labels of a small number of labeled nodes, a cross-entropy loss function is used for model training. For each labeled node, the cross-entropy between its true label and the predicted probability distribution is calculated. The cross-entropy losses of all labeled nodes are summed to obtain the total loss. By minimizing the loss function, optimization algorithms such as gradient descent are used to update the weight parameters of the graph convolutional network. Due to the information propagation characteristics of graph convolution, the state information of labeled nodes will propagate to unlabeled nodes through the network topology, thus achieving semi-supervised learning.

[0041] Based on node state estimation results and historical operational data, a self-supervised learning method is employed to further optimize the state assessment model. The purpose of self-supervised learning is to leverage the inherent structure and temporal characteristics of the data itself to improve model performance without additional labeling. A temporal prediction agent task is constructed to predict the load changes of nodes in future moments using their historical features and states. For each node, its feature sequence and state estimation sequence over a continuous time period are extracted as input to predict its load change at the next moment. The load change is defined as the difference between the load rate at the next moment and the load rate at the current moment. Based on the node representation extracted by the graph convolutional neural network, a temporal prediction branch is added. A fully connected layer maps the node representation to the predicted load change value. The weight vector of the prediction layer is used to perform a dot product with the node representation vector, and a bias term is added to obtain the predicted load change value. Based on the calculated actual load change and predicted value from historical data, a self-supervised loss function is constructed for model optimization. The self-supervised loss function is the sum of squares of the differences between the actual load change and the predicted load change for all nodes. A joint loss function is constructed by combining the self-supervised loss and the semi-supervised loss, where the weight coefficient of the self-supervised loss is set to 0.3. By minimizing the joint loss function, the time-series information of unlabeled data is fully utilized while using labeled data, resulting in an optimized node state evaluation result.

[0042] Based on the optimized node status assessment results and the logistics network diagram, graph analysis algorithms are used to identify bottleneck nodes and critical paths in the network. For bottleneck node identification, a multi-index comprehensive scoring method is employed. The centrality index for each node is calculated, including degree centrality, betweenness centrality, and proximity centrality. Degree centrality measures the number of direct connections a node has, betweenness centrality measures the importance of a node as a transit hub in the network, and proximity centrality measures the ease with which a node can reach other nodes. For each node, its betweenness centrality is obtained by calculating the frequency of its occurrence in the shortest paths between all node pairs. Nodes with high betweenness centrality will have a significant impact on the entire network if they experience failures or congestion. Based on the node's centrality index and status assessment results, a bottleneck risk index is calculated. This index comprehensively considers the normalized betweenness centrality, the node's congestion score (derived from the status category: 0 for normal, 0.33 for mild congestion, 0.67 for moderate congestion, and 1 for severe congestion), and the node's load rate, with weighting coefficients of 0.3, 0.4, and 0.3, respectively. When calculating the bottleneck risk index, data preprocessing is required to ensure the rationality of the weighted calculation because betweenness centrality, congestion score, and load rate have different dimensions and numerical ranges. Betweenness centrality is transformed using maximum and minimum normalization methods to make its value range 0 to 1. Load rate is transformed using Z-score normalization and mapped to the 0-1 range using the sigmoid function. Congestion score has been converted from classification status to a 0-1 range using ordered encoding. The bottleneck risk index comprehensively considers the network importance of nodes, current congestion status, and load level; nodes with high risk indices are identified as bottleneck nodes. For critical path identification, path flow analysis is used. Based on historical transportation records, the cargo flow and usage frequency of each edge are statistically analyzed to identify high-flow edges. For paths containing multiple high-flow edges, the comprehensive flow value of the path is calculated, and the top 10% of paths with the highest flow values ​​are identified as critical paths. Based on the bottleneck node set, critical path set, and the status assessment results of all nodes, a logistics network status assessment report is generated. The logistics network status assessment report includes an overall network health score, status categories and risk indices for each node, a list of bottleneck nodes and their early warning information, congestion status of critical paths, and network capacity margin analysis.

[0043] Furthermore, dynamic graph neural networks can be used instead of static graph convolutional neural networks to better capture the temporal evolution characteristics of the logistics network. The purpose of dynamic graph neural networks is to model the changes in network states over time and improve the accuracy of state prediction. Specifically, the logistics network is modeled as a temporal graph sequence, with each time step corresponding to a network snapshot containing the features and states of each node at that time. A graph recurrent neural network structure is adopted, adding a recurrent neural network layer after the graph convolutional layer, with the recurrent neural network layer using a gated recurrent unit structure. The graph convolutional layer is responsible for aggregating neighborhood information in the spatial dimension, while the recurrent neural network layer is responsible for aggregating historical information in the temporal dimension. By using the hidden states of nodes from previous time steps as input to the current time step, the temporal information is effectively transferred. During training, a temporal prediction task is used, using the network states of the previous T time steps to predict the state at time step T+1. The model parameters are optimized through backpropagation, thereby improving the ability to model the temporal evolution of network states.

[0044] Step 3: Extract order time-series features based on the logistics network status assessment report and the fused logistics dataset, construct an external factor feature matrix, predict logistics demand, and obtain logistics demand prediction results; The purpose of this step is to predict the logistics demand and spatiotemporal distribution characteristics of various regions in the future, so as to provide a basis for resource scheduling and route planning.

[0045] Based on historical order data in the fused logistics dataset, time-series features of order volume are extracted. The historical order data is aggregated by time granularity (e.g., hourly) and spatial granularity (e.g., delivery area) to obtain time series of order volume for each region in each time period. The order volume time series contains an order volume vector from the first time point to the Tth time point, and each element in the vector at each time point corresponds to the order volume of a region.

[0046] Based on the order volume time series, a time series decomposition method is used to divide the sequence into three parts: trend component, periodic component, and residual component. The trend component is extracted using the moving average method, with a calculation time window of 30 days. The periodic component is obtained through periodic detection and extraction methods, detecting the existence of daily, weekly, and monthly periods. For detected periods, Fourier transform is used to extract the corresponding periodic component. The residual component is the remainder after subtracting the trend component and the periodic component from the original sequence.

[0047] Based on the three components obtained from the decomposition, multi-scale time features are constructed. Daily-scale features are extracted, including the number of hours in a day, whether it is working hours, and whether it is a peak meal period; weekly-scale features are extracted, including the day of the week and whether it is a weekend; and monthly-scale features are extracted, including the month and whether it is the beginning or end of the month. The trend component, periodic component, residual component, and multi-scale time features are concatenated to obtain the order time-series feature vector. The order time-series feature vector is a B-row, 20-column matrix, where B represents the number of time steps and 20 represents the dimension of the time-series features.

[0048] An external factor feature matrix is ​​constructed based on promotional activity data, holiday data, weather data, and social event data. Promotional activity data includes information such as promotion time, promotion intensity, and participating product categories; holiday data includes information such as holiday type and number of holiday days; weather data includes information such as temperature, rainfall, and extreme weather warnings; and social event data includes information such as major events and traffic control.

[0049] For continuous factors (such as temperature, rainfall, and promotional discounts), numerical normalization encoding is used to scale the values ​​to between 0 and 1. The normalization method uses min-max normalization, and the specific steps are as follows: Historical data statistics: By analyzing historical data, the minimum and maximum values ​​of each continuous factor are calculated, and a table of characteristic value ranges is established; Difference calculation: For the original feature value to be processed, calculate the difference between it and the minimum value of that feature; Range calculation: Calculate the difference between the maximum and minimum values ​​of the feature, and use it as the denominator for normalization; Normalization transformation: Divide the difference obtained in step 2 by the range value obtained in step 3 to obtain the normalized feature value between 0 and 1.

[0050] The normalized eigenvalues ​​represent the relative positions of the original eigenvalues ​​within their historical value range. The minimum and maximum values ​​obtained through historical data statistics ensure the stability of the normalization.

[0051] For discrete factors (such as holiday type, weather type, and promotional activity type), embedding vector encoding is used. A unique integer index is assigned to each discrete category, and the embedding layer maps these integer indices to dense real-valued vectors. The embedding vectors have a dimension of 8, and the parameters of the embedding layer are learned through end-to-end training.

[0052] Based on the normalized continuous factors and the discrete factors encoded by the embedding vectors, an external factor embedding matrix is ​​constructed. This matrix is ​​B rows and 30 columns, where B represents the time step and 30 represents the external factor feature dimension. Each row of the matrix corresponds to the external factor feature vector at a given time step, containing the encoded values ​​of all external factors at that time step.

[0053] Based on the external factor embedding matrix, a deep factor interaction network is constructed to capture the nonlinear coupling relationships between external factors. The interaction between factors often has a greater impact on demand than the independent effects of each factor, such as the synergistic effect of promotional activities combined with holidays, or the coupled impact of rainfall and traffic congestion.

[0054] Specifically, a multi-layer feature cross structure is used to extract factor interaction features. The first layer performs second-order interaction, calculating the interaction features between any two factors. For the feature vectors of the first and second factors, their second-order interaction features are calculated through the following steps: Element-wise multiplication calculation: Perform element-wise multiplication (Hadamard product) on the eigenvectors of the first and second factors to obtain the element-wise product vector of the two factors; Linear transformation: The product vector is linearly transformed through the weight matrix to capture the linear combination relationship of interactive features; Bias addition: Add a bias vector to the linear transformation result to enhance the expressive power of the model; Interactive feature output: Obtain the second-order interactive feature vectors of the first and second factors.

[0055] Element-wise multiplication can effectively capture the multiplicative interactions between features. By concatenating the second-order interaction features of all factor pairs, a second-order interaction feature matrix is ​​obtained.

[0056] Based on the second-order interaction feature matrix and the original factor embedding matrix, the second layer performs third-order interactions to capture the higher-order coupling relationships between the three factors. The third-order interaction features are calculated through the following steps: Feature concatenation: The second-order interactive feature matrix is ​​concatenated with the original factor feature matrix to form a comprehensive feature representation; Weight transformation: The concatenated features are linearly transformed using a weight matrix to learn higher-order interaction patterns; Bias adjustment: Add bias vectors to adjust parameters and improve the model's fitting ability; Third-order feature output: Obtains a third-order interactive feature matrix containing high-order coupling relationships between the three factors.

[0057] Based on features obtained from multi-layered interactions, an attention pooling mechanism is used to aggregate the contributions of different interaction features. The attention weights are adaptively calculated according to the importance of the features. The calculation method for the attention weights is as follows: Attention score calculation: For the k-th interaction feature vector, the feature vector is mapped to the attention space through a linear transformation, which uses a weight matrix and a bias vector for the transformation; Nonlinear activation: Applying the hyperbolic tangent activation function to the linear transformation result restricts the value range to the (-1,1) interval, thereby enhancing the expressive power; Rating vectorization: Multidimensional features are compressed into scalar scores by linearly combining attention vectors, reflecting the importance of the feature; Weight normalization: The scores of all interaction features are normalized by applying the softmax function to ensure that the sum of all attention weights is 1, thus obtaining the final attention weight of the k-th interaction feature.

[0058] The specific implementation of the function is as follows: an attention scoring function is used to calculate the importance score of the k-th interactive feature; the feature vector is mapped to the attention space through a linear transformation, in which a weight matrix and a bias vector are used for the transformation; a hyperbolic tangent activation function is applied for a nonlinear transformation to restrict the value range to the interval between negative one and positive one; finally, a scalar score is obtained by linearly combining the vectors; this score reflects the importance of the feature to the final prediction result.

[0059] Based on attention weights, a weighted summation is used to obtain the factor interaction feature vector. The specific steps are as follows: Weighting application: Multiply each interaction feature vector by its corresponding attention weight to obtain a weighted feature vector; Feature aggregation: Summing all weighted feature vectors to achieve effective fusion of different interactive features; Interaction feature output: The final factor interaction feature vector is obtained by comprehensively considering the interaction relationships of all factors.

[0060] Through the above multi-layer interaction and attention aggregation, a factor interaction feature vector that fully captures the nonlinear coupling relationship between factors is obtained, where the interaction feature dimension is 64.

[0061] Based on order time-series feature vectors and factor interaction feature vectors, a multi-head attention mechanism is used for feature fusion, enabling the model to dynamically focus on important information from different time points and different factors.

[0062] Specifically, order time-series features are used as queries, and factor interaction features are used as keys and values. An attention mechanism is used to calculate the correlation between time-series features and factor features. For a time-series feature at time t, the steps for calculating its attention score for factor features at all times are as follows: The temporal features at time t and the factor features at time t1 are linearly transformed using the corresponding projection matrices to obtain the query vector and the key vector, respectively. The dot product of the query vector and the key vector is calculated to obtain the original attention score. The original score is divided by the square root of the attention projection dimension to avoid the gradient vanishing problem caused by the attention score being too large. The final attention score of the temporal features at time t to the factor features at time t1 is obtained.

[0063] The dimension of the attention projection is 32, and t1 represents the time index of the factor feature. Dividing by the square root is to avoid the gradient vanishing effect caused by excessively large attention scores.

[0064] Based on the attention score, the attention weight is calculated using the softmax function. This weight represents the attention weight from time t to time t1, and is obtained through exponential function and normalization.

[0065] Based on attention weights and factor eigenvalues, the fused feature vector is calculated. This vector is the fused feature vector at time t, obtained by weighting and summing the factor features at all times using attention weights, with the projection matrix of the mean value in the attention mechanism used for transformation.

[0066] To enhance the model's expressive power, a multi-head attention mechanism is adopted, using eight attention heads to compute in parallel. Each head learns a different attention pattern. Finally, the outputs of each head are concatenated and projected through a linear layer to obtain a fused feature sequence. This sequence is a matrix of T rows and 128 columns, where T represents the time step and 128 represents the fused feature dimension.

[0067] Based on fused feature vectors, a sequence-to-sequence (Seq2Seq) model is used to predict the logistics demand in various regions over future time periods. The Seq2Seq model consists of two parts: an encoder and a decoder. The encoder extracts historical information, and the decoder generates future predictions.

[0068] The encoder employs a bidirectional long short-term memory (BiLSTM) network to extract temporal features from both forward and backward directions. For the input sequence, the sequence contains a fused feature vector from the first time step to the a-th time step, where a represents the input time step, with a value of 168 (representing hourly data from the past 7 days). The hidden states of the forward LSTM and backward LSTM are calculated using the forward and backward LSTM networks, respectively.

[0069] The forward and backward hidden states are concatenated to obtain the output of the bidirectional encoder, which contains bidirectional temporal information at time t.

[0070] The encoder's final hidden state, as a comprehensive representation of historical information, is passed to the decoder.

[0071] The decoder uses a unidirectional LSTM to progressively generate demand predictions for future time steps. At each time step t1, the decoder receives the prediction output from the previous time step and the external factor features of the current time step as input, updates the hidden state, and generates the prediction for the current time step. The hidden state represents the decoder's internal state at time t1.

[0072] The decoder's initial hidden state is initialized by the encoder's final hidden state.

[0073] An attention mechanism is introduced during the decoding process, enabling the decoder to dynamically monitor historical information from different encoder times. For decoder time t1, the attention weights from all encoder times are calculated using softmax normalization; these weights represent the attention weights from decoder time t1 to encoder time t.

[0074] Based on the attention weights, the context vector is calculated. This vector is the context vector of the decoder at time t1, which is obtained by weighting and summing the hidden states of the encoder at all times using the attention weights.

[0075] The context vector is concatenated with the decoder's hidden state, and the predicted logistics demand for each region at the current moment is generated through a fully connected layer. The predicted logistics demand is a vector, and each element of the vector corresponds to the predicted order volume for a region. It is calculated using the weight matrix and bias vector of the output layer of the sequence-to-sequence model.

[0076] By decoding step by step, demand forecasts for the next T time steps (e.g., the next 72 hours) are generated, resulting in a logistics demand forecasting result. This result contains a vector of predicted order volumes from the first time step to the Tth time step.

[0077] To improve prediction accuracy, the model training employs a combination of teacher-mandated and free-running strategies. Initially, a teacher-mandated strategy is used, where the decoder uses real historical values ​​as input at each time step. Later in training, the model gradually transitions to a free-running strategy, using its own predictions as input, thereby enhancing its generalization ability in real-world applications.

[0078] Furthermore, probabilistic prediction can be used instead of point prediction to quantify the uncertainty of prediction. The purpose of probabilistic prediction is not only to predict the expected value of demand, but also the probability distribution of demand, providing a basis for robust decision-making. Specifically, in the output layer of the decoder, not only the mean of demand is predicted, but also the standard deviation of demand. The output layer contains two parallel fully connected branches, one branch outputs the mean, and the other branch outputs the logarithm of the standard deviation (using logarithms ensures that the standard deviation is positive). During training, a negative log-likelihood loss function is used, assuming that demand follows a normal distribution. The loss function is calculated by dividing the squared difference between the true demand value and the predicted mean by twice the predicted variance, and then adding the logarithm of the predicted standard deviation.

[0079] By minimizing the negative log-likelihood, the model learns to predict the mean and uncertainty of demand. During prediction, it outputs a probability distribution of demand, which is a normal distribution with the mean of the predicted demand and the variance being the square of the standard deviation of the predicted demand.

[0080] The predicted range of demand can be obtained by sampling or calculating confidence intervals, thereby supporting subsequent robust resource scheduling.

[0081] Step 4: Based on the logistics demand forecast results, construct a set of delivery tasks, predict the travel time of road segments, optimize delivery routes and task allocation, and obtain a dynamic delivery route plan; The purpose of this step is to generate an optimal delivery route plan that takes into account real-time constraints, based on demand forecasting results and network status assessment.

[0082] Based on logistics demand forecasting results, logistics network status assessment reports, and current pending order information, a set of delivery tasks is constructed. All pending orders are retrieved from the order management system, and each order includes attributes such as order number, product information, weight, volume, delivery address, and expected delivery time.

[0083] Based on the order's delivery address and warehouse location, a nearest warehouse matching method is used to assign a shipping warehouse to each order. For each order, the straight-line distance and estimated transit time between its delivery address and each warehouse are calculated, and the warehouse with the best overall score is selected as the shipping point. The overall score considers distance, transit time, warehouse inventory availability, and current load.

[0084] Based on order information and warehouse allocation results, a delivery task is constructed. The delivery task includes attributes such as origin (shipping warehouse), destination (delivery address), cargo weight, cargo volume, time window (including earliest delivery time and latest delivery time), and customer priority.

[0085] Organize all pending delivery tasks into a delivery task set, which contains all delivery tasks that need to be processed.

[0086] Based on real-time and historical traffic data from the fused logistics dataset, a spatiotemporal prediction model is used to predict the travel time of each road segment in the future. The road network is divided into multiple road segments, and for the travel time of each road segment at a specific time, the Kalman filter method is used to predict the travel time of that road segment based on its historical travel time series and current traffic conditions.

[0087] The state equations and observation equations for Kalman filtering are implemented through the following steps: Based on the state and state transition matrix of the previous moment, the actual travel time state at the current moment is predicted, while considering the influence of process noise; the predicted state is mapped to the observation space through the observation matrix, and observation noise is added to obtain the theoretical observation value; the difference between the theoretical observation value and the actual observation value is compared, and state correction is performed through Kalman gain; the optimal state estimate that integrates the predicted information and the observation information is obtained.

[0088] The state vector represents the actual travel time state at time t, the observations come from real-time traffic data, the state transition matrix describes the state evolution law, the observation matrix establishes the mapping relationship between the state and the observation, and the process noise and observation noise represent the model uncertainty and measurement error, respectively.

[0089] By using the prediction and update steps of Kalman filtering, the predicted travel time for each road segment in the future time period is obtained, forming a spatiotemporal traffic condition prediction matrix.

[0090] Based on customs clearance records in the customs dataset, a probability distribution model of clearance time is established using statistical analysis methods. For each port, the clearance time in different time periods (weekdays / weekends, peak / off-peak periods) is statistically analyzed, and a log-normal distribution is fitted to obtain the distribution parameters (mean and standard deviation). During route planning, the corresponding clearance time distribution is queried based on the expected arrival time at the port and the port number, and the expected value is used as the clearance time estimate.

[0091] Based on vehicle status data, obtain information on all available transport vehicles, including their current location, load capacity, remaining capacity, available time period, and transport type (land / sea / air). Organize the available vehicles into an available resource list.

[0092] Based on the scale of the delivery task set and the spatial distribution characteristics of the tasks, a task decomposition strategy is adopted to break down the large-scale path planning problem into multiple sub-problems. Specifically, a spatial clustering method is used, based on the start and end coordinates of the tasks, and the K-means clustering algorithm is used to divide the tasks into K regional clusters, with tasks within each cluster being relatively concentrated geographically. The number of clusters K is determined according to the total number of tasks and the number of available computing nodes on the cloud platform, and is usually set to twice the number of computing nodes to achieve load balancing.

[0093] For each task cluster, a regional subproblem is constructed. The subproblem includes all delivery tasks within the cluster, a subset of vehicles serving the region, the region's road network information, and traffic conditions. Subproblems are coupled with boundary tasks and shared resources. Boundary tasks refer to tasks whose origin and destination belong to different clusters, while shared resources refer to vehicles that can be dispatched across regions.

[0094] To ensure consistency in the solutions to subproblems, a Lagrange relaxation method is employed to handle the coupling constraints between them. For the allocation of boundary tasks and shared resources, Lagrange multipliers are introduced to relax the global consistency constraint into local penalty terms for each subproblem. This results in an independent set of subproblems and an initial Lagrange multiplier vector.

[0095] The distributed computing architecture based on the cloud platform distributes each sub-problem to different computing nodes for parallel solution. Each computing node runs an improved genetic algorithm to solve the sub-problem it is responsible for.

[0096] The improved genetic algorithm process is as follows: An initial population is generated, with each individual representing a path planning scheme. Diverse initial solutions are generated using a combination of random generation and heuristic construction. Individual encoding uses task sequence encoding, where each element in the sequence represents a delivery task, and the sequence order represents the service order of the tasks. The sequence is divided into multiple subsequences based on vehicle capacity constraints, with each subsequence corresponding to a delivery route for one vehicle.

[0097] The fitness function is calculated based on individual encoding. The fitness function comprehensively considers multiple optimization objectives, including transportation cost, delivery time, and time window violation penalties. The fitness value is calculated through the following steps: The total transportation cost is calculated, including fuel cost, vehicle usage cost, customs fees, and other expenses; the total delivery time is calculated, including transportation time, loading and unloading time, and waiting time; the penalty value for time window violation is calculated, and penalties are imposed on tasks that exceed the time window; the transportation cost, time cost, and service quality penalty items are weighted and summed according to weight coefficients of 0.4, 0.4, and 0.2 respectively; the negative value of the weighted sum is taken to obtain the final fitness value, and the higher the fitness value, the better the solution.

[0098] Data preprocessing instructions: When calculating the fitness function, since transportation cost (unit: yuan), delivery time (unit: hours), and penalty value (unit: minutes) have different dimensions and numerical ranges, data preprocessing is required to ensure the rationality of the weighted calculation: Transportation cost is transformed using the maximum and minimum normalization methods, using the minimum and maximum historical costs as normalization boundaries; delivery time is transformed using the maximum and minimum normalization methods, using the theoretical shortest time and the longest allowable time as normalization boundaries; penalty value is transformed using the maximum and minimum normalization methods, using the maximum possible penalty value as the normalization boundary.

[0099] The normalized fitness function is a weighted negative of cost, time, and penalty value, with weighting coefficients of 0.4, 0.4, and 0.2, respectively.

[0100] A higher fitness value indicates a better solution.

[0101] Based on fitness values, a tournament selection strategy is used to select parent individuals for reproduction. Crossover operations are performed using sequential crossover and partially mapped crossover operators to generate offspring individuals. A mutation strategy combining insertion mutation and exchange mutation is employed to mutate offspring individuals, increasing population diversity.

[0102] Through multiple generations of evolution, each computing node obtains a local optimum solution to its subproblem. Each computing node uploads its local solution to the master node, which coordinates the solutions to the subproblems based on the Lagrange multiplier update strategy. The Lagrange multiplier update uses the gradient ascent method with a step size parameter of 0.1. The constraint violation gradient is obtained by calculating the degree of inconsistency between the solutions to the subproblems.

[0103] The specific implementation of the constraint violation gradient is as follows: calculate the allocation difference of each subproblem on the boundary task, quantify the degree of constraint violation by comparing the processing schemes of adjacent subproblems on the same boundary task; sum the violation degrees of each boundary task by weight, with the weights determined according to the importance and urgency of the task; finally, obtain the violation gradient of each constraint by gradient calculation, where the set of adjacent subproblems represents the subproblems adjacent to the current subproblem, and the local solution and the neighbor solution represent the allocation values ​​of the local solution and the neighbor solution on the boundary task, respectively.

[0104] Based on the local optimal solutions of each subproblem and the updated Lagrange multipliers, a global coordination algorithm is used to generate a consistent global path planning scheme.

[0105] For boundary tasks, a competitive bidding mechanism is used to determine the final allocation. Each relevant subproblem submits a service offer for the boundary task, taking into account both the task's impact on the subproblem's objective function and resource utilization efficiency. The offer calculation includes the cost of the subproblem serving the task, the corresponding Lagrange multiplier for the task, and the amount of resources required to serve the task. The subproblem with the lowest offer is selected to undertake the boundary task.

[0106] For shared resources, a resource allocation optimization algorithm is used to determine the optimal allocation scheme. Based on the resource requirements and service capabilities of each sub-problem, a linear programming model for resource allocation is constructed. The objective is to minimize the total resource allocation cost. Constraints include that each resource can be allocated to at most one sub-problem, the resource requirement of each sub-problem does not exceed its demand, and the decision variables are binary variables. By solving the linear programming model, the optimal resource allocation scheme is obtained.

[0107] Based on the results of boundary task allocation and shared resource allocation, the solutions to each subproblem are updated to obtain a globally consistent path planning scheme. The global scheme is then optimized using 2-opt and 3-opt local search algorithms to optimize the delivery routes of each vehicle, reducing the total travel distance and delivery time.

[0108] Based on real-time traffic and emergency information, a dynamic replanning mechanism is employed to adjust route plans in real time. An event monitoring module is established to monitor emergencies such as traffic congestion, road closures, and vehicle breakdowns in real time.

[0109] When a sudden event affecting delivery is detected, the impact of the event on the current route plan is assessed. Impact assessment indicators include the number of vehicles affected, delay time, and additional costs. If the impact exceeds a preset threshold, a dynamic replanning process is triggered.

[0110] Dynamic replanning employs an incremental optimization strategy, replanning only the affected parts of the path while keeping other paths unchanged, thus reducing computation time and scheme changes. For affected vehicles, the optimal path from the current location to the remaining delivery points is recalculated, taking into account real-time traffic conditions and time window constraints.

[0111] The replanning process employs fast heuristic algorithms, including the nearest neighbor algorithm, the saving algorithm, and the insertion algorithm. The nearest neighbor algorithm starts from the current position and selects the next delivery point that is closest to the current location and satisfies the constraints at each step. The saving algorithm reduces the total travel distance by merging paths. The insertion algorithm inserts unserved delivery points into the optimal positions on the existing paths.

[0112] The system receives real-time updated route plans and pushes the adjustment information to relevant vehicles and delivery personnel to ensure timely execution of the route plans.

