Intelligent scheduling method and system applied to marine transportation assembly box

Abstract

The invention provides an intelligent scheduling method and system applied to marine transportation assembled boxes, and the method comprises the steps: firstly obtaining a historical assembled box scheduling data set under a target marine transportation line, the historical assembled box scheduling data set comprises a plurality of scheduling sequences, and each sequence is composed of the initial position of an assembled box, a target port, cargo loading attributes, a historical path track and other characteristics; performing dynamic path feature extraction on the historical assembly box scheduling data set to obtain real-time environment association and path optimization features, and performing dynamic strategy matching on the real-time environment association and path optimization features based on a preset dynamic strategy matching network to generate a path labeling result; and finally, an optimization strategy set is generated according to a path labeling result and is used for training a dynamic strategy matching network to update parameters, so that intelligent and efficient scheduling of the marine transportation assembly box is realized.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an intelligent scheduling method and system for shipping LCL containers. Background Art

[0002] In the field of maritime logistics, the scheduling of LCL is an extremely complex and critical task, and its efficiency and accuracy directly affect the operational effectiveness of the entire maritime supply chain. Traditional LCL scheduling methods mainly rely on manual experience and fixed scheduling rules, which are incapable of coping with the growing demand for shipping, complex shