Port transportation network operation index prediction method based on big data mining

Abstract

The present application relates to the technical field of port intelligent management, and discloses a port transportation network operation index prediction method based on big data mining, comprising the following steps: obtaining and preprocessing port multi-source operation data to form a historical state sequence of port nodes; constructing port transportation network structure data and inputting a DCRNN model; separating time memory update and space diffusion reception in a DCGRU unit to obtain an improved DCRNN model; predicting future port node operation states; loading the prediction results to a digital twin port scene for future state playback to obtain operation pressure evolution results; calculating future port transportation network operation indexes according to the operation pressure evolution results, and determining index increase reason chains through single-factor recovery playback; inputting the indexes and reason chains into a large language model to generate operation index prediction sentences. The present application realizes accurate prediction of port operation indexes and is suitable for intelligent port scheduling and safety control scenes.

Description

Technical Field

[0001] This invention relates to the field of port intelligent management technology, and in particular to a method for predicting the port transportation network operation index based on big data mining. Background Technology

[0002] With the development of intelligent port construction, port management platforms typically collect port operation data through video surveillance, BeiDou positioning, ship AIS, gate identification, equipment monitoring, operation management and hidden danger management, and use GIS maps, 3D visualization interfaces and digital twin scenes to display the status of personnel, vehicles, ships, equipment and operations.

[0003] Current technologies largely focus on data display, alarm recording, and post-event statistics, lacking sufficient depth in exploring the spatiotemporal correlations between multi-source operational data in port areas. Within the port transportation network, different port area nodes are interconnected through vehicle traffic, cargo flow, vessel berthing, operational coordination, and risk transmission. Anomalies at a single node can propagate continuously along the business chain. Existing forecasting methods typically generate operational indices directly based on historical metrics, making it difficult to distinguish between continuous changes within a single node and the impact of transmission from related nodes. This can easily lead to misjudgments of local fluctuations or omissions of transmitted risks.

[0004] Current port operation index predictions typically output only numerical values ​​or levels, lacking integration with future state replays of digital twin port area scenarios, making it difficult to present the evolution of operational pressure. Regarding the causes of index increases, existing methods largely rely on the number of alarms, human experience, or simple rules for judgment, lacking a process of verifying the impact of key factors through replay analysis, resulting in insufficient interpretability of the prediction results.

[0005] Therefore, how to provide a method for predicting the port transportation network operation index based on big data mining is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a method for predicting the port transportation network operation index based on big data mining. This invention utilizes an improved DCRNN model, digital twin playback, and counterfactual tracing technology to perform spatiotemporal prediction, state playback, and cause analysis on multi-source operation data of the port area. It has the advantages of high prediction accuracy, clear pressure transmission process, and traceable causes of index increases.

[0007] The port transportation network operation index prediction method based on big data mining according to an embodiment of the present invention includes the following steps: Acquire multi-source port operation data, perform preprocessing on the multi-source port operation data, and obtain the historical state sequence of port area nodes; Based on the transportation relationships between key spatial units in the port area, port transportation network structure data is constructed, and the port transportation network structure data and the historical state sequence of port area nodes are input into the DCRNN model. In the DCGRU unit of the DCRNN model, a time memory update channel and a spatial diffusion reception channel are set up to extract the continuous running trend of a single node, receive the influence of related nodes, and perform separate calculations based on the node's own running trend and external transmission. Predict the operational status of port area nodes within multiple future time slices based on the improved DCRNN model; The predicted port area node operation status is loaded into the digital twin port area scenario, and the future state is replayed according to the port area business transmission relationship. The port area operation pressure evolution result is determined based on the replay process. Based on the evolution of port area operation pressure, calculate the future port transportation network operation index, replicate the digital twin port area scenario to construct a replay scenario, perform single-factor recovery replay on the key factors that cause the index to rise, and determine the cause chain of the index rise based on the changes in the index and the changes in the transmission path before and after the recovery. Input the cause chain of the index increase and the future port transportation network operation index into the large language model, and output the port transportation network operation index prediction statement.

[0008] Optionally, the preprocessing of the port's multi-source operational data specifically includes: Collect multi-source port operation data, divide the multi-source port operation data into time segments according to the preset sampling time, delete duplicate and abnormal records, fill in missing fields, map data with spatial location to the corresponding port area nodes, perform standardization processing on data fields with different dimensions, and arrange the processed multi-source port operation data according to the port area node number and time slice order to obtain the historical state sequence of port area nodes.

[0009] Optionally, the construction of port transportation network structure data based on the transportation relationships between key spatial units in the port area specifically includes: Key spatial units in the port area are designated as port area nodes, and the comprehensive transportation impact relationship between port area nodes is determined based on vehicle traffic routes, ship entry and exit routes, cargo flow routes, operation connection routes, and risk transmission routes. Directed connection edges between port area nodes are constructed based on the transmission direction of the comprehensive transportation impact relationship, and the connection strength of the directed connection edges is determined based on the frequency of passage, circulation, number of operation connections and risk transmission between port area nodes. Port transportation network structure data is generated according to the port area node number, the direction of the directed connection edge, and the connection strength; The port transportation network structure data is used as the spatial diffusion input of the DCRNN model, and the historical state sequence of port nodes is used as the time series input of the DCRNN model.

[0010] Optionally, the time memory update channel specifically includes: Read the node state vector of the same port area node in consecutive time slices from the historical state sequence of port area nodes. Use the current time slice node state vector, the previous time slice node state vector and the previous time slice node memory state as input to the time memory update channel. Perform differential processing on the current time slice node state vector and the previous time slice node state vector to obtain the node state change. Based on the changes in node state, determine whether the node state is continuously strengthening, continuously weakening, or remaining stable, and generate a state change identifier. Based on the state change identifier and the memory state of the previous time slice node, a memory retention coefficient and a memory update coefficient are generated. The part of the memory state of the previous time slice node that is consistent with the continuous trend is retained according to the memory retention coefficient. The part of the current time slice node state vector that corresponds to the continuous trend change is written according to the memory update coefficient, so as to obtain the memory state of the current time slice node. The memory state of the current time slice node obtained within a continuous time slice is output in chronological order as the continuous running trend of a single node.

