A business intelligence processing method and system for a supply chain integrated service platform

Abstract

The application discloses a kind of business intelligent processing method and system for supply chain comprehensive service platform, including the following steps: collecting each data and pre-processing;Based on the related feature data of task and road network;Constitute the space-time scheduling network adjusted by confidence degree;Divide frozen area, rolling area and look-ahead area, construct initial scheduling scheme;Using improved ant colony algorithm to search the task sequence and driving path of each vehicle;Using the local improvement of partial vehicle scheduling based on tabu strategy neighborhood search algorithm, obtain target scheduling scheme;According to target scheduling scheme, execute vehicle scheduling.The application can obtain vehicle scheduling scheme with shorter driving time and lower waiting time in large-scale port collection and distribution scene by constructing space-time scheduling network and combining improved ant colony global search and tabu neighborhood local fine optimization, significantly improve the overall efficiency of port collection and distribution vehicle scheduling.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of port transportation, and in particular to a business intelligent processing method and system for a supply chain comprehensive service platform. BACKGROUND

[0002] With the continuous improvement of the scale and automation level of port container terminals, the scheduling problem of port container transportation vehicles is becoming increasingly complex. In the prior art, the common practice is to abstract vehicle scheduling as a vehicle path planning problem with time windows based on port road network, yard and ship schedule information, and to generate a static scheduling scheme under given parameters using rule strategies, mathematical programming or heuristic algorithms. Some researches introduce task volume prediction or traffic prediction, but mostly use simple time series models or statistical models to give point prediction values of future task volume and road section travel time, and then use the point prediction as a fixed input to participate in scheduling optimization.

[0003] However, the prior art generally assumes that task arrival and road network traffic state are relatively certain, lacks quantification and utilization of prediction uncertainty, and cannot dynamically adjust task time windows, penalty coefficients and safety buffer quantities according to prediction confidence, resulting in large-scale delays and frequent rework in actual execution of the scheduling scheme. At the same time, most methods still use a single period or simple rolling method in the time dimension, without zoning management of the scheduling time axis, without distinguishing between frozen tasks that should be kept stable, rolling tasks that can be adjusted in the medium term, and prospect tasks that are roughly planned in the long term, which can easily disturb the issued plan when adjusting locally and affect the stability of on-site execution. In addition, in terms of optimization, common methods usually focus on one-time global construction or rely only on local search, lacking a collaborative optimization mechanism that combines global search with differentiated local refinement for rolling and prospect zones, resulting in insufficient real-time performance, robustness and overall operational efficiency of the scheduling scheme in large-scale, strongly uncertain port transportation scenarios. SUMMARY

[0004] To solve the technical problems in the background art, the present application proposes a business intelligent processing method and system for a supply chain comprehensive service platform.

[0005] The business intelligent processing method for a supply chain comprehensive service platform proposed by the present application comprises the following steps:

[0006] S1, collect and preprocess various data in the port supply chain production system to obtain task feature vectors, road network feature vectors and time feature vectors;

[0007] S2, obtaining a task arrival quantity prediction value, a task quantity prediction value, a task prediction confidence, a road segment expected travel time, a road segment congestion probability, and a road network prediction confidence based on the task feature vector, the road network feature vector, and the time feature vector;

[0008] S3, constructing a basic space-time scheduling network, adjusting weights of space-time edges of the basic space-time scheduling network based on the task prediction confidence and the road network prediction confidence to obtain a confidence-adjusted space-time scheduling network;

[0009] S4, dividing a frozen area, a rolling area, and a look-ahead area, assigning tasks to each vehicle based on the confidence-adjusted space-time scheduling network to construct an initial scheduling scheme;

[0010] S5, taking the initial scheduling scheme as an initial solution, searching for a task sequence and a travel path of each vehicle in the rolling area and the look-ahead area based on the confidence-adjusted space-time scheduling network using an improved ant colony algorithm to obtain a candidate scheduling scheme set;

[0011] S6, based on the candidate scheduling scheme set, using a neighborhood search algorithm based on a tabu strategy to locally improve scheduling of part of the vehicles in the rolling area and the look-ahead area to obtain a target scheduling scheme;

[0012] S7, executing vehicle scheduling according to the target scheduling scheme.

[0013] Preferably, each item of data in the port supply chain production system is collected and preprocessed to obtain the task feature vector, the road network feature vector, and the time feature vector as follows:

[0014] The ship plan arrival and departure times, the actual ship arrival and departure times, the yard handling operation records, the storage occupancy rate, the collection and distribution vehicle positioning trajectories, the collection and distribution vehicle operation logs, the port road network structure information, the historical travel time records of each road segment, the road event information, and the weather information in the port production system are collected;

[0015] The collected data is added with a time mark according to a unified time reference, and different sources of data are aligned according to the time mark, so that each type of data corresponds on the same time axis;

[0016] The aligned data is subjected to outlier identification and elimination to remove data records that are obviously beyond the reasonable range of business and will affect the accuracy of subsequent modeling, and then aggregated by time to form a task original sequence divided by time slice and a road network original sequence divided by time slice;

[0017] The numerical fields representing quantity, duration, and occupancy rate in the task original sequence and the road network original sequence are subjected to normalization processing or standardization processing to eliminate differences between different dimensions;

[0018] encoding and converting category type fields representing ship type, container type category, yard area, road event type and weather condition into numerical features available for model input;

[0019] reorganizing the processed task data according to ship identification, yard identification, container type category and operation link to obtain task feature vectors representing operation scale and operation state of different task units at each time slice;

[0020] reorganizing the processed road network data according to topological connection relationship, functional attribute, length, speed limit and historical passing time statistics of road segments to obtain road network feature vectors representing structural attribute and passing characteristics of each road segment;

[0021] constructing time features based on hour information, shift information, date information, holiday information and peak period information, and arranging the constructed time features into time feature vectors representing operation rhythm and periodicity.

[0022] Preferably, the task arrival quantity prediction value, the operation quantity prediction value, the task prediction confidence, the road segment expected passing time, the road segment congestion probability and the road network prediction confidence are obtained based on the task feature vector, the road network feature vector and the time feature vector, as follows:

[0023] grouping task data corresponding to different ships, different yards and different container types according to preset time slices, and arranging task feature vectors and time feature vectors belonging to the same object and at consecutive time slices in chronological order to construct corresponding time series input sample sets;

[0024] constructing a task-side time series prediction model, the task-side time series prediction model comprising a sequence encoding layer for encoding joint feature sequences, an attention transformation layer for capturing dependency relationships between different time slices, and a prediction output layer for outputting task prediction results of each future time slice, wherein the attention transformation layer highlights historical time slices having greater influence on current prediction by assigning different weights to different time slices;

[0025] inputting the task feature vector and the time feature vector sequence of the current time into the task-side time series prediction model to obtain task arrival quantity prediction values and operation quantity prediction values of each future time slice;

[0026] predicting the same task feature vector and time feature vector sequence by constructing multiple parallel task-side time series prediction models with the same structure but different initialization parameters, and calculating the dispersion degree of each prediction result at each time slice to convert the dispersion degree into the task prediction confidence of the corresponding time slice;

[0027] The specific method for converting dispersion into task prediction confidence is to first calculate the average of these prediction results, then calculate the degree of deviation of each prediction result from the average and sum these deviations. The smaller the deviation, the higher the confidence, and vice versa.

[0028] A spatiotemporal prediction model for the road network is constructed. The spatiotemporal prediction model for the road network includes an input encoding layer for receiving the joint feature sequence of each road segment and extracting features within the time slice; a spatial association modeling layer for aggregating information of adjacent road segments based on the adjacency relationship of road segments; a time dependency modeling layer for modeling the evolution relationship of traffic status in consecutive time slices of the same road segment; and a prediction output layer for outputting the prediction results of the traffic status of road segments in future time slices. The time dependency modeling layer assigns different weights to different time slices to highlight historical time slices that have a greater impact on future status.

[0029] Input the current road network feature vector and time feature vector into the road network-side spatiotemporal prediction model, and output the expected passage time and congestion probability of road segments for each future time slice;

[0030] By constructing multiple parallel road network-side spatiotemporal prediction models with identical structures but different initialization parameters, predictions are made on the same road network feature vector and time feature vector sequence. The fluctuation degree of each prediction result in each time slice is statistically analyzed, and the fluctuation degree is converted into the road network prediction confidence of the corresponding time slice. The principle is the same as obtaining the task prediction confidence, so that time slices with smaller fluctuations obtain higher road network prediction confidence, and time slices with larger fluctuations obtain lower road network prediction confidence.

[0031] Preferably, a basic spatiotemporal scheduling network is constructed, and the weights of the spatiotemporal edges of the basic spatiotemporal scheduling network are adjusted based on the task prediction confidence and the road network prediction confidence to obtain a confidence-adjusted spatiotemporal scheduling network, as follows:

[0032] Obtain the maximum allowable scheduling delay, and set a uniform scheduling time step based on the maximum allowable scheduling delay. Divide the future preset time range into several consecutive scheduling time steps, with each scheduling time step corresponding to a non-overlapping time interval.

[0033] The predicted arrival and workload values ​​of the tasks output by time slices are mapped according to the overlap between the time slices and the scheduling time steps. When the length of the time slice is greater than the length of the scheduling time step, the predicted value corresponding to the time slice is split into multiple scheduling time steps according to the time coverage ratio. When the length of the time slice is less than the length of the scheduling time step, the predicted value corresponding to the time slice is evenly distributed within the scheduling time step to obtain a task prediction sequence with the scheduling time step as the time index. At the same time, the expected passage time and congestion probability of the road segment output by time slices are mapped according to the overlap between the time slices and the scheduling time steps to obtain a road network prediction sequence with the scheduling time step as the time index.

[0034] Based on the port area road network structure, a set of physical nodes is constructed by selecting yard nodes, berth nodes, gate nodes, and hub intersection nodes. For each physical node in the set of physical nodes, it is copied at each scheduling time step. Each physical node and its scheduling time step are combined into a spatiotemporal node. All spatiotemporal nodes are combined to form a set of spatiotemporal nodes covering the entire preset time range.

[0035] A spatiotemporal edge set is constructed based on the movement and waiting process of vehicles between adjacent spatiotemporal nodes, and the basic weight of each spatiotemporal edge is determined by the expected travel time of the road segment and the road segment congestion probability at the corresponding scheduling time step, thus forming a basic spatiotemporal scheduling network.

[0036] Taking the vehicle's operation within the port area as a reference, in the set of spatiotemporal nodes, a moving spatiotemporal edge is established for pairs of spatiotemporal nodes that are directly connected in space by a single road segment and belong to adjacent scheduling time steps in time. This edge is used to indicate that the vehicle travels along the corresponding road segment between adjacent scheduling time steps. A waiting spatiotemporal edge is established for pairs of spatiotemporal nodes corresponding to the same physical node in adjacent scheduling time steps. This edge is used to indicate that the vehicle stays at that node between adjacent scheduling time steps. A service spatiotemporal edge is established between the spatiotemporal nodes corresponding to the yard node and the berth node. This edge is used to indicate that the vehicle performs loading and unloading operations at that node. Thus, a set of spatiotemporal edges including moving spatiotemporal edges, waiting spatiotemporal edges, and service spatiotemporal edges is constructed.

[0037] For each moving spatiotemporal edge, the basic weight of the moving spatiotemporal edge is determined based on the expected travel time of its associated road segment in the corresponding scheduling time step and the road segment congestion probability. The expected travel time is used as the basic part, and the congestion probability is converted into an additional penalty term through mapping and weighted and superimposed with the basic part. For each waiting spatiotemporal edge, the basic weight of the waiting spatiotemporal edge is determined based on the length of the scheduling time step it covers and the type of the corresponding physical node. For each service spatiotemporal edge, the basic weight of the service spatiotemporal edge is determined based on the expected loading and unloading operation time and operation load of the corresponding node.

[0038] The basic time window for each task is obtained. The parameters of the basic time window are determined based on the planned time of the vessel to which the task belongs, the operation rhythm of the corresponding yard, and the expected loading and unloading operation time. Then, the basic time window is adjusted in combination with the task prediction confidence at each scheduling time step.

[0039] Set a base lateness penalty coefficient for each task, determine the magnitude of the lateness penalty coefficient based on the task type, task priority and business requirements, and then adjust the lateness penalty coefficient based on the task prediction confidence.

[0040] For each road segment, the risk penalty adjustment coefficient is calculated based on the road network prediction confidence at each scheduling time step. The basic weights of the spatiotemporal edges of the corresponding road segment in the basic spatiotemporal scheduling network are combined with the risk penalty adjustment coefficient to obtain the risk penalty weights of the spatiotemporal edges, thereby obtaining the spatiotemporal scheduling network adjusted by confidence.

[0041] For each scheduling time step, the overall confidence level is obtained by combining the statistical results of the prediction confidence of all tasks and the prediction confidence of the road network within the scheduling time step. The safety buffer time is then set based on the overall confidence level.

[0042] Preferably, the system divides the area into a frozen zone, a rolling zone, and a prospective zone. Based on a time-space scheduling network adjusted for confidence levels, tasks are assigned to each vehicle to construct an initial scheduling scheme, as follows:

[0043] The scheduling timeline is set according to the maximum allowable scheduling delay. Starting from the current moment, the first preset time point, the second preset time point, and the third preset time point are determined sequentially on the scheduling timeline. The time interval from the current moment to the first preset time point is divided into the frozen zone, the time interval from the first preset time point to the second preset time point is divided into the rolling zone, the time interval from the second preset time point to the third preset time point is divided into the prospective zone, and the time interval after the third preset time point is marked as the time interval outside the current planning range.

[0044] By comparing the start and end times of each scheduling time step on the scheduling time axis with the boundaries of the frozen zone, rolling zone, and prospective zone, all scheduling time steps are marked as frozen time steps, rolling time steps, and prospective time steps, respectively. For scheduling time steps whose start and end times cross the boundaries of different zones, they are classified into the time zone with the largest proportion according to the proportion of their time length in each zone.

[0045] For all tasks, they are marked as frozen tasks, rolling tasks, and prospective tasks respectively by comparing the execution time range of the tasks with the frozen zone, rolling zone, and prospective zone on the scheduling time axis.

[0046] For tasks marked as frozen tasks, a frozen flag is set in the task attributes, indicating that the vehicle allocation result and execution order corresponding to the task will remain unchanged in subsequent optimization processes or will only be allowed to be adjusted in abnormal situations. For tasks marked as rolling tasks, an adjustable flag is set in the task attributes, indicating that the task will allow the reallocation of vehicles and adjustment of the execution order in subsequent optimization processes. For tasks marked as prospective tasks, a coarse planning flag is set in the task attributes, indicating that the task only needs to determine the vehicle or service time range it belongs to in the current scheduling cycle, without having to refine it to the specific scheduling time step.

[0047] Based on the confidence-adjusted time-space scheduling network, tasks are assigned to each vehicle and the initial task execution order is determined to construct an initial scheduling scheme. The part corresponding to the frozen task is set as the frozen part, and the parts corresponding to the rolling task and the prospective task are set as the optimizable part.

[0048] Preferably, the initial scheduling scheme is used as the initial solution. Based on the time-space scheduling network adjusted with confidence, an improved ant colony algorithm is used to search for the task sequences and driving paths of each vehicle in the rolling zone and the lookout zone, resulting in a set of candidate scheduling schemes, as follows:

[0049] Only the spatiotemporal nodes and edges corresponding to the optimizable parts are included in the subsequent ant colony search, while the spatiotemporal nodes and edges corresponding to the frozen parts are set to read-only and do not participate in path construction.

[0050] A comprehensive optimization objective is constructed based on the basic weights, risk penalty weights, and safety buffer durations of each spatiotemporal edge in the confidence-adjusted spatiotemporal scheduling network. This objective is used to evaluate the merits of different vehicle scheduling schemes, and the scheme with the smaller comprehensive optimization objective value is considered the better scheme.

[0051] The pheromone intensity parameter is initialized for each spatiotemporal edge in the optimizable part. Calculation rules are set for heuristic information that includes factors such as distance, time window, congestion level, and task priority. This allows the heuristic information to reflect the attractiveness of the current spatiotemporal edge in shortening travel time, reducing waiting time, and ensuring the timely completion of critical tasks. The combination of pheromone and heuristic information improves the ant colony algorithm's guidance to the optimal path.

[0052] At the beginning of each iteration, a number of ants are assigned to each vehicle in the optimizable part. The initial position of each ant is set as the starting spatiotemporal node of the vehicle in the spatiotemporal scheduling network. The frozen part of the task sequence of the vehicle in the initial scheduling scheme is solidified into the prefix path that the ants must follow. On this basis, the ants are only allowed to construct paths on the spatiotemporal nodes and spatiotemporal edges corresponding to the rolling zone and the prospect zone.

[0053] After each ant completes the path construction from the starting spatiotemporal node to the vehicle's ending spatiotemporal node, the comprehensive optimization target value is calculated based on the constructed complete scheduling scheme. The comprehensive optimization target values ​​of all ants are sorted, and the candidate scheduling scheme with the smallest comprehensive optimization target value in this iteration is recorded.

[0054] A set of candidate scheduling schemes is obtained by iteratively executing the path construction process.

[0055] Preferably, during the path construction process, when the ant of a vehicle is located at the current spatiotemporal node, candidate spatiotemporal edges that would violate vehicle capacity constraints are eliminated based on the set of next spatiotemporal nodes reachable from the current spatiotemporal node. Transfer reference values ​​are calculated for the remaining candidate spatiotemporal edges. The transfer reference values ​​are weighted combinations based on the pheromone intensity, heuristic information, congestion penalty weight, and corresponding task priority on the current candidate spatiotemporal edge, and the transfer probability of each candidate spatiotemporal edge is obtained through normalization processing. Thus, the next spatiotemporal node is selected probabilistically, thereby guiding the path construction process.

[0056] Preferably, in the pheromone update phase, the improved ant colony algorithm performs pheromone evaporation processing on the spatiotemporal edges participating in path construction, reduces the pheromone intensity on each spatiotemporal edge according to a preset evaporation ratio, and sets differentiated pheromone adjustment strategies for the rolling zone and the prospecting zone. A larger pheromone increment is applied to the spatiotemporal edges in the rolling zone that belong to the better solution of the comprehensive optimization objective, so as to accelerate the convergence of path selection in the rolling zone. A higher pheromone evaporation ratio is set for the spatiotemporal edges in the prospecting zone, so as to enhance the diversity and exploration capability of path search in the prospecting zone.

[0057] Preferably, based on the candidate scheduling scheme set, a neighborhood search algorithm based on a tabu strategy is used to locally improve the scheduling of some vehicles in the rolling zone and the prospective zone, resulting in the target scheduling scheme, as follows:

[0058] Based on the candidate scheduling scheme set, the candidate scheduling scheme with the smallest comprehensive optimization objective value is selected as the initial solution for neighborhood search. Then, according to the candidate scheduling scheme, the task sequence, driving path and task start and end time corresponding to each vehicle in the candidate scheduling scheme are extracted to obtain the current scheduling solution.

[0059] Based on the current scheduling solution, identify the task segments that belong to the frozen part and the task segments that belong to the optimizable part. Mark the task segments corresponding to the frozen tasks as non-adjustable segments, and mark the task segments corresponding to the rolling tasks and the prospective tasks as adjustable segments.

[0060] For adjustable segments, define task swapping, task insertion, and path replacement operations;

[0061] The task exchange operation is used to exchange two rolling tasks or prospective tasks between different vehicles or between different positions of the same vehicle. By selecting a task to be exchanged in the current scheduling solution, a new task allocation combination is constructed to explore the impact of different vehicles undertaking different tasks and different task execution orders on the comprehensive optimization objective value.

[0062] The task insertion operation is used to extract a rolling task or prospective task from the task sequence of a certain vehicle and attempt to insert the task into a different position in the task sequence of another vehicle or the same vehicle. At each candidate insertion position, the start and end times of the task sequence of the vehicle are recalculated to find a task insertion method that is more conducive to reducing travel time and waiting time.

[0063] The path replacement operation is used to select different driving paths for a vehicle between two adjacent tasks without changing the vehicle task order. The spatiotemporal edges traversed by the corresponding path in the current scheduling solution are replaced with the spatiotemporal edges traversed by the alternative path. This allows for the exploration of a better combination of paths with better travel risk and travel time under the premise of feasible time constraints, by utilizing the preset alternative path priority and risk penalty weight in the confidence-adjusted spatiotemporal scheduling network.

[0064] At the beginning of each iteration, based on the current scheduling solution, and according to the preset neighborhood size, one operation is selected from task swapping, task insertion, and path replacement to randomly perturb the tasks and paths in the adjustable segment, generating several neighborhood candidate scheduling solutions.

[0065] Calculate the comprehensive optimization objective value for each neighborhood candidate scheduling solution, and select the neighborhood candidate scheduling solution with the smallest comprehensive optimization objective value as the preferred neighborhood solution in this iteration.

[0066] Construct a tabu list and record the task swapping, task insertion, and path replacement operations that have been executed in the iteration. Operations recorded in the tabu list are considered taboo operations and are temporarily prohibited from being executed repeatedly in subsequent iterations to avoid the search process repeatedly going back and forth in the same set of solutions, thereby enhancing the exploration ability of neighborhood search in the solution space.

[0067] In each iteration, it is determined whether the preferred neighborhood solution obtained in this round was generated by a taboo operation. If the corresponding operation is a taboo operation and the overall optimization objective value of the preferred neighborhood solution is not better than the currently known global optimal scheduling solution, then the preferred neighborhood solution is abandoned according to the taboo strategy, and a neighborhood solution that is not taboo and has the second best overall optimization objective value is selected as the new current scheduling solution. If the preferred neighborhood solution is generated by a taboo operation but the overall optimization objective value is better than the currently known global optimal scheduling solution, then the preferred neighborhood solution is updated as the new current scheduling solution and the new global optimal scheduling solution.

[0068] After the iteration process is completed, the globally optimal scheduling solution is output as the target scheduling scheme.

[0069] A business intelligence processing system for a supply chain integrated service platform includes:

[0070] The data acquisition and feature construction module is used to collect various data from the port supply chain production system and preprocess them to obtain task feature vectors, road network feature vectors, and time feature vectors.

[0071] The time series and spatiotemporal prediction module is used to obtain the predicted values ​​of task arrival, workload, task prediction confidence, expected travel time of road segments, road segment congestion probability, and road network prediction confidence based on task feature vectors, road network feature vectors, and time feature vectors.

[0072] The temporal and spatial scheduling network construction module is used to construct a basic temporal and spatial scheduling network. Based on the task prediction confidence and the road network prediction confidence, the weights of the spatiotemporal edges of the basic temporal and spatial scheduling network are adjusted to obtain a confidence-adjusted temporal and spatial scheduling network.

[0073] The initial scheduling generation module is used to divide the frozen area, rolling area and prospective area, and to assign tasks to each vehicle based on the time-space scheduling network adjusted by confidence, and to construct the initial scheduling scheme.

[0074] The ant colony global optimization module is used to take the initial scheduling scheme as the initial solution, and based on the confidence-adjusted time-space scheduling network, adopts an improved ant colony algorithm to search the task sequence and driving path of each vehicle in the rolling zone and the prospect zone to obtain a set of candidate scheduling schemes.

[0075] The neighborhood search local optimization module is used to improve the scheduling of some vehicles in the rolling zone and the prospect zone based on the set of candidate scheduling schemes and the neighborhood search algorithm based on the tabu strategy, so as to obtain the target scheduling scheme.

[0076] The scheme execution module is used to execute vehicle scheduling according to the target scheduling scheme.

[0077] The business intelligence processing method and system for a supply chain integrated service platform proposed in this invention have the following beneficial technical effects:

[0078] 1. This application constructs a dual-channel spatiotemporal prediction system for both the task side and the road network side under a unified scheduling time step. Based on this, a spatiotemporal scheduling network covering storage yards, berths, gates, and road network nodes is formed. This achieves collaborative modeling of the process of generating collection and distribution tasks and the evolution of road network traffic status. By utilizing the dispersion of multi-model prediction results, task prediction confidence and road network prediction confidence are constructed. A clear mapping relationship is established between the confidence and the time window adjustment ratio, penalty adjustment coefficient, and safety buffer adjustment coefficient. This ensures that the task time window, lateness penalty, road segment risk penalty, and buffer duration of each scheduling time step are no longer static fixed parameters, but are adaptively adjusted according to the prediction reliability. Thus, while ensuring feasibility, the impact of prediction errors on the executability and stability of the scheduling scheme is effectively reduced, and the robustness of the scheme to uncertain disturbances is improved.

[0079] 2. This application achieves hierarchical rolling optimization of short-term and long-term tasks by adopting a time partitioning mechanism of frozen zone, rolling zone, and prospective zone, and a corresponding hierarchical management mechanism for frozen tasks, rolling tasks, and prospective tasks. It improves the ant colony algorithm by combining pheromones with heuristic information considering congestion risk, time windows, and task priorities on a time-spaced scheduling network with adjusted confidence to achieve differentiated search. This guides the rolling zone to converge faster and the prospective zone to continue exploring. Neighborhood search refines candidate solutions under strict time window and capacity constraints through operations such as task exchange, task insertion, and path replacement. Combined with real-time deviation monitoring and rolling updates of prediction confidence during execution, this application can continuously output vehicle scheduling solutions with lower travel and waiting times, less sensitivity to state disturbances, and controllable adjustment costs in large-scale port collection and distribution scenarios. This significantly improves the problems of rigid scheduling solutions, easy failure, and insufficient global performance in existing technologies. Attached Figure Description

[0080] Figure 1 This is a flowchart of a business intelligence processing method for a supply chain integrated service platform according to the present invention;

[0081] Figure 2 This is a schematic diagram of a business intelligence processing system for a supply chain integrated service platform according to the present invention. Detailed Implementation

[0082] Embodiments of the present invention are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar symbols denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0083] like Figure 1 The business intelligence processing method shown is for a supply chain integrated service platform and includes the following steps:

[0084] S1. Collect and preprocess various data from the port supply chain production system to obtain task feature vector, road network feature vector, and time feature vector;

[0085] In an optional embodiment, various data from the port supply chain production system are collected and preprocessed to obtain task feature vectors, road network feature vectors, and time feature vectors, as follows:

[0086] Collect data from the port production system, including planned arrival and departure times of ships, actual arrival and departure times of ships, yard loading and unloading operation records, storage occupancy rate, location trajectory of transport vehicles, operation logs of transport vehicles, port area road network structure information, historical passage time records of each road section, road event information, and weather information.

[0087] The collected data is time-stamped according to a unified time base, and data from different sources are aligned based on the time stamps so that all types of data correspond on the same timeline.

[0088] Outlier identification and removal are performed on the aligned data to remove data records that are clearly outside the reasonable range of business and will affect the accuracy of subsequent modeling. Then, the data is aggregated by time to form the original task sequence divided by time slice and the original road network sequence divided by time slice.

[0089] Normalize or standardize the numerical fields representing quantity, duration, and occupancy rate in the original task sequence and the original road network sequence to eliminate differences between different units.

[0090] The categorical fields representing ship type, container type, yard area, road event type, and weather conditions are encoded and converted into numerical features that can be used as model input.

[0091] The processed task data is reorganized according to ship identification, yard identification, container type and operation stage to obtain task feature vectors that characterize the operation scale and operation status of different task units in each time slice.

[0092] The processed road network data is reorganized according to the topological connection relationship, functional attributes, length, speed limit and historical travel time statistics of the road segments to obtain the road network feature vector used to characterize the structural attributes and traffic characteristics of each road segment.

[0093] Time features are constructed based on hourly information, shift information, date information, whether it is a holiday or peak period information. The constructed time features are then organized into time feature vectors to characterize the work rhythm and periodicity.

[0094] S2. Based on the task feature vector, road network feature vector, and time feature vector, obtain the predicted value of task arrival, the predicted value of workload, the task prediction confidence, the expected travel time of road segment, the probability of road segment congestion, and the road network prediction confidence.

[0095] In an optional embodiment, the predicted arrival value of the task, the predicted workload, the predicted confidence level of the task, the expected travel time of the road segment, the probability of road segment congestion, and the predicted confidence level of the road network are obtained based on the task feature vector, the road network feature vector, and the time feature vector, as follows:

[0096] The task data corresponding to different ships, different storage yards and different container types are grouped according to the preset time slices. The task feature vectors and time feature vectors that belong to the same object and are in consecutive time slices are arranged in chronological order to construct the corresponding time series input sample set.

[0097] A task-side time-series prediction model is constructed, which includes a sequence encoding layer for encoding joint feature sequences, an attention transformation layer for capturing the dependencies between different time slices, and a prediction output layer for outputting the task prediction results for each future time slice. The attention transformation layer assigns different weights to different time slices to highlight historical time slices that have a greater impact on the current prediction.

[0098] Input the current task feature vector and time feature vector sequence into the task-side time series prediction model to obtain the predicted values ​​of task arrival and workload for each future time slice;

[0099] By constructing multiple parallel task-side time-series prediction models with the same structure but different initialization parameters, the same task feature vector and time feature vector sequence are predicted. The dispersion of each prediction result at each time slice is statistically analyzed, and the dispersion is converted into the task prediction confidence of the corresponding time slice.

[0100] The specific method for converting dispersion into task prediction confidence is to first calculate the average of these prediction results, then calculate the degree of deviation of each prediction result from the average and sum these deviations. The smaller the deviation, the higher the confidence, and vice versa.

[0101] A spatiotemporal prediction model for the road network is constructed. The spatiotemporal prediction model for the road network includes an input encoding layer for receiving the joint feature sequence of each road segment and extracting features within the time slice; a spatial association modeling layer for aggregating information of adjacent road segments based on the adjacency relationship of road segments; a time dependency modeling layer for modeling the evolution relationship of traffic status in consecutive time slices of the same road segment; and a prediction output layer for outputting the prediction results of the traffic status of road segments in future time slices. The time dependency modeling layer assigns different weights to different time slices to highlight historical time slices that have a greater impact on future status.

[0102] Input the current road network feature vector and time feature vector into the road network-side spatiotemporal prediction model, and output the expected passage time and congestion probability of road segments for each future time slice;

[0103] By constructing multiple parallel road network-side spatiotemporal prediction models with identical structures but different initialization parameters, predictions are made on the same road network feature vector and time feature vector sequence. The fluctuation degree of each prediction result in each time slice is statistically analyzed, and the fluctuation degree is converted into the road network prediction confidence of the corresponding time slice. The principle is the same as obtaining the task prediction confidence, so that time slices with smaller fluctuations obtain higher road network prediction confidence, and time slices with larger fluctuations obtain lower road network prediction confidence.

[0104] S3. Construct a basic time-space scheduling network. Adjust the weights of the spatiotemporal edges of the basic time-space scheduling network based on the task prediction confidence and the road network prediction confidence to obtain a confidence-adjusted time-space scheduling network.

[0105] In an optional embodiment, a basic spatiotemporal scheduling network is constructed. The weights of the spatiotemporal edges of the basic spatiotemporal scheduling network are adjusted based on the task prediction confidence and the road network prediction confidence to obtain a confidence-adjusted spatiotemporal scheduling network, as follows:

[0106] Obtain the maximum allowable scheduling delay, and set a uniform scheduling time step based on the maximum allowable scheduling delay. Divide the future preset time range into several consecutive scheduling time steps, with each scheduling time step corresponding to a non-overlapping time interval.

[0107] The predicted arrival and workload values ​​of the tasks output by time slices are mapped according to the overlap between the time slices and the scheduling time steps. When the length of the time slice is greater than the length of the scheduling time step, the predicted value corresponding to the time slice is split into multiple scheduling time steps according to the time coverage ratio. When the length of the time slice is less than the length of the scheduling time step, the predicted value corresponding to the time slice is evenly distributed within the scheduling time step to obtain a task prediction sequence with the scheduling time step as the time index. At the same time, the expected passage time and congestion probability of the road segment output by time slices are mapped according to the overlap between the time slices and the scheduling time steps to obtain a road network prediction sequence with the scheduling time step as the time index.

[0108] Based on the port area road network structure, a set of physical nodes is constructed by selecting yard nodes, berth nodes, gate nodes, and hub intersection nodes. For each physical node in the set of physical nodes, it is copied at each scheduling time step. Each physical node and its scheduling time step are combined into a spatiotemporal node. All spatiotemporal nodes are combined to form a set of spatiotemporal nodes covering the entire preset time range.

[0109] A spatiotemporal edge set is constructed based on the movement and waiting process of vehicles between adjacent spatiotemporal nodes, and the basic weight of each spatiotemporal edge is determined by the expected travel time of the road segment and the road segment congestion probability at the corresponding scheduling time step, thus forming a basic spatiotemporal scheduling network.

[0110] Taking the vehicle's operation within the port area as a reference, in the set of spatiotemporal nodes, a moving spatiotemporal edge is established for pairs of spatiotemporal nodes that are directly connected in space by a single road segment and belong to adjacent scheduling time steps in time. This edge is used to indicate that the vehicle travels along the corresponding road segment between adjacent scheduling time steps. A waiting spatiotemporal edge is established for pairs of spatiotemporal nodes corresponding to the same physical node in adjacent scheduling time steps. This edge is used to indicate that the vehicle stays at that node between adjacent scheduling time steps. A service spatiotemporal edge is established between the spatiotemporal nodes corresponding to the yard node and the berth node. This edge is used to indicate that the vehicle performs loading and unloading operations at that node. Thus, a set of spatiotemporal edges including moving spatiotemporal edges, waiting spatiotemporal edges, and service spatiotemporal edges is constructed.

[0111] For each moving spatiotemporal edge, the basic weight of the moving spatiotemporal edge is determined based on the expected travel time of its associated road segment in the corresponding scheduling time step and the road segment congestion probability. The expected travel time is used as the basic part, and the congestion probability is converted into an additional penalty term through mapping and weighted and superimposed with the basic part. For each waiting spatiotemporal edge, the basic weight of the waiting spatiotemporal edge is determined based on the length of the scheduling time step it covers and the type of the corresponding physical node. For each service spatiotemporal edge, the basic weight of the service spatiotemporal edge is determined based on the expected loading and unloading operation time and operation load of the corresponding node.

[0112] The basic time window for each task is obtained. The parameters of the basic time window are determined based on the planned time of the vessel to which the task belongs, the operation rhythm of the corresponding yard, and the expected loading and unloading operation time. Then, the basic time window is adjusted in combination with the task prediction confidence at each scheduling time step.

[0113] Set a base lateness penalty coefficient for each task, determine the magnitude of the lateness penalty coefficient based on the task type, task priority and business requirements, and then adjust the lateness penalty coefficient based on the task prediction confidence.

[0114] For each road segment, the risk penalty adjustment coefficient is calculated based on the road network prediction confidence at each scheduling time step. The basic weights of the spatiotemporal edges of the corresponding road segment in the basic spatiotemporal scheduling network are combined with the risk penalty adjustment coefficient to obtain the risk penalty weights of the spatiotemporal edges, thereby obtaining the spatiotemporal scheduling network adjusted by confidence.

[0115] For each scheduling time step, the overall confidence level is obtained by combining the statistical results of the prediction confidence of all tasks and the prediction confidence of the road network within the scheduling time step. The safety buffer time is then set based on the overall confidence level.

[0116] S4. Divide the area into a frozen zone, a rolling zone, and a prospective zone. Assign tasks to each vehicle based on the time-space scheduling network adjusted by confidence level and construct an initial scheduling scheme.

[0117] In an optional embodiment, a frozen zone, a rolling zone, and a prospective zone are divided, and tasks are assigned to each vehicle based on a confidence-adjusted time-space scheduling network to construct an initial scheduling scheme, as follows:

[0118] The scheduling timeline is set according to the maximum allowable scheduling delay. Starting from the current moment, the first preset time point, the second preset time point, and the third preset time point are determined sequentially on the scheduling timeline. The time interval from the current moment to the first preset time point is divided into the frozen zone, the time interval from the first preset time point to the second preset time point is divided into the rolling zone, the time interval from the second preset time point to the third preset time point is divided into the prospective zone, and the time interval after the third preset time point is marked as the time interval outside the current planning range.

[0119] By comparing the start and end times of each scheduling time step on the scheduling time axis with the boundaries of the frozen zone, rolling zone, and prospective zone, all scheduling time steps are marked as frozen time steps, rolling time steps, and prospective time steps, respectively. For scheduling time steps whose start and end times cross the boundaries of different zones, they are classified into the time zone with the largest proportion according to the proportion of their time length in each zone.

[0120] For all tasks, they are marked as frozen tasks, rolling tasks, and prospective tasks respectively by comparing the execution time range of the tasks with the frozen zone, rolling zone, and prospective zone on the scheduling time axis.

[0121] For tasks marked as frozen tasks, a frozen flag is set in the task attributes, indicating that the vehicle allocation result and execution order corresponding to the task will remain unchanged in subsequent optimization processes or will only be allowed to be adjusted in abnormal situations. For tasks marked as rolling tasks, an adjustable flag is set in the task attributes, indicating that the task will allow the reallocation of vehicles and adjustment of the execution order in subsequent optimization processes. For tasks marked as prospective tasks, a coarse planning flag is set in the task attributes, indicating that the task only needs to determine the vehicle or service time range it belongs to in the current scheduling cycle, without having to refine it to the specific scheduling time step.

[0122] Based on the confidence-adjusted time-space scheduling network, tasks are assigned to each vehicle and the initial task execution order is determined to construct an initial scheduling scheme. The part corresponding to the frozen task is set as the frozen part, and the parts corresponding to the rolling task and the prospective task are set as the optimizable part.

[0123] S5. Using the initial scheduling scheme as the initial solution, and based on the time-space scheduling network adjusted with confidence, an improved ant colony algorithm is used to search for the task sequence and driving path of each vehicle in the rolling zone and the prospect zone to obtain a set of candidate scheduling schemes.

[0124] In an optional embodiment, the initial scheduling scheme is used as the initial solution. Based on the confidence-adjusted time-space scheduling network, an improved ant colony algorithm is used to search for the task sequences and driving paths of each vehicle in the rolling zone and the lookout zone, resulting in a set of candidate scheduling schemes, as follows:

[0125] Only the spatiotemporal nodes and edges corresponding to the optimizable parts are included in the subsequent ant colony search, while the spatiotemporal nodes and edges corresponding to the frozen parts are set to read-only and do not participate in path construction.

[0126] A comprehensive optimization objective is constructed based on the basic weights, risk penalty weights, and safety buffer durations of each spatiotemporal edge in the confidence-adjusted spatiotemporal scheduling network. This objective is used to evaluate the merits of different vehicle scheduling schemes, and the scheme with the smaller comprehensive optimization objective value is considered the better scheme.

[0127] The pheromone intensity parameter is initialized for each spatiotemporal edge in the optimizable part. Calculation rules are set for heuristic information that includes factors such as distance, time window, congestion level, and task priority. This allows the heuristic information to reflect the attractiveness of the current spatiotemporal edge in shortening travel time, reducing waiting time, and ensuring the timely completion of critical tasks. The combination of pheromone and heuristic information improves the ant colony algorithm's guidance to the optimal path.

[0128] At the beginning of each iteration, a number of ants are assigned to each vehicle in the optimizable part. The initial position of each ant is set as the starting spatiotemporal node of the vehicle in the spatiotemporal scheduling network. The frozen part of the task sequence of the vehicle in the initial scheduling scheme is solidified into the prefix path that the ants must follow. On this basis, the ants are only allowed to construct paths on the spatiotemporal nodes and spatiotemporal edges corresponding to the rolling zone and the prospect zone.

[0129] After each ant completes the path construction from the starting spatiotemporal node to the vehicle's ending spatiotemporal node, the comprehensive optimization target value is calculated based on the constructed complete scheduling scheme. The comprehensive optimization target values ​​of all ants are sorted, and the candidate scheduling scheme with the smallest comprehensive optimization target value in this iteration is recorded.

[0130] A set of candidate scheduling schemes is obtained by iteratively executing the path construction process;

[0131] In an optional embodiment, during the path construction process, when the ant of a vehicle is located at the current spatiotemporal node, candidate spatiotemporal edges that would violate vehicle capacity constraints are eliminated based on the set of next spatiotemporal nodes reachable from the current spatiotemporal node. A transition reference value is calculated for the remaining candidate spatiotemporal edges. The transition reference value is a weighted combination based on the pheromone intensity, heuristic information, congestion penalty weight, and corresponding task priority on the current candidate spatiotemporal edge. The transition probability of each candidate spatiotemporal edge is obtained through normalization, thereby selecting the next spatiotemporal node in a probabilistic manner to guide the path construction process.

[0132] In an optional embodiment, during the pheromone update phase, the improved ant colony algorithm performs pheromone evaporation processing on the spatiotemporal edges participating in path construction, reducing the pheromone intensity on each spatiotemporal edge according to a preset evaporation ratio, and setting differentiated pheromone adjustment strategies for the rolling zone and the prospecting zone. A larger pheromone increment is applied to the spatiotemporal edges in the rolling zone that belong to the better solution of the comprehensive optimization objective, so as to accelerate the convergence of path selection in the rolling zone. A higher pheromone evaporation ratio is set for the spatiotemporal edges in the prospecting zone, so as to enhance the diversity and exploration capability of path search in the prospecting zone.

[0133] S6. Based on the candidate scheduling scheme set, a neighborhood search algorithm based on tabu strategy is used to locally improve the scheduling of some vehicles in the rolling area and the prospective area to obtain the target scheduling scheme.

[0134] In an optional embodiment, based on the candidate scheduling scheme set, a neighborhood search algorithm based on a tabu strategy is used to locally improve the scheduling of some vehicles in the rolling zone and the prospective zone to obtain the target scheduling scheme, as follows:

[0135] Based on the candidate scheduling scheme set, the candidate scheduling scheme with the smallest comprehensive optimization objective value is selected as the initial solution for neighborhood search. Then, according to the candidate scheduling scheme, the task sequence, driving path and task start and end time corresponding to each vehicle in the candidate scheduling scheme are extracted to obtain the current scheduling solution.

[0136] Based on the current scheduling solution, identify the task segments that belong to the frozen part and the task segments that belong to the optimizable part. Mark the task segments corresponding to the frozen tasks as non-adjustable segments, and mark the task segments corresponding to the rolling tasks and the prospective tasks as adjustable segments.

[0137] For adjustable segments, define task swapping, task insertion, and path replacement operations;

[0138] The task exchange operation is used to exchange two rolling tasks or prospective tasks between different vehicles or between different positions of the same vehicle. By selecting a task to be exchanged in the current scheduling solution, a new task allocation combination is constructed to explore the impact of different vehicles undertaking different tasks and different task execution orders on the comprehensive optimization objective value.

[0139] The task insertion operation is used to extract a rolling task or prospective task from the task sequence of a certain vehicle and attempt to insert the task into a different position in the task sequence of another vehicle or the same vehicle. At each candidate insertion position, the start and end times of the task sequence of the vehicle are recalculated to find a task insertion method that is more conducive to reducing travel time and waiting time.

[0140] The path replacement operation is used to select different driving paths for a vehicle between two adjacent tasks without changing the vehicle task order. The spatiotemporal edges traversed by the corresponding path in the current scheduling solution are replaced with the spatiotemporal edges traversed by the alternative path. This allows for the exploration of a better combination of paths with better travel risk and travel time under the premise of feasible time constraints, by utilizing the preset alternative path priority and risk penalty weight in the confidence-adjusted spatiotemporal scheduling network.

[0141] At the beginning of each iteration, based on the current scheduling solution, and according to the preset neighborhood size, one operation is selected from task swapping, task insertion, and path replacement to randomly perturb the tasks and paths in the adjustable segment, generating several neighborhood candidate scheduling solutions.

[0142] Calculate the comprehensive optimization objective value for each neighborhood candidate scheduling solution, and select the neighborhood candidate scheduling solution with the smallest comprehensive optimization objective value as the preferred neighborhood solution in this iteration.

[0143] Construct a tabu list and record the task swapping, task insertion, and path replacement operations that have been executed in the iteration. Operations recorded in the tabu list are considered taboo operations and are temporarily prohibited from being executed repeatedly in subsequent iterations to avoid the search process repeatedly going back and forth in the same set of solutions, thereby enhancing the exploration ability of neighborhood search in the solution space.

[0144] In each iteration, it is determined whether the preferred neighborhood solution obtained in this round was generated by a taboo operation. If the corresponding operation is a taboo operation and the overall optimization objective value of the preferred neighborhood solution is not better than the currently known global optimal scheduling solution, then the preferred neighborhood solution is abandoned according to the taboo strategy, and a neighborhood solution that is not taboo and has the second best overall optimization objective value is selected as the new current scheduling solution. If the preferred neighborhood solution is generated by a taboo operation but the overall optimization objective value is better than the currently known global optimal scheduling solution, then the preferred neighborhood solution is updated as the new current scheduling solution and the new global optimal scheduling solution.

[0145] After the iteration process is completed, the globally optimal scheduling solution is output as the target scheduling scheme.

[0146] S7. Execute vehicle dispatching according to the target dispatching plan.

[0147] like Figure 2 The business intelligence processing system shown is for a supply chain integrated service platform, comprising:

[0148] The data acquisition and feature construction module is used to collect various data from the port supply chain production system and preprocess them to obtain task feature vectors, road network feature vectors, and time feature vectors.

[0149] The time series and spatiotemporal prediction module is used to obtain the predicted values ​​of task arrival, workload, task prediction confidence, expected travel time of road segments, road segment congestion probability, and road network prediction confidence based on task feature vectors, road network feature vectors, and time feature vectors.

[0150] The temporal and spatial scheduling network construction module is used to construct a basic temporal and spatial scheduling network. Based on the task prediction confidence and the road network prediction confidence, the weights of the spatiotemporal edges of the basic temporal and spatial scheduling network are adjusted to obtain a confidence-adjusted temporal and spatial scheduling network.

[0151] The initial scheduling generation module is used to divide the frozen area, rolling area and prospective area, and to assign tasks to each vehicle based on the time-space scheduling network adjusted by confidence, and to construct the initial scheduling scheme.

[0152] The ant colony global optimization module is used to take the initial scheduling scheme as the initial solution, and based on the confidence-adjusted time-space scheduling network, adopts an improved ant colony algorithm to search the task sequence and driving path of each vehicle in the rolling zone and the prospect zone to obtain a set of candidate scheduling schemes.

[0153] The neighborhood search local optimization module is used to improve the scheduling of some vehicles in the rolling zone and the prospect zone based on the set of candidate scheduling schemes and the neighborhood search algorithm based on the tabu strategy, so as to obtain the target scheduling scheme.

[0154] The scheme execution module is used to execute vehicle scheduling according to the target scheduling scheme.

[0155] Table 1 shows a comparison of scheduling effects during peak hours on a single day.

[0156] Table 1 Comparison of scheduling effects during peak hours on a single day

[0157]

[0158] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A business intelligence processing method for a supply chain integrated service platform, characterized in that, Includes the following steps: S1. Collect and preprocess various data from the port supply chain production system. These data include planned arrival and departure times of ships, actual arrival and departure times of ships, yard loading and unloading operation records, storage occupancy rate, location trajectory of collection and distribution vehicles, operation log of collection and distribution vehicles, port area road network structure, historical passage time records of each road section, road event information, and weather information. Obtain task feature vector, road network feature vector, and time feature vector. S2. Based on the task feature vector, road network feature vector, and time feature vector, obtain the predicted value of task arrival, the predicted value of workload, the task prediction confidence, the expected travel time of road segment, the probability of road segment congestion, and the road network prediction confidence. S3. Construct a basic spatiotemporal scheduling network. Adjust the weights of the spatiotemporal edges of the basic spatiotemporal scheduling network based on the task prediction confidence and the road network prediction confidence to obtain a confidence-adjusted spatiotemporal scheduling network. The specific steps are as follows: Obtain the maximum allowable scheduling delay and set a uniform scheduling time step based on the maximum allowable scheduling delay; The predicted arrival and workload values ​​of the tasks output by time slices are mapped according to the overlap relationship between time slices and scheduling time steps to obtain a task prediction sequence with scheduling time steps as time indexes. At the same time, the expected passage time and congestion probability of road segments output by time slices are mapped according to the overlap relationship between time slices and scheduling time steps to obtain a road network prediction sequence with scheduling time steps as time indexes. Based on the port area road network structure, a set of physical nodes is constructed. For each physical node in the set of physical nodes, it is copied at each scheduling time step. Each physical node and its scheduling time step are combined into a spatiotemporal node. All spatiotemporal nodes are combined to form a set of spatiotemporal nodes. A spatiotemporal edge set is constructed based on the movement and waiting process of vehicles between adjacent spatiotemporal nodes, and the basic weight of each spatiotemporal edge is determined by the expected travel time of the road segment and the road segment congestion probability at the corresponding scheduling time step, thus forming a basic spatiotemporal scheduling network. Obtain the base time window for each task, and then adjust the base time window based on the task prediction confidence at each scheduling time step; Set a base lateness penalty coefficient for each task, and then adjust the lateness penalty coefficient based on the task prediction confidence. For each road segment, the risk penalty adjustment coefficient is calculated based on the road network prediction confidence at each scheduling time step. The basic weights of the spatiotemporal edges of the corresponding road segment in the basic spatiotemporal scheduling network are combined with the risk penalty adjustment coefficient to obtain the risk penalty weights of the spatiotemporal edges, thereby obtaining the spatiotemporal scheduling network adjusted by confidence. For each scheduling time step, the overall confidence level is obtained by combining the statistical results of the prediction confidence of all tasks and the prediction confidence of the road network within the scheduling time step. The safety buffer time is then set based on the overall confidence level. S4. Divide the area into a frozen zone, a rolling zone, and a prospective zone. Assign tasks to each vehicle based on the time-space scheduling network adjusted for confidence, and construct an initial scheduling scheme. The specific steps are as follows: The scheduling timeline is set according to the maximum allowable scheduling delay. The first preset time point, the second preset time point, and the third preset time point are determined sequentially from the current time on the scheduling timeline, and the frozen area, the rolling area, and the prospective area are divided in sequence. By comparing the start and end times of each scheduling time step on the scheduling time axis with the boundaries of the frozen zone, rolling zone, and prospecting zone, all scheduling time steps are marked as frozen time steps, rolling time steps, and prospecting time steps, respectively. For all tasks, they are marked as frozen tasks, rolling tasks, and prospective tasks respectively by comparing the execution time range of the tasks with the frozen zone, rolling zone, and prospective zone on the scheduling time axis. Based on the confidence-adjusted time-space scheduling network, tasks are assigned to each vehicle and the initial task execution order is determined. An initial scheduling scheme is constructed, in which the part corresponding to frozen tasks is set as the frozen part, and the parts corresponding to rolling tasks and prospective tasks are set as the optimizable part. S5. Using the initial scheduling scheme as the initial solution, and based on the time-space scheduling network adjusted with confidence, an improved ant colony algorithm is used to search for the task sequence and driving path of each vehicle in the rolling zone and the prospect zone to obtain a set of candidate scheduling schemes. S6. Based on the candidate scheduling scheme set, a neighborhood search algorithm based on tabu strategy is used to locally improve the scheduling of some vehicles in the rolling area and the prospective area to obtain the target scheduling scheme. S7. Execute vehicle dispatching according to the target dispatching plan.

2. The business intelligence processing method for a supply chain integrated service platform according to claim 1, characterized in that, Based on the task feature vector, road network feature vector, and time feature vector, the predicted values ​​of task arrival volume, workload, task prediction confidence, expected travel time of road segments, road segment congestion probability, and road network prediction confidence are obtained, as follows: Input the current task feature vector and time feature vector sequence into the task-side time series prediction model to obtain the predicted values ​​of task arrival and workload for each future time slice; By constructing multiple parallel task-side time-series prediction models with the same structure but different initialization parameters, the same task feature vector and time feature vector sequence are predicted. The dispersion of each prediction result at each time slice is statistically analyzed, and the dispersion is converted into the task prediction confidence of the corresponding time slice. Input the current road network feature vector and time feature vector into the road network-side spatiotemporal prediction model, and output the expected passage time and congestion probability of road segments for each future time slice; By constructing multiple parallel road network-side spatiotemporal prediction models with the same structure but different initialization parameters, predictions are made on the same road network feature vector and time feature vector sequence. The fluctuation degree of each prediction result in each time slice is statistically analyzed, and the fluctuation degree is converted into the road network prediction confidence of the corresponding time slice.

3. The business intelligence processing method for a supply chain integrated service platform according to claim 1, characterized in that, Using the initial scheduling scheme as the initial solution, an improved ant colony algorithm is employed based on the confidence-adjusted time-space scheduling network to search the task sequences and travel paths of each vehicle in the rolling zone and the lookout zone, resulting in a set of candidate scheduling schemes, as follows: A comprehensive optimization objective is constructed based on the basic weights, risk penalty weights, and safety buffer durations of each spatiotemporal edge in the spatiotemporal scheduling network adjusted for confidence. Initialize the pheromone intensity parameters for each spatiotemporal edge in the optimizable part; At the beginning of each iteration, a number of ants are assigned to each vehicle in the optimizable part. The initial position of each ant is set as the starting spatiotemporal node of the vehicle in the spatiotemporal scheduling network. The frozen part of the task sequence of the vehicle in the initial scheduling scheme is solidified into the prefix path that the ants must follow. On this basis, the ants are only allowed to construct paths on the spatiotemporal nodes and spatiotemporal edges corresponding to the rolling zone and the prospect zone. After each ant completes the path construction from the starting spatiotemporal node to the vehicle's ending spatiotemporal node, the comprehensive optimization target value is calculated based on the constructed complete scheduling scheme. The comprehensive optimization target values ​​of all ants are sorted, and the candidate scheduling scheme with the smallest comprehensive optimization target value in this iteration is recorded. A set of candidate scheduling schemes is obtained by iteratively executing the path construction process.

4. The business intelligence processing method for a supply chain integrated service platform according to claim 3, characterized in that, During the path construction process, when an ant of a vehicle is located at the current spatiotemporal node, candidate spatiotemporal edges that would violate vehicle capacity constraints are eliminated based on the set of next spatiotemporal nodes reachable from the current spatiotemporal node. Transfer reference values ​​are calculated for the remaining candidate spatiotemporal edges, and the transfer probability of each candidate spatiotemporal edge is obtained through normalization processing. The next spatiotemporal node is selected in a probabilistic manner.

5. The business intelligence processing method for a supply chain integrated service platform according to claim 1, characterized in that, In the pheromone update phase, the improved ant colony algorithm performs pheromone evaporation processing on the spatiotemporal edges involved in path construction, reduces the pheromone intensity on each spatiotemporal edge according to a preset evaporation ratio, and sets differentiated pheromone adjustment strategies for the rolling zone and the prospect zone.

6. The business intelligence processing method for a supply chain integrated service platform according to claim 1, characterized in that, Based on the candidate scheduling scheme set, a neighborhood search algorithm based on tabu strategy is used to locally improve the scheduling of some vehicles in the rolling zone and the prospective zone, resulting in the target scheduling scheme, as follows: Based on the candidate scheduling scheme set, the candidate scheduling scheme with the smallest comprehensive optimization objective value is selected as the initial solution for neighborhood search, and the current scheduling solution is obtained according to the candidate scheduling scheme; Based on the current scheduling solution, identify the task segments that belong to the frozen part and the task segments that belong to the optimizable part. Mark the task segments corresponding to the frozen tasks as non-adjustable segments, and mark the task segments corresponding to the rolling tasks and the prospective tasks as adjustable segments. At the beginning of each iteration, based on the current scheduling solution, and according to the preset neighborhood size, one operation is selected from task swapping, task insertion, and path replacement to randomly perturb the tasks and paths in the adjustable segment, generating several neighborhood candidate scheduling solutions. Calculate the comprehensive optimization objective value for each neighborhood candidate scheduling solution, and select the neighborhood candidate scheduling solution with the smallest comprehensive optimization objective value as the preferred neighborhood solution in this iteration. Construct a tabu list and record the task swapping, task insertion, and path replacement operations performed during the iteration; In each iteration, it is determined whether the preferred neighborhood solution obtained in this round was generated by a taboo operation. If the corresponding operation is a taboo operation and the overall optimization objective value of the preferred neighborhood solution is not better than the currently known global optimal scheduling solution, then the preferred neighborhood solution is abandoned according to the taboo strategy, and a neighborhood solution that is not taboo and has the second best overall optimization objective value is selected as the new current scheduling solution. If the preferred neighborhood solution is generated by a taboo operation but the overall optimization objective value is better than the currently known global optimal scheduling solution, then the preferred neighborhood solution is updated as the new current scheduling solution and the new global optimal scheduling solution. After the iteration process is completed, the globally optimal scheduling solution is output as the target scheduling scheme.

7. A business intelligence processing system for a supply chain integrated service platform, used with the business intelligence processing method for a supply chain integrated service platform according to any one of claims 1 to 6, characterized in that, include: The data acquisition and feature construction module is used to collect various data from the port supply chain production system and preprocess them to obtain task feature vectors, road network feature vectors, and time feature vectors. The time series and spatiotemporal prediction module is used to obtain the predicted values ​​of task arrival, workload, task prediction confidence, expected travel time of road segments, road segment congestion probability, and road network prediction confidence based on task feature vectors, road network feature vectors, and time feature vectors. The temporal and spatial scheduling network construction module is used to construct a basic temporal and spatial scheduling network. Based on the task prediction confidence and the road network prediction confidence, the weights of the spatiotemporal edges of the basic temporal and spatial scheduling network are adjusted to obtain a confidence-adjusted temporal and spatial scheduling network. The initial scheduling generation module is used to divide the frozen area, rolling area and prospective area, and to assign tasks to each vehicle based on the time-space scheduling network adjusted by confidence, and to construct the initial scheduling scheme. The ant colony global optimization module is used to take the initial scheduling scheme as the initial solution, and based on the confidence-adjusted time-space scheduling network, adopts an improved ant colony algorithm to search the task sequence and driving path of each vehicle in the rolling zone and the prospect zone to obtain a set of candidate scheduling schemes. The neighborhood search local optimization module is used to improve the scheduling of some vehicles in the rolling zone and the prospect zone based on the set of candidate scheduling schemes and the neighborhood search algorithm based on the tabu strategy, so as to obtain the target scheduling scheme. The scheme execution module is used to execute vehicle scheduling according to the target scheduling scheme.

Images