[0113] Step 5: Based on the logistics demand forecast results and dynamic delivery route plan, conduct supply and demand matching analysis, optimize resource scheduling according to the supply and demand matching analysis results, and obtain a resource pre-scheduling plan; The purpose of this step is to proactively pre-schedule resources based on demand forecasting and network status, ensuring the robustness of resource allocation.

[0114] Based on logistics demand forecasting results, if probabilistic forecasting is used, the mean and standard deviation of demand are directly obtained. If point forecasting is used, uncertainty is estimated based on historical forecasting error statistics. Specifically, the mean absolute percentage error of historical forecasts is calculated, and the standard deviation of demand is estimated by multiplying the mean absolute percentage error by the demand mean. This method can convert the relative magnitude of historical forecasting errors into an absolute measure of uncertainty.

[0115] Based on the demand mean and standard deviation, a confidence interval for demand is constructed. At a 95% confidence level, the upper bound of demand is calculated as follows: multiply the standard deviation by 1.96 (corresponding to the quantile of the standard normal distribution at the 95% confidence level), and add the result to the demand mean. The lower bound of demand is calculated as follows: multiply the standard deviation by 1.96, and subtract the result from the demand mean. To conservatively plan resources, the upper bound of demand is used as the basis for resource allocation.

[0116] Based on the upper bound of demand and the current resource distribution, a supply-demand matching analysis is performed. For each region and time period, the upper bound of demand is compared with the current available resource capacity. When the upper bound of demand is greater than the current available resource capacity, there is a potential resource shortage in that region during that time period. The shortage is calculated by subtracting the current available resource capacity from the upper bound of demand; the positive value obtained is the resource shortage.

[0117] When the upper bound of demand is less than the current available resource capacity, there is a resource surplus in the region. The surplus is calculated by subtracting the upper bound of demand from the current available resource capacity, and the positive value obtained is the resource surplus.

[0118] Based on shortages and surpluses, identify the set of source regions (surplus regions) and target regions (shortage regions) for the transfer of demand resources.

[0119] Based on the source region set, target region set, and logistics network status assessment report, calculate the transfer cost of resources from the source region to the target region.

[0120] Resource transfer costs consist of three parts: transportation costs, time costs, and opportunity costs. Transportation costs are calculated based on the distance between the source and target regions and the mode of transportation. For the unit resource transportation cost between regions, the optimal route and estimated transportation costs are obtained from the route planning module. Time costs are calculated based on the time required for resource transfer and the time value of time. The transfer time is obtained from the route planning results, and the time cost is calculated by multiplying the time value coefficient (set at 50 yuan per hour) by the transfer time. Opportunity costs represent the losses incurred when the source region's future demand cannot be met after the resource transfer. Calculated based on the uncertainty of future demand forecasts in the source region and the scarcity of resources, opportunity costs are calculated by multiplying the opportunity cost coefficient (set at 0.1) by the source region's demand standard deviation and the amount of resource surplus.

[0121] Based on the three-part cost structure, the total transfer cost is calculated as follows: multiply the quantity of resources transferred by the unit transportation cost and the time cost respectively, add the two products together, and finally add the opportunity cost to obtain the total transfer cost. For all source and target region pairs, the transfer cost is calculated to form a resource transfer cost matrix.

[0122] Based on the resource transfer cost matrix, the uncertainty information of demand forecasting, and the supply and demand matching analysis results, a robust resource scheduling optimization model is constructed.

[0123] A robust optimization objective function is constructed based on the worst-case demand scenario. The optimization objective is to minimize the total cost and maximize the demand satisfaction rate in the worst-case scenario (demand upper bound). The decision variable is the resource transfer amount, representing the amount of resources transferred from the source region to the target region. The steps for constructing the objective function are as follows: calculate the total resource transfer cost from all source regions to the target region, i.e., sum the transfer costs for all source and target region pairs; calculate the penalty cost for unmet demand, i.e., for each target region, calculate the positive value of the demand gap and multiply it by a penalty coefficient (valued at 200 yuan / unit), where the demand gap is the upper bound of demand minus the resource supply after scheduling, and the resource supply after scheduling equals the current available resource capacity plus the total amount of resources transferred into that region; finally, the transfer cost and penalty cost are added together to obtain the overall objective function. The purpose of calculating the positive value of the demand gap in the function is: when the supply is insufficient to meet the upper bound of demand, return the gap amount; when the supply is sufficient, return 0, ensuring that only actual resource shortages are penalized. The first term represents the total cost of resource transfer, and the second term represents the penalty for unmet demand. The trade-off between cost and service quality is achieved by minimizing the sum of the two terms.

[0124] The constraints include: resource availability constraint, meaning the total transfer out of the source region cannot exceed its resource surplus; resource transfer time constraint, meaning the resource transfer time must be within the available transfer time window; and non-negativity constraint, meaning all decision variables must be non-negative. The first constraint ensures that the transfer out of the source region does not exceed its surplus, the second constraint ensures that resources can arrive before demand occurs, and the third constraint ensures that the decision variables are non-negative.

[0125] Based on the above optimization model, a linear programming solver is used to solve it. Since the constraints and objective function in the model are linear, the simplex method can be used for efficient solution. The solution yields the optimal resource transfer scheme, which includes the optimal resource transfer amount between each pair of source and target regions. Based on the optimal transfer scheme, a robust resource pre-scheduling scheme is generated, which details which resources need to be transferred from where and where, the transfer time, and the route.

[0126] Based on the actual execution of the global dynamic delivery route plan, a real-time monitoring system is used to obtain actual data during the execution process, including actual order arrivals, actual delivery completion times, and actual resource consumption.

[0127] The demand deviation is calculated based on actual order arrivals and forecasted demand: the average forecasted demand is subtracted from the actual order arrivals; the difference is the demand deviation. For each region and time period, the demand deviation value is recorded. A sliding window statistical method is used to calculate the statistical characteristics of the deviation within the most recent time window (e.g., the last 7 days), including the mean deviation, standard deviation, and systematic trend of the deviation.

[0128] Based on the statistical characteristics of the deviation, diagnose whether the prediction model has systematic bias. If the mean deviation deviates significantly from zero (as determined by a t-test), it indicates that the prediction model has a systematic bias of overestimating or underestimating, and the model parameters need to be adjusted. The adjustment method is to add a deviation correction term to the output layer of the prediction model. The value of the correction term is equal to the negative of the mean deviation, and the corrected mean demand prediction is equal to the original mean prediction minus the mean deviation.

[0129] The uncertainty estimate is dynamically adjusted based on changes in the standard deviation of the deviation. If the standard deviation of the actual deviation is greater than the standard deviation of the forecast, it indicates that the model's estimate of uncertainty is insufficient, and the uncertainty parameter needs to be increased. The adjustment method is to select the larger of the original demand standard deviation and the actual deviation standard deviation as the adjusted demand standard deviation. The purpose of selecting the larger of the two standard deviations in the function is: if the standard deviation of the actual deviation is greater than the original forecast standard deviation, it indicates that the actual uncertainty is underestimated, and a larger standard deviation value needs to be used; if the original forecast standard deviation is already large enough, the original value is maintained to ensure the conservatism and accuracy of the uncertainty estimate.

[0130] Based on the updated demand forecast mean and uncertainty estimate, the process is repeated to generate an adjusted resource scheduling plan. The adjusted plan more accurately reflects actual demand characteristics and improves the effectiveness of resource allocation. This results in an adjusted robust resource scheduling plan.

[0131] Furthermore, the safety stock strategy from inventory theory can be used to replace the resource reservation method based on confidence intervals to manage demand uncertainty more systematically. The purpose of the safety stock strategy is to add a certain safety margin to the expected demand, the size of which is determined based on demand uncertainty and the target service level. Specifically, for the target service level (e.g., a 95% probability of demand satisfaction), the corresponding z-value is obtained by consulting the standard normal distribution table (z=1.65 for a 95% service level). Based on the z-value and the demand standard deviation, the safety stock is calculated by multiplying the z-value by the demand standard deviation. The resource allocation is set as the expected demand plus the safety stock. By adjusting the target service level, a trade-off can be struck between resource cost and service quality; a higher service level requires a larger safety stock and higher resource costs, but the risk of unmet demand is lower.

[0132] Step 6: Identify key scenarios based on dynamic delivery route schemes and resource pre-scheduling schemes, simulate the execution process, calculate the expected performance and risk indicators of the decision-making scheme, and generate an intelligent decision support report; The purpose of this step is to provide managers with multi-dimensional decision support information through scenario simulation and risk assessment.

[0133] Based on historical logistics data, scenario feature vectors are constructed to describe the logistics operation status at different times. The feature vectors include multi-dimensional information such as demand distribution characteristics (demand volume, demand volatility, and demand correlation in each region), supply capacity characteristics (resource availability, resource utilization rate, and resource allocation flexibility in each region), and external environmental characteristics (weather conditions, traffic conditions, impact of holidays, and impact of emergencies).

[0134] For continuous features, Z-score standardization is used for data preprocessing. The standardization method involves subtracting the mean of the feature from its original value and dividing by its standard deviation. For discrete features, one-hot encoding is used to convert each category into a binary vector representation.

[0135] Based on the standardized feature vectors, the K-means clustering algorithm is used to identify typical scenario categories. The clustering process includes: initializing cluster centers, calculating the Euclidean distance from each sample to each cluster center, assigning samples to the nearest cluster center, updating the cluster centers to the mean of all samples within that category, and repeating the iteration until the cluster centers converge. The optimal number of clusters is determined using the elbow rule, selecting the number of clusters that slows down the decrease in the sum of squares within each cluster.

[0136] A set of typical scenario categories is obtained, with each category representing a logistics operation state with similar characteristics. For each category, its characteristic statistics are calculated, including the mean, standard deviation, and distribution range of each dimension, forming a characteristic profile of the scenario category.

[0137] Based on the real-time status of the current logistics system, a current state feature vector is extracted. The feature extraction method extracts multi-dimensional information including current demand distribution, supply capacity, and external environment. Numerical features are Z-score standardized, and categorical features are one-hot encoded to ensure that the data format remains consistent with historical scenario feature vectors.

[0138] Calculate the similarity between the current state and each typical scenario category. The similarity calculation uses the cosine similarity method. The steps are as follows: calculate the dot product of the current state feature vector and the typical scenario feature vector; calculate the magnitude of each vector (i.e., the square root of the sum of the squares of the elements in each vector); divide the dot product by the product of the two magnitudes to obtain the cosine similarity value. The closer the cosine similarity value is to 1, the more similar the two vectors are; the closer it is to 0, the lower the similarity.

[0139] Based on the similarity calculation results, scenario categories with similarity exceeding a preset threshold (e.g., 0.7) are selected as key scenarios. For the selected key scenarios, probability estimation is performed based on the similarity values. The probability estimation method is as follows: the similarity value of each key scenario is normalized, that is, the similarity value of each scenario is divided by the sum of the similarity values ​​of all key scenarios to obtain the probability of occurrence of each scenario. This ensures that the sum of the probabilities of all key scenarios equals 1. The resulting set of key scenarios and their corresponding probabilities provides a foundation for subsequent scenario sample generation and parallel simulation.

[0140] Based on the feature distribution and occurrence probability of key scenarios, a Monte Carlo method is used to generate scenario samples. For each key scenario, the number of samples generated is determined according to its occurrence probability. During sample generation, for continuous features, random samples are generated in the standardized space using Gaussian sampling, and then transformed back to the original feature space through inverse standardization (multiplying the standardized value by the original standard deviation and adding the original mean). For discrete features, multinomial sampling is used to generate samples based on the historical frequency of each category, and then converted into one-hot encoded vectors.

[0141] For each generated scenario sample, a corresponding logistics system simulation model is constructed. The simulation model includes a demand generation module (generating demand sequences for each region based on scenario characteristics), a resource scheduling module (generating resource scheduling schemes based on robust optimization algorithms), and an execution simulation module (simulating the execution process of the resource scheduling scheme and calculating indicators such as delivery time, cost, and service quality).

[0142] Parallel computing technology is employed to run simulation models of multiple scenario samples simultaneously, thereby improving computational efficiency. During parallel simulation, each computing node is responsible for the simulation calculation of one or more scenario samples. The simulation results include key performance indicators such as total cost, service level, resource utilization, and risk metrics.

[0143] The simulation results of all scenario samples are collected to form a scenario simulation results dataset, which provides data support for the comprehensive evaluation of decision-making schemes.

[0144] Based on the simulation results and the probability of occurrence of each scenario, a weighted evaluation method is used to calculate the overall performance of the decision-making scheme. For performance indicators (such as transportation costs and delivery time), the expected value is calculated as follows: the probability of occurrence of each key scenario category is multiplied by the average performance indicator of all samples in that category, and the weighted values ​​of all key scenario categories are summed to obtain the expected value of the performance indicator. The expected value reflects the average performance of the decision-making scheme under different scenarios.

[0145] The standard deviation of performance metrics measures the volatility and uncertainty of decision outcomes. The standard deviation is calculated as follows: for each scenario sample, calculate the squared difference between its performance metric value and the expected value; then, take a weighted average of the squared differences of all samples according to the probability of the scenario occurring; finally, take the square root to obtain the standard deviation.

[0146] Calculate the worst-case values ​​of performance metrics to assess the resilience of decision-making options. The worst-case value is determined by finding the scenario among all sample scenarios that results in the worst possible performance metric. For metrics that need to be minimized (e.g., cost), the worst-case scenario is the one with the highest metric value; for metrics that need to be maximized (e.g., on-time performance), the worst-case scenario is the one with the lowest metric value. The purpose of the worst-case selection function is to help identify the system's performance under extreme conditions, providing a basis for risk management.

[0147] Based on expected value, standard deviation, and worst-case scenario value, a comprehensive decision support report is generated, which includes decision recommendations, risk warnings, key indicator predictions, and visualization charts.

[0148] Based on the comprehensive evaluation results of the integrated decision support report and simulation data, sensitivity analysis was used to identify the key factors that have the greatest impact on the decision outcome.

[0149] Specifically, for each external factor (such as weather, traffic, promotional intensity, etc.), while keeping other factors at baseline, the value of this factor is changed, and the changes in performance indicators are observed. The correlation coefficient or elasticity coefficient between the factor change and the performance change is calculated. The elasticity coefficient is calculated as follows: calculate the percentage change in performance indicator, calculate the percentage change in external factor, and divide the percentage change in performance by the percentage change in factor to obtain the elasticity coefficient.

[0150] The larger the absolute value of the elasticity coefficient, the more significant the impact of that factor on performance. Based on the elasticity coefficient, key influencing factors are identified, namely the top 5 factors with the largest absolute values ​​of the elasticity coefficient.

[0151] Based on the comprehensive evaluation results of decision-making options, risk analysis results, and sensitivity analysis results, an intelligent decision support report is generated using visualization technology. The report adopts a modular structure and includes the following modules: The decision-making overview module displays the key points of recommended delivery route and resource scheduling plans, including the number of routes, resource allocation direction, estimated costs, and delivery time.

[0152] The performance expectation module uses bar charts and radar charts to display the expected values, confidence intervals, and worst-case values ​​of various performance indicators, allowing for a direct comparison of the performance of different indicators.

[0153] The risk assessment module uses a heatmap to display the performance distribution under different scenarios, with the color intensity indicating the performance level, helping managers identify high-risk scenarios.

[0154] The sensitivity analysis module uses a tornado diagram to display key influencing factors and their degree of impact, sorted by degree of impact, to help managers identify factors that need to be focused on.

[0155] The optimization suggestion module provides targeted optimization suggestions based on simulation results and system analysis, such as improvement measures for bottleneck nodes and emergency plans for high-risk scenarios.

[0156] Receive a visual intelligent decision support report, output in HTML or PDF format, supporting interactive queries and detailed information expansion.

[0157] Furthermore, digital twin technology can be used to replace offline simulation to achieve higher fidelity scenario simulation. The purpose of digital twin technology is to construct a virtual mirror of the logistics system, which remains synchronized with the real system in real time. Specifically, a three-dimensional visualization model of the logistics system is established, containing virtual representations of physical entities such as warehousing facilities, transport vehicles, and delivery routes. Real-time status data of the real system is collected through IoT devices and synchronized to the virtual model, maintaining consistency between the virtual model and the real system. Scenario simulations in the virtual environment allow for testing of different decision-making schemes and response strategies without affecting real operations. Digital twin technology provides an immersive visualization experience, helping managers to more intuitively understand system status and decision-making effectiveness, thereby improving the scientific rigor and credibility of decision-making.

[0158] Cloud-based logistics data analysis and decision-making systems, such as Figure 2 As shown, the cloud-based logistics data analysis and decision-making method described above includes: The multi-source data fusion module is used to acquire multi-source datasets, perform spatiotemporal standardization on data at different time granularities, calculate confidence scores, and use an adaptive weight fusion method to obtain the fused logistics dataset. The network status assessment module constructs a logistics network graph based on the fused logistics dataset, predicts the status of each node in the logistics network graph, identifies bottleneck nodes and critical paths, and obtains a logistics network status assessment report. The demand forecasting module extracts order time-series features and constructs an external factor feature matrix based on the logistics network status assessment report and the fused logistics dataset to predict logistics demand and obtain logistics demand forecasting results. The route planning module constructs a set of delivery tasks based on logistics demand forecasting results, predicts road segment travel time, optimizes delivery routes and task allocation, and obtains dynamic delivery route solutions. The resource scheduling module performs supply and demand matching analysis based on logistics demand forecasting results and dynamic delivery route plans, and optimizes resource scheduling based on the supply and demand matching analysis results to obtain a resource pre-scheduling plan. The decision support module identifies key scenarios based on dynamic delivery route plans and resource pre-scheduling plans, simulates the execution process, calculates the expected performance and risk indicators of the decision-making plan, and generates an intelligent decision support report.

[0159] In one embodiment of the present invention, a specific example is provided: A cross-border e-commerce logistics company used the system of this invention for logistics data analysis and decision optimization during a promotional period. The company has five main warehousing centers in Southeast Asia, located in Singapore, Kuala Lumpur (Malaysia), Bangkok (Thailand), Ho Chi Minh City (Vietnam), and Jakarta (Indonesia), and needs to serve customer orders covering these five countries.

[0160] One week before the promotion, the system collected data from multiple sources. Examples of some order and warehouse data are shown in Table 1: Table 1: Examples of partial order and warehouse data;

[0161] Table 2 shows some examples of traffic and customs data: Table 2: Examples of some traffic and customs data;

[0162] Multi-source data fusion processed approximately 500GB of heterogeneous data. Through an adaptive weighted fusion algorithm, data integrity was improved from 75% to 92%, and data consistency from 68% to 89%. Logistics network status assessment identified three potential bottleneck nodes (Singapore warehousing center, Kuala Lumpur distribution station, and Bangkok forward warehouse) and five critical paths. A semi-supervised learning algorithm achieved 85% status prediction accuracy with only 20% labeled data. The demand forecasting model predicted a 350% increase in order volume during the period, with the most significant growth in Thailand and Vietnam. A deep interactive network successfully captured the synergistic effect of promotional activities and holidays, achieving a prediction accuracy of 78%. Dynamic route planning handled approximately 80,000 delivery tasks, reducing the solution time from 6 hours to 45 minutes through cloud parallel computing. The real-time route adjustment function handled 127 unexpected events during the promotion period, with an average response time of 3.2 minutes. The robust optimization model considered the uncertainty of demand fluctuations of ±30% and formulated a comprehensive solution including resource pre-configuration and dynamic scheduling. Sensitivity analysis showed that appropriately increasing resource pre-configuration by 15% could improve the service level under worst-case conditions from 82% to 94%. The critical scenario identification algorithm identified "high demand + severe weather + customs delays" as the highest-risk scenario, and digital twin simulation verified the effectiveness of the response strategy. The intelligent decision support system provided management with real-time monitoring and early warning capabilities.

[0163] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A cloud-based logistics data analysis and decision-making method, characterized in that, include: We acquire multi-source datasets, perform spatiotemporal standardization on data at different time granularities, calculate confidence scores, and use an adaptive weight fusion method to obtain the fused logistics dataset. Based on the fused logistics dataset, a logistics network graph is constructed, the status of each node in the logistics network graph is predicted, bottleneck nodes and critical paths are identified, and a logistics network status assessment report is obtained. Based on the logistics network status assessment report and the fused logistics dataset, order time-series features are extracted and an external factor feature matrix is ​​constructed to predict logistics demand and obtain logistics demand prediction results. Based on the logistics demand forecasting results, a set of delivery tasks is constructed, the travel time of road segments is predicted, and the delivery routes and task allocation are optimized to obtain a dynamic delivery route scheme. Based on the logistics demand forecast results and dynamic delivery route plans, supply and demand matching analysis is performed, and resource scheduling is optimized according to the supply and demand matching analysis results to obtain a resource pre-scheduling plan. Based on dynamic delivery route schemes and resource pre-scheduling schemes, key scenarios are identified, the execution process is simulated, and the expected performance and risk indicators of the decision-making schemes are calculated to generate intelligent decision support reports.

2. The cloud-based logistics data analysis and decision-making method according to claim 1, characterized in that, The calculation of the credibility score specifically includes: Acquire data from the same spatiotemporal location in a multi-source dataset, calculate the Pearson correlation coefficient between different data sources in the same spatiotemporal location, and evaluate the degree of consistency between data sources based on the correlation coefficient. A basic credibility score is established based on historical accuracy statistics from various data sources; Timeliness weights are calculated based on data update frequency and collection time intervals. The consistency level, basic credibility score, and timeliness weight are weighted and combined to obtain a comprehensive credibility score.

3. The cloud-based logistics data analysis and decision-making method according to claim 1, characterized in that, The state of each node in the predicted logistics network diagram specifically includes: Obtain the adjacency matrix of the logistics network graph, and normalize the adjacency matrix to obtain the normalized adjacency matrix. Based on the fused logistics dataset, features of warehouse center nodes, delivery station nodes, and customer nodes are extracted, and node feature vectors are constructed. Two graph convolutional layers are used to process the node feature vectors. Each layer aggregates the neighborhood node information and updates the node representation through a normalized adjacency matrix. Based on the real state labels of a small number of labeled nodes, a semi-supervised learning method is used to train the updated node representations to predict the state of all nodes.

4. The cloud-based logistics data analysis and decision-making method according to claim 1, characterized in that, The predicted logistics demand for each region specifically includes: Based on the logistics network status assessment report and the fused logistics dataset, an external factor feature matrix is ​​extracted. Element-wise multiplication is performed on any two features in the external factor feature matrix to obtain second-order interactive features. Third-order interaction features are obtained by further cross-referencing the second-order interaction features with the original features. The attention pooling mechanism is used to weight and aggregate the second-order and third-order interaction features to obtain the fused features; Based on fusion features, a sequence-to-sequence model is used to predict the logistics demand in various regions over future time periods.

5. The cloud-based logistics data analysis and decision-making method according to claim 4, characterized in that, The forecast of logistics demand in each region also includes: The fused feature sequence is obtained, and a bidirectional long short-term memory network encoder is used to encode the fused feature sequence. The forward and backward dependencies of the temporal features are extracted to obtain the encoder output. The Long Short-Term Memory Network Decoder employs an attention mechanism, which focuses on the relevant temporal information of the encoder output in each decoding step. The decoder output is mapped to the predicted logistics demand for each region through a fully connected layer.

6. The cloud-based logistics data analysis and decision-making method according to claim 1, characterized in that, The optimization of delivery routes and task allocation specifically includes: Based on the logistics demand forecast results, a set of delivery tasks is constructed, and the K-means clustering algorithm is used to divide the delivery tasks into multiple sub-regions according to geographical location and delivery time window constraints; For each sub-region, a vehicle routing planning sub-problem is constructed. The objective function of the sub-problem includes transportation cost, time cost, and service quality penalty. The Lagrange relaxation method is used to handle the coupling constraints between subproblems, decomposing the original problem into independent subproblems; Each subproblem is allocated independent computing resources on the cloud platform for parallel optimization and solution.

7. The cloud-based logistics data analysis and decision-making method according to claim 1, characterized in that, The resource scheduling optimization includes: Based on the historical demand data in the fused logistics dataset, statistical analysis is performed to calculate the mean and standard deviation of demand in each region and to construct a probability distribution model for demand uncertainty. A robust optimization objective function is constructed based on the worst-case demand scenario. The objective function includes a weighted combination of resource scheduling cost and demand satisfaction. Establish resource availability constraints, resource transfer time constraints, and non-negativity constraints; The optimal resource scheduling scheme is obtained by solving the robust optimization model using a linear programming solver.

8. The cloud-based logistics data analysis and decision-making method according to claim 1, characterized in that, The generation of the intelligent decision support report specifically includes: Based on historical logistics operation data in the fused logistics dataset, K-means clustering algorithm is used for cluster analysis to identify typical logistics operation scenario categories; Based on the cosine similarity calculation between the current state and each typical scenario, the key scenario most similar to the current state is selected. The Monte Carlo method is used to generate random samples of key scenarios, and parallel simulation analysis is performed on a cloud platform. Based on the simulation results, a weighted evaluation method is used to comprehensively evaluate different decision-making schemes, generating a decision support report that includes risk analysis and sensitivity analysis.

9. The cloud-based logistics data analysis and decision-making method according to claim 8, characterized in that, The weighted evaluation method based on simulation results is used to comprehensively evaluate different decision-making schemes, specifically including: The weighted comprehensive score of the decision-making scheme in terms of cost, efficiency, and service quality is calculated based on the simulation results of each scenario. Sensitivity analysis was used to identify the key parameters and uncertainties that have the greatest impact on decision-making outcomes. Based on a risk assessment model, the potential risks and benefits of different decision-making options are quantified; Generate comprehensive decision support reports that include decision recommendations, risk warnings, key indicator forecasts, and visualization charts.

10. A cloud-based logistics data analysis and decision-making system, characterized in that: The method for performing the cloud-based logistics data analysis and decision-making method according to any one of claims 1-9 includes: The multi-source data fusion module is used to acquire multi-source datasets, perform spatiotemporal standardization on data at different time granularities, calculate confidence scores, and use an adaptive weight fusion method to obtain the fused logistics dataset. The network status assessment module constructs a logistics network graph based on the fused logistics dataset, predicts the status of each node in the logistics network graph, identifies bottleneck nodes and critical paths, and obtains a logistics network status assessment report. The demand forecasting module extracts order time-series features and constructs an external factor feature matrix based on the logistics network status assessment report and the fused logistics dataset to predict logistics demand and obtain logistics demand forecasting results. The route planning module constructs a set of delivery tasks based on logistics demand forecasting results, predicts road segment travel time, optimizes delivery routes and task allocation, and obtains dynamic delivery route solutions. The resource scheduling module performs supply and demand matching analysis based on logistics demand forecasting results and dynamic delivery route plans, and optimizes resource scheduling based on the supply and demand matching analysis results to obtain a resource pre-scheduling plan. The decision support module identifies key scenarios based on dynamic delivery route plans and resource pre-scheduling plans, simulates the execution process, calculates the expected performance and risk indicators of the decision-making plan, and generates an intelligent decision support report.

Images