[0011] Optionally, the spatial diffusion receiving channel specifically includes: Read the incoming directed edges of the target port area nodes, obtain the single node continuous operation trend and connection strength of the associated nodes based on the incoming directed edges, multiply the single node continuous operation trend of the associated nodes by the corresponding connection strength and sum them to obtain the transmission influence of the associated nodes. The influence of associated nodes is the change in external state that is transmitted from the associated node to the target port area node through the incoming directed connection edge; The continuous operating trend of a single node in the target port area is taken as the node's own operating trend, and the influence of related nodes is taken as external transmission. The consistency of the trend direction between the node's own operating trend and external transmission is judged. When the node's own operating trend is consistent with the trend of external transmission, the external transmission is incorporated into the node's own operating trend to obtain the diffusion memory state of the target port area node. When the node's own operating trend is inconsistent with the external transmission trend, the external transmission write is blocked, and the node's own operating trend is used as the diffusion memory state of the target port area node.

[0012] Optionally, the prediction of the port area node's operating status within multiple future time slices based on the improved DCRNN model specifically includes: The diffusion memory state of the target port area node is input into the decoding unit of the improved DCRNN model. The port area node operation state for the next time slice is generated according to the preset prediction step size. The port area node operation state for the next time slice is used as the subsequent prediction input for iterative decoding until the preset prediction time length is reached, so as to obtain the port area node operation state for multiple future time slices.

[0013] Optionally, the execution of future state replay according to the port area business transmission relationship specifically includes: A digital twin port area scenario is constructed based on port area spatial data and port area business object data, and a mapping relationship is established between port area nodes and digital twin port area scenario objects. The operational status of port area nodes within multiple future time slices is loaded into the corresponding digital twin port area scene object according to the port area node number and the predicted time slice. The operational status of the digital twin port area scene object is updated frame by frame according to the order of the predicted time slices. The pressure level of each port node in each predicted time slot is determined based on the degree of deviation between the port node's operating status and the preset normal operating threshold. Based on the comparison of the pressure level changes of related port area nodes within adjacent forecast time slots according to the port area business transmission relationship, the pressure transmission relationship is recorded. Based on the pressure level, pressure transmission relationship, pressure duration, and number of affected port area nodes, the evolution result of port area operational pressure is formed.

[0014] Optionally, the calculation of the future port transportation network operation index based on the port area operation pressure evolution results specifically includes: The future port transportation network operation index is calculated based on the port area operation pressure evolution results, including converting the pressure level of port area nodes in each forecast time segment into node pressure values; The conduction amplification factor is determined based on the continuous pressure transmission relationship, the duration correction factor is determined based on the pressure duration, and the range correction factor is determined based on the number of affected port area nodes. After summarizing the node pressure values ​​of each port area node, the values ​​are normalized by combining the transmission amplification coefficient, the continuous correction coefficient, and the range correction coefficient to obtain the future port transportation network operation index.

[0015] Optionally, the determination of the causal chain of the index increase based on the changes in the index before and after recovery and the changes in the transmission path specifically includes: The digital twin port area scene that has completed the replay of the future state is replicated, retaining the predicted time slice, port area node mapping relationship and port area business transmission relationship, and constructing a recovery replay scene corresponding to the original replay process; From the port area operation pressure evolution results, select port area nodes whose pressure level reaches the restricted or blocked as nodes to be restored, adjust the port area node operation status of the nodes to be restored in the recovery playback scenario to the normal operation threshold range, and keep the operation status of other port area nodes unchanged. Based on the adjusted recovery replay scenario, the future state replay is re-executed and the port transportation network operation index is recalculated. The changes in the port transportation network operation index and the changes in pressure transmission relationship before and after the recovery are compared. The node to be recovered where the index drops to a preset threshold and the pressure transmission relationship is interrupted is identified as the node causing the index increase. The index increase cause chain is generated according to the transmission sequence of the index increase cause node in the original playback process.

[0016] Optionally, the output port transportation network operation index prediction statement specifically includes: The future port transportation network operation index, the corresponding prediction time slice, the chain of reasons for the index increase, the nodes of the reasons for the index increase, and the pressure transmission relationship are compiled into a structured prompt text; The structured prompt text is input into the large language model, which then generates a prediction statement for the port transportation network operation index according to a preset statement template. The port transportation network operation index prediction statement should include at least the prediction time, the future port transportation network operation index, the operation status level, the chain of reasons for the index increase, and the handling prompts.

[0017] The beneficial effects of this invention are: (1) The improved DCRNN model handles the continuous operating trend of the port area nodes themselves and the transmission influence of related nodes respectively, which can reduce the interference of local short-term fluctuations on the overall prediction results and improve the stability of the port transportation network operation index prediction.

[0018] (2) Load the predicted port area node operation status into the digital twin port area scenario and execute the future state playback according to the port area business transmission relationship. This can reflect the formation, continuity and transmission process of operation pressure and improve the matching degree between the operation index and the actual operation process of the port area.

[0019] (3) By performing single-factor recovery playback of the restored playback scenario, the cause chain of the index increase is determined based on the changes in the index and the changes in the transmission path before and after recovery. The prediction statement is generated by combining the large language model to improve the interpretability of the prediction results and the availability of management decisions. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1This is a flowchart of the port transportation network operation index prediction method based on big data mining proposed in this invention; Figure 2 This is a digital twin future state playback diagram of the port transportation network operation index prediction method based on big data mining proposed in this invention. Figure 3 This is a counterfactual tracing diagram of the port transportation network operation index prediction method based on big data mining proposed in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figures 1-3 The method for predicting the port transportation network operation index based on big data mining includes the following steps: Acquire multi-source port operation data, perform preprocessing on the multi-source port operation data, and obtain the historical state sequence of port area nodes; Based on the transportation relationships between key spatial units in the port area, port transportation network structure data is constructed, and the port transportation network structure data and the historical state sequence of port area nodes are input into the DCRNN model. In the DCGRU unit of the DCRNN model, a time memory update channel and a spatial diffusion reception channel are set up to extract the continuous running trend of a single node, receive the influence of related nodes, and perform separate calculations based on the node's own running trend and external transmission. Predict the operational status of port area nodes within multiple future time slices based on the improved DCRNN model; The predicted port area node operation status is loaded into the digital twin port area scenario, and the future state is replayed according to the port area business transmission relationship. The port area operation pressure evolution result is determined based on the replay process. Based on the evolution of port area operation pressure, calculate the future port transportation network operation index, replicate the digital twin port area scenario to construct a replay scenario, perform single-factor recovery replay on the key factors that cause the index to rise, and determine the cause chain of the index rise based on the changes in the index and the changes in the transmission path before and after the recovery. Input the cause chain of the index increase and the future port transportation network operation index into the large language model, and output the port transportation network operation index prediction statement.

[0023] In this embodiment, the preprocessing of multi-source port operation data specifically includes: Collect multi-source port operation data, divide the multi-source port operation data into time segments according to the preset sampling time, delete duplicate and abnormal records, fill in missing fields, map data with spatial location to the corresponding port area nodes, perform standardization processing on data fields with different dimensions, and arrange the processed multi-source port operation data according to the port area node number and time slice order to obtain the historical state sequence of port area nodes.

[0024] In this embodiment, the port's multi-source operational data includes personnel BeiDou positioning data, vehicle gate passage data, port area road passage data, ship AIS positioning data, berth occupancy data, equipment operation data, operational task data, hazard handling data, meteorological monitoring data, and AI visual recognition data. The AI ​​visual recognition data is output by recognition models deployed in port area video surveillance, drone inspection footage, and mobile surveillance cameras. It includes at least the following data: identification data of vehicles illegally entering the front edge, identification data of drivers illegally getting out of their vehicles, identification data of small fishing boats illegally entering the port basin, identification data of personnel entering the electronic fence, and identification data of personnel staying for extended periods. Each piece of AI visual recognition data includes alarm type, alarm time, alarm location, target number, and recognition confidence level. During preprocessing, the port's multi-source operational data is divided into continuous time slices with a 5-minute sampling interval. For data within the same time slice, statistics are performed on counting, mean, maximum, and duration. Records of the same target repeatedly generated within the same time slice are deleted, as are AI visual recognition data with missing timestamps, spatial coordinates exceeding port area boundaries, empty target numbers, and recognition confidence scores below 0.80. Missing fields are filled with valid values ​​from the previous time slice; data that cannot be filled are marked as missing. Personnel positioning coordinates, vehicle positioning coordinates, ship AIS coordinates, drone alarm coordinates, and video surveillance point coordinates are uniformly converted to the port area coordinate system, and port area node numbers are determined based on the port area node boundaries into which the coordinates fall. For data without real-time coordinates, mapping is completed based on the binding relationship between equipment number, camera number, gate number, berth number, and port area node number. Maximum and minimum normalization are performed on quantitative fields, and standard deviation standardization is performed on speed, waiting time, occupancy time, and meteorological values.

[0025] In this embodiment, constructing port transportation network structure data based on the transportation relationships between key spatial units in the port area specifically includes: Key spatial units within the port area are designated as port area nodes. The comprehensive transportation impact relationships between these nodes are determined based on vehicle traffic routes, vessel entry and exit routes, cargo transfer routes, operational connection routes, and risk transmission routes. Key spatial units are defined by artificial spatial zoning on the port area's electronic map, with each node possessing a unique node number and spatial boundaries. Vehicle traffic routes are determined by vehicle positioning trajectories and gate passage records; vessel entry and exit routes are determined by vessel AIS trajectories and berthing / departure records; cargo transfer routes are determined by loading / unloading operation records and yard transfer records; operational connection routes are determined by the preceding and following processes of the operational tasks; and risk transmission routes are determined by historical alarm locations, hazard locations, and areas affected by handling. When there are records of target movement, cargo transfer, operational handover, or risk impact between two port area nodes within the same time window, a comprehensive transportation impact relationship is established between the two nodes. Directed connections between port nodes are constructed based on the transmission direction of the comprehensive transportation impact relationship. The connection strength of the directed connections is determined based on the frequency of passage, flow, number of operation connections, and number of risk transmissions between port nodes. The direction of the directed connections is determined according to the target movement direction, cargo flow direction, operation sequence, and risk impact expansion direction. The connection strength is calculated within a preset statistical period, which can be set to 7 days. The frequency of passage is the number of times vehicles and ships enter the terminal port node from the starting port node. The frequency of flow is the number of times cargo is transferred from the starting port node to the terminal port node. The number of operation connections is the number of times the next operation node starts after the previous operation node is completed. The number of risk transmissions is the number of times adjacent port nodes generate related alarms or have their impacts handled within a preset time interval after an alarm or potential hazard is generated at a port node. The connection strength is normalized to a value between 0 and 1. Port transportation network structure data is generated according to the port area node number, the direction of the directed connection edge, and the connection strength. The port transportation network structure data is represented by a directed adjacency matrix. The row number of the directed adjacency matrix represents the starting port area node, the column number represents the ending port area node, and the matrix element represents the connection strength from the starting port area node to the ending port area node. When there is no comprehensive transportation influence relationship between two port area nodes, the corresponding matrix element is set to 0. When there is a comprehensive transportation influence relationship between two port area nodes, the corresponding matrix element is set to the normalized connection strength. The port transportation network structure data is used as the spatial diffusion input of the DCRNN model, and the historical state sequence of port nodes is used as the time series input of the DCRNN model. The DCRNN model receives the directed adjacency matrix and the historical state sequence of port nodes. The port transportation network structure data does not encode the relationship type separately. The impact of vehicle passage, ship entry and exit, cargo flow, operation connection and risk transmission is uniformly converted into the connection strength of the directed connection edge.

[0026] In this embodiment, the time memory update channel specifically includes: Read the node state vector of the same port area node in consecutive time slices from the historical state sequence of port area nodes. Use the current time slice node state vector, the previous time slice node state vector and the previous time slice node memory state as input to the time memory update channel. Perform differential processing on the current time slice node state vector and the previous time slice node state vector to obtain the node state change. Based on the changes in node state, the system determines whether the node state is continuously increasing, continuously decreasing, or stable, and generates a state change identifier. A change threshold is set for each dimension of the operational characteristic in the node state change data; this threshold can be set to 10% of the historical standard deviation of the corresponding operational characteristic. When the node state change data for three consecutive time slices is greater than the change threshold, the corresponding operational characteristic is marked as continuously increasing. When the node state change data for three consecutive time slices is less than the negative change threshold, the corresponding operational characteristic is marked as continuously decreasing. When the absolute value of the node state change data for three consecutive time slices does not exceed the change threshold, the corresponding operational characteristic is marked as stable. Based on the state change identifier and the memory state of the previous time slice node, a memory retention coefficient and a memory update coefficient are generated. The portion of the previous time slice node's memory state consistent with the continuous trend is retained according to the memory retention coefficient, and the portion corresponding to the continuous trend change is written into the current time slice node's state vector according to the memory update coefficient, thus obtaining the current time slice node's memory state. The memory retention coefficient and memory update coefficient are vectors ranging from 0 to 1, with the same vector dimension as the node's state vector. When the state change identifier is continuous enhancement or continuous weakening, the memory update coefficient of the corresponding dimension is increased and the memory retention coefficient of the corresponding dimension is decreased. When the state change identifier is stable maintenance, the memory retention coefficient of the corresponding dimension is increased and the memory update coefficient of the corresponding dimension is decreased. The current time slice... The node memory state is obtained by combining the retained portion of the node memory state from the previous time slice and the written portion of the node state vector from the current time slice. In this embodiment, the memory update baseline value is first determined based on the state change identifier. When the state change identifier is continuously increasing or continuously decreasing, the memory update baseline value for the corresponding dimension is set to 0.70. When the state change identifier is stable, the memory update baseline value for the corresponding dimension is set to 0.30. Then, the node state change amount and the node memory state from the previous time slice are input into the linear mapping layer, and a fine-tuning coefficient with a value range of 0 to 1 is generated through the Sigmoid gating function. The memory update baseline value and the fine-tuning coefficient are fused according to a preset ratio to obtain the memory update coefficient. The memory retention coefficient is obtained by subtracting the memory update coefficient from 1. The memory state of the current time slice node obtained within a continuous time slice is output in chronological order as the continuous running trend of a single node.

[0027] In this embodiment, the spatial diffusion receiving channel specifically includes: Read the incoming directed edges of the target port area node, obtain the single-node continuous operation trend and connection strength of the associated node based on the incoming directed edges, multiply the single-node continuous operation trend of the associated node by the corresponding connection strength and sum them to obtain the transmission influence of the associated node; the incoming directed edges are the directed edges in the port transportation network structure data whose terminating node is the target port area node, and the associated nodes are the starting nodes of the incoming directed edges; when there are multiple associated nodes in the target port area node, read the single-node continuous operation trend and corresponding connection strength of each associated node in the same time slice, multiply the single-node continuous operation trend of each associated node by the corresponding connection strength in each dimension and sum them to obtain the transmission influence of the associated node with the same dimension as the single-node continuous operation trend; The influence of the associated node is the external state change that is transmitted from the associated node to the target port area node through the incoming directed connection edge. The external state change is used to represent the transmission state of the associated node to the target port area node caused by changes in vehicle queuing, ship waiting, yard congestion, berth occupancy, operation delay, equipment abnormality and safety alarm. The external state change does not directly replace the continuous operation trend of a single node in the target port area. The continuous operational trend of a single node in the target port area is taken as the node's own operational trend, and the influence transmitted by related nodes is taken as external transmission. A consistency judgment is made on the trend direction of the node's own operational trend and external transmission. This consistency judgment is performed separately for each dimension of the node's state vector. When the node's own operational trend and external transmission have the same sign on the same operational feature dimension, and both absolute values ​​are greater than the change threshold corresponding to that operational feature, the trend direction is determined to be consistent in that dimension. When the signs are opposite, or either absolute value is not greater than the change threshold corresponding to that operational feature, the trend direction is determined to be inconsistent in that dimension. The change threshold can be set to 10% of the historical standard deviation of the corresponding operational feature. When the node's own operating trend is consistent with the trend direction of external transmission, the external transmission is incorporated into the node's own operating trend to obtain the diffusion memory state of the target port area node; for operating feature dimensions with consistent trend directions, the node's own operating trend and the external transmission are added together to obtain the diffusion memory state of that dimension. When the node's own operating trend is inconsistent with the external transmission trend, the external transmission write is blocked, and the node's own operating trend is used as the diffusion memory state of the target port area node. Blocking the external transmission write means that the external transmission is not written into the diffusion memory state in the corresponding operating feature dimension, and the single node continuous operating trend of the target port area node is preserved.

[0028] In this embodiment, predicting the operating status of port area nodes within multiple future time slices based on the improved DCRNN model specifically includes: The diffused memory state of the target port area node is input into the decoding unit of the improved DCRNN model. The port area node operation state for the next time slice is generated according to the preset prediction step size. The port area node operation state for the next time slice is used as the input for subsequent predictions and the decoding is repeated until the preset prediction time length is reached to obtain the port area node operation state for multiple future time slices. When the preset prediction step size is 10 minutes and the preset prediction time length is 60 minutes, the number of iterations is 6. The port area node operation state values ​​include vehicle passage status value, ship waiting status value, berth occupancy status value, yard turnover status value, equipment operation status value, operation status value, and safety alarm status value.

[0029] In this implementation method, the execution of future state replay according to the business transmission relationship in the port area specifically includes: A digital twin port area scenario is constructed based on port area spatial foundation data and port area business object data, establishing a mapping relationship between port area nodes and digital twin port area scenario objects. The port area spatial foundation data can include port area electronic maps and road topology data; the port area business object data can include the numbering and location data of personnel, vehicles, ships, equipment, work tasks, alarm points, and hidden danger points; the mapping relationship is formed by binding port area node numbers with digital twin port area scenario object numbers, so that each port area node corresponds to a spatial area or set of business objects in the digital twin port area scenario. The operational status of port area nodes within multiple future time slices is loaded into the corresponding digital twin port area scene objects according to the port area node number and the predicted time slice. The operational status of the digital twin port area scene objects is updated frame by frame according to the order of the predicted time slices. The predicted time slices can be in 10-minute intervals. Frame by frame update means that the operational status of the port area nodes corresponding to each predicted time slice is written into the digital twin port area scene object according to the order of the predicted time slices, and the display status and business status of the corresponding objects are refreshed. The operational status of port area nodes can be mapped to color, identifier, quantity, speed, occupancy duration, alarm intensity, and work load value. The pressure level of each port node in each predicted time slot is determined based on the deviation between its operating status and the preset normal operating threshold. The preset normal operating threshold is determined based on historical normal operating samples of the port nodes. When the deviation of the port node's operating status from the preset normal operating threshold does not exceed 10%, the pressure level is set to normal; when the deviation is greater than 10% but not more than 30%, the pressure level is set to busy; when the deviation is greater than 30% but not more than 60%, the pressure level is set to restricted; and when the deviation is greater than 60%, the pressure level is set to blocked. The pressure transmission relationship is recorded by comparing the pressure level changes of related port area nodes in adjacent prediction time slots according to the port area business transmission relationship. The port area business transmission relationship is consistent with the direction of the directed connection edges in the port transportation network structure data. When the pressure level of the starting port area node increases in the previous prediction time slot and the pressure level of the ending port area node increases in the subsequent prediction time slot, the pressure transmission relationship from the starting port area node to the ending port area node is recorded. When the pressure level increases in the same direction in multiple consecutive prediction time slots, the pressure transmission relationships in the same direction are merged into a continuous pressure transmission relationship. Based on the pressure level, pressure transmission relationship, pressure duration, and number of affected port area nodes, the port area operation pressure evolution result is formed; the pressure duration is the product of the number of consecutive predicted time slots in which the pressure level of the same port area node reaches busy, restricted, or blocked, and the interval between predicted time slots; the number of affected port area nodes is the total number of port area nodes in which the pressure level reaches busy, restricted, or blocked within the predicted time range; the port area operation pressure evolution result includes the pressure level distribution, continuous pressure transmission relationship, pressure duration, and number of affected port area nodes in each predicted time slot.

[0030] In this embodiment, the calculation of the future port transportation network operation index based on the port area operation pressure evolution results specifically includes: The future port transportation network operation index is calculated based on the port area operation pressure evolution results, including converting the pressure level of port area nodes in each prediction time slot into node pressure value; normal corresponds to node pressure value of 0, busy corresponds to node pressure value of 40, restricted corresponds to node pressure value of 70, and congested corresponds to node pressure value of 100; for all port area nodes in the same prediction time slot, the average node pressure value is calculated according to the node pressure value, which is used as the base pressure value for the corresponding prediction time slot. The conduction amplification factor is determined based on the continuous pressure conduction relationship, the duration correction factor is determined based on the pressure duration, and the range correction factor is determined based on the number of affected port area nodes. The conduction amplification factor is determined based on the number of continuous pressure conduction relationships: it is set to 1 when there are no continuous pressure conduction relationships, 1.1 when there are 1 to 2 continuous pressure conduction relationships, and 1.2 when there are 3 or more continuous pressure conduction relationships. The duration correction factor is determined based on the pressure duration: it is set to 1 when the pressure duration is less than 10 minutes, between 1 and 1.2 when the pressure duration is between 10 and 30 minutes, and 1.2 when the pressure duration is greater than 30 minutes. The range correction factor is determined based on the proportion of the number of affected port area nodes to the total number of port area nodes: it is set to 1 when the proportion is less than 10%, between 1 and 1.2 when the proportion is between 10% and 30%, and 1.2 when the proportion is greater than 30%. After summarizing the node pressure values ​​of each port area node, normalization is performed using the transmission amplification coefficient, the persistence correction coefficient, and the range correction coefficient to obtain the future port transportation network operation index. In this embodiment, the base pressure value for each prediction time slice is first calculated, and then the base pressure value is multiplied by the transmission amplification coefficient, the persistence correction coefficient, and the range correction coefficient respectively to obtain the prediction time slice pressure value. When the prediction time slice pressure value is greater than 100, the prediction time slice pressure value is truncated to 100. The prediction time slice pressure values ​​of multiple future prediction time slices are output in chronological order as the future port transportation network operation index sequence. When a single future port transportation network operation index needs to be output, the maximum value among the multiple future prediction time slice pressure values ​​is taken as the future port transportation network operation index.

[0031] In this embodiment, the chain of causes for the increase in the index, determined based on changes in the index before and after recovery and changes in the transmission path, specifically includes: The digital twin port area scene that has completed the replay of the future state is replicated, retaining the prediction time slice, port area node mapping relationship and port area business transmission relationship, and constructing a recovery replay scene corresponding to the original replay process; the recovery replay scene and the original replay process use the same prediction time range, prediction time slice interval, port area node number, digital twin port area scene object number and directed transmission relationship, so that the recovery replay scene and the original replay process differ only in the port area node running status of the node to be recovered; From the port area operational pressure evolution results, port area nodes with pressure levels reaching restricted or blocked are selected as nodes to be restored. The operational status of the nodes to be restored in the recovery playback scenario is adjusted to within the normal operation threshold range, while keeping the operational status of other port area nodes unchanged. The nodes to be restored are sorted from high to low pressure level, and when the pressure levels are the same, they are sorted from long to short pressure duration. Only one node to be restored is adjusted in a single recovery playback. The normal operation threshold range adopts the preset normal operation threshold, and the operational status of the port area nodes in the prediction time slice corresponding to the node to be restored is adjusted to a value that does not exceed the preset normal operation threshold. Based on the adjusted recovery replay scenario, the future state replay is re-executed and the port transportation network operation index is recalculated. The changes in the port transportation network operation index and the pressure transmission relationship before and after the recovery are compared. When re-executing the future state replay, the same port area business transmission relationship and index calculation method as the original replay process are used. The change in the port transportation network operation index is the port transportation network operation index obtained in the original replay process minus the port transportation network operation index obtained in the recovery replay scenario. Changes in pressure transmission relationships include a reduction in the number of pressure transmission relationships, changes in the pressure transmission start node, changes in the pressure transmission end node, and interruption of continuous pressure transmission relationships. Nodes to be recovered where the index decreases to a preset threshold and the pressure transmission relationship is interrupted are identified as nodes causing the index increase. An index increase cause chain is generated according to the transmission sequence of these nodes in the original playback process. The preset threshold can be set to 10 index points. An interrupted pressure transmission relationship means that at least one continuous pressure transmission relationship following the node to be recovered in the restored playback scene no longer occurs. When multiple nodes to be recovered meet the index increase cause node condition, they are arranged according to the pressure transmission sequence in the original playback process. When multiple nodes to be recovered are in the same transmission sequence, they are arranged according to the corresponding index decrease from largest to smallest, generating an index increase cause chain.

[0032] In this embodiment, the output port transportation network operation index prediction statement specifically includes: The future port transportation network operation index, the corresponding prediction time slice, the chain of reasons for the index increase, the nodes of the reasons for the index increase, and the pressure transmission relationship are compiled into a structured prompt text; The structured prompt text is input into the large language model, which generates a port transportation network operation index prediction statement according to a preset statement template. The preset statement template includes a time description field, an index description field, a status judgment field, a cause chain description field, and a handling prompt field. The large language model only organizes the statements based on the field content in the structured prompt text. The port transportation network operation index prediction statement must include at least the prediction time, the future port transportation network operation index, the operation status level, the cause chain of the index increase, and the handling prompts. The handling prompts are obtained through a retrieval-enhanced generation method. The large language model retrieves matching text from a pre-set port safety management knowledge base, operation specification base, and emergency response plan base based on the cause nodes, operation status level, and pressure transmission relationship in the cause chain of the index increase. The retrieved matching text and the structured prompt text are input into the large language model to generate the handling prompts. The handling prompts are only an auxiliary expression of the port transportation network operation index prediction statement.

[0033] Example 1: To verify the feasibility of this invention in practice, it was applied to a scenario of predicting the operation index of a transportation network in a comprehensive port area. This port area simultaneously experiences various operational activities, including vehicle entry and exit from gates, ship berthing and departure, cargo flow in the yard, quay crane operations, personnel inspections, AI visual alarms, and hazard handling. In actual operation, vehicle queuing, berth occupancy, equipment malfunctions, or safety alarms do not exist in isolation. Congestion at one node often propagates along roads, berths, yards, and operational chains to other nodes. Existing methods often calculate the operation index based on historical alarm counts or single business indicators, making it difficult to predict the direction of pressure transmission in advance or explain the reasons for index increases. Therefore, problems such as prediction lag, misjudgment of high-risk states, and unclear cause analysis are prone to occur.

[0034] In this embodiment, multi-source port operation data is collected, including vehicle traffic data, vessel AIS data, berth occupancy data, equipment operation data, work task data, personnel positioning data, AI visual alarm data, hazard handling data, and meteorological monitoring data. The system divides the above data according to a unified prediction time slice, processes duplicate data, missing fields, and abnormal coordinates, and maps vehicles, vessels, equipment, personnel, alarm points, and hazard points to corresponding port area nodes, forming a historical state sequence of port area nodes. The historical state sequence of port area nodes is used to describe the operational changes of each node within a continuous time slice, including the number of vehicles, average speed, vessel waiting status, berth occupancy status, equipment status, work load, number of safety alarms, and hazard status.

[0035] Port transportation network structure data is constructed based on the transportation relationships between key spatial units within the port area. The relationships between port nodes are jointly determined by vehicle traffic paths, vessel entry and exit paths, cargo transfer paths, operational connection paths, and risk transmission paths. Directed connections are generated based on the transmission direction, and connection strength is determined based on traffic frequency, cargo transfer frequency, number of operational connections, and number of risk transmissions. This port transportation network structure data reflects the likelihood and intensity of operational pressure from one node to another, rather than treating port nodes as isolated objects.

[0036] In the model application process, the historical state sequence of port area nodes and the port transportation network structure data are input into the DCRNN model. A time-memory update channel and a spatial diffusion reception channel are set up in the DCGRU unit. The time-memory update channel mainly extracts the continuous operating trend of a single node to determine whether a node is continuously busy, continuously easing, or remaining stable. The spatial diffusion reception channel reads the incoming connection edges of the target node, receives the influence transmitted by related nodes, and performs separate calculations from the target node's own operating trend. When the direction of the influence transmitted by related nodes is consistent with the target node's own trend, the external transmission is incorporated into the target node's state; when the two directions are inconsistent, the external transmission is blocked to avoid irrelevant fluctuations affecting the prediction results.

[0037] The improved DCRNN model outputs the operational status of port nodes over multiple future time slices. The prediction results are loaded into a digital twin port scenario. The system updates the operational status of vehicles, ships, berths, equipment, tasks, and alarm objects frame-by-frame according to the predicted time slices, and determines the pressure level based on the deviation between the predicted status of each node and the normal operating threshold. Pressure levels include normal, busy, restricted, and congested. The operational index is derived from the pressure level: normal corresponds to low pressure, busy to medium pressure, restricted to high pressure, and congested to the highest pressure. This index is then adjusted based on pressure transmission relationships, pressure duration, and the range of affected nodes, and finally normalized to form the future port transportation network operational index. A higher index indicates greater operational pressure on the future port transportation network.

[0038] The system replicates the digital twin port area scenario to construct a recovery replay scenario. Nodes with pressure levels reaching restricted or congested levels are sequentially restored to normal operating thresholds, while maintaining the operating status of other nodes. The future state replay is then re-executed, and the operating index is recalculated. If the operating index of a node significantly decreases after recovery, and the original pressure transmission path is interrupted, that node is identified as the cause of the index increase. Multiple cause nodes form an index increase cause chain according to the transmission sequence in the original replay process. Finally, the system organizes the future port transportation network operating index, predicted time slice, operating status level, pressure transmission relationship, and index increase cause chain into a structured prompt text, which is then input into a large language model to generate an operating index prediction statement for management personnel.

[0039] Table 1: Comparison of Prediction Results of Port Transportation Network Operation Index

[0040] As shown in Table 1, the method of this invention outperforms traditional statistical methods, LSTM models, and the standard DCRNN model in all five indicators. Regarding the operational index error, the method of this invention has an error of 3.1 index points, significantly lower than the 5.2 index points of the standard DCRNN model. This indicates that the method of this invention reduces errors caused by irrelevant spatial diffusion by separating the node's own operational trend from the influence of related nodes, making the prediction results closer to the actual operational index. In terms of the accuracy rate of high-risk state identification, the method of this invention reaches 93.5%, higher than the 87.6% of the standard DCRNN model, indicating that the method of this invention can more accurately identify busy, restricted, and congested states. Regarding the high-risk state false alarm rate, the method of this invention reduces it to 4.6%, lower than the 9.8% of the standard DCRNN model, indicating that the method of this invention has a stronger ability to detect potential operational pressure. Regarding the average early warning time, the method of this invention reaches 39 minutes, providing more sufficient response time for scheduling and safety management. Regarding the hit rate of the cause of index increase, the method of this invention reaches 88.4%, indicating that single-factor recovery replay can improve the credibility of the cause chain judgment.

[0041] 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 method for predicting the port transportation network operation index based on big data mining, characterized in that, Includes the following steps: Acquire multi-source port operation data, perform preprocessing on the multi-source port operation data, and obtain the historical state sequence of port area nodes; Based on the transportation relationships between key spatial units in the port area, port transportation network structure data is constructed, and the port transportation network structure data and the historical state sequence of port area nodes are input into the DCRNN model. In the DCGRU unit of the DCRNN model, a time memory update channel and a spatial diffusion reception channel are set up to extract the continuous running trend of a single node, receive the influence of related nodes, and perform separate calculations based on the node's own running trend and external transmission. Predict the operational status of port area nodes within multiple future time slices based on the improved DCRNN model; The predicted port area node operation status is loaded into the digital twin port area scenario, and the future state is replayed according to the port area business transmission relationship. The port area operation pressure evolution result is determined based on the replay process. Based on the evolution of port area operation pressure, calculate the future port transportation network operation index, replicate the digital twin port area scenario to construct a replay scenario, perform single-factor recovery replay on the key factors that cause the index to rise, and determine the cause chain of the index rise based on the changes in the index and the changes in the transmission path before and after the recovery. Input the cause chain of the index increase and the future port transportation network operation index into the large language model, and output the port transportation network operation index prediction statement.

2. The method for predicting the port transportation network operation index based on big data mining according to claim 1, characterized in that, The preprocessing of multi-source port operation data specifically includes: Collect multi-source port operation data, divide the multi-source port operation data into time segments according to the preset sampling time, delete duplicate and abnormal records, fill in missing fields, map data with spatial location to the corresponding port area nodes, perform standardization processing on data fields with different dimensions, and arrange the processed multi-source port operation data according to the port area node number and time slice order to obtain the historical state sequence of port area nodes.

3. The method for predicting the port transportation network operation index based on big data mining according to claim 1, characterized in that, The specific data for constructing the port transportation network structure based on the transportation relationships between key spatial units in the port area includes: Key spatial units in the port area are designated as port area nodes, and the comprehensive transportation impact relationship between port area nodes is determined based on vehicle traffic routes, ship entry and exit routes, cargo flow routes, operation connection routes, and risk transmission routes. Directed connection edges between port area nodes are constructed based on the transmission direction of the comprehensive transportation impact relationship, and the connection strength of the directed connection edges is determined based on the frequency of passage, circulation, number of operation connections and risk transmission between port area nodes. Port transportation network structure data is generated according to the port area node number, the direction of the directed connection edge, and the connection strength; The port transportation network structure data is used as the spatial diffusion input of the DCRNN model, and the historical state sequence of port nodes is used as the time series input of the DCRNN model.

4. The method for predicting the port transportation network operation index based on big data mining according to claim 3, characterized in that, The time memory update channel specifically includes: Read the node state vector of the same port area node in consecutive time slices from the historical state sequence of port area nodes. Use the current time slice node state vector, the previous time slice node state vector and the previous time slice node memory state as input to the time memory update channel. Perform differential processing on the current time slice node state vector and the previous time slice node state vector to obtain the node state change. Based on the changes in node state, determine whether the node state is continuously strengthening, continuously weakening, or remaining stable, and generate a state change identifier. Based on the state change identifier and the memory state of the previous time slice node, a memory retention coefficient and a memory update coefficient are generated. The part of the memory state of the previous time slice node that is consistent with the continuous trend is retained according to the memory retention coefficient. The part of the current time slice node state vector that corresponds to the continuous trend change is written according to the memory update coefficient, so as to obtain the memory state of the current time slice node. The memory state of the current time slice node obtained within a continuous time slice is output in chronological order as the continuous running trend of a single node.

5. The method for predicting the port transportation network operation index based on big data mining according to claim 4, characterized in that, The spatial diffusion receiving channel specifically includes: Read the incoming directed edges of the target port area nodes, obtain the single node continuous operation trend and connection strength of the associated nodes based on the incoming directed edges, multiply the single node continuous operation trend of the associated nodes by the corresponding connection strength and sum them to obtain the transmission influence of the associated nodes. The influence of associated nodes is the change in external state that is transmitted from the associated node to the target port area node through the incoming directed connection edge; The continuous operating trend of a single node in the target port area is taken as the node's own operating trend, and the influence of related nodes is taken as external transmission. The consistency of the trend direction between the node's own operating trend and external transmission is judged. When the node's own operating trend is consistent with the trend of external transmission, the external transmission is incorporated into the node's own operating trend to obtain the diffusion memory state of the target port area node. When the node's own operating trend is inconsistent with the external transmission trend, the external transmission write is blocked, and the node's own operating trend is used as the diffusion memory state of the target port area node.

6. The method for predicting the port transportation network operation index based on big data mining according to claim 5, characterized in that, The prediction of the port area node's operating status within multiple future time slices based on the improved DCRNN model specifically includes: The diffusion memory state of the target port area node is input into the decoding unit of the improved DCRNN model. The port area node operation state for the next time slice is generated according to the preset prediction step size. The port area node operation state for the next time slice is used as the subsequent prediction input for iterative decoding until the preset prediction time length is reached, so as to obtain the port area node operation state for multiple future time slices.

7. The method for predicting the port transportation network operation index based on big data mining according to claim 6, characterized in that, The specific implementation of future state replay according to the port area business transmission relationship includes: A digital twin port area scenario is constructed based on port area spatial data and port area business object data, and a mapping relationship is established between port area nodes and digital twin port area scenario objects. The operational status of port area nodes within multiple future time slices is loaded into the corresponding digital twin port area scene object according to the port area node number and the predicted time slice. The operational status of the digital twin port area scene object is updated frame by frame according to the order of the predicted time slices. The pressure level of each port node in each predicted time slot is determined based on the degree of deviation between the port node's operating status and the preset normal operating threshold. Based on the comparison of the pressure level changes of related port area nodes within adjacent forecast time slots according to the port area business transmission relationship, the pressure transmission relationship is recorded. Based on the pressure level, pressure transmission relationship, pressure duration, and number of affected port area nodes, the evolution result of port area operational pressure is formed.

8. The method for predicting the port transportation network operation index based on big data mining according to claim 7, characterized in that, The calculation of the future port transportation network operation index based on the evolution of port area operational pressure specifically includes: The future port transportation network operation index is calculated based on the port area operation pressure evolution results, including converting the pressure level of port area nodes in each forecast time segment into node pressure values; The conduction amplification factor is determined based on the continuous pressure transmission relationship, the duration correction factor is determined based on the pressure duration, and the range correction factor is determined based on the number of affected port area nodes. After summarizing the node pressure values ​​of each port area node, the values ​​are normalized by combining the transmission amplification coefficient, the continuous correction coefficient, and the range correction coefficient to obtain the future port transportation network operation index.

9. The method for predicting the port transportation network operation index based on big data mining according to claim 8, characterized in that, The chain of causes for the rise in the index, determined based on changes in the index before and after recovery and changes in the transmission path, specifically includes: The digital twin port area scene that has completed the replay of the future state is replicated, retaining the predicted time slice, port area node mapping relationship and port area business transmission relationship, and constructing a recovery replay scene corresponding to the original replay process; From the port area operation pressure evolution results, select port area nodes whose pressure level reaches the restricted or blocked as nodes to be restored, adjust the port area node operation status of the nodes to be restored in the recovery playback scenario to the normal operation threshold range, and keep the operation status of other port area nodes unchanged. Based on the adjusted recovery replay scenario, the future state replay is re-executed and the port transportation network operation index is recalculated. The changes in the port transportation network operation index and the changes in pressure transmission relationship before and after the recovery are compared. The node to be recovered where the index drops to a preset threshold and the pressure transmission relationship is interrupted is identified as the node causing the index increase. The index increase cause chain is generated according to the transmission sequence of the index increase cause node in the original playback process.

10. The method for predicting the port transportation network operation index based on big data mining according to claim 9, characterized in that, The output port transportation network operation index prediction statement specifically includes: The future port transportation network operation index, the corresponding prediction time slice, the chain of reasons for the index increase, the nodes of the reasons for the index increase, and the pressure transmission relationship are compiled into a structured prompt text; The structured prompt text is input into the large language model, which then generates a prediction statement for the port transportation network operation index according to a preset statement template. The port transportation network operation index prediction statement should include at least the prediction time, the future port transportation network operation index, the operation status level, the chain of reasons for the index increase, and the handling prompts.

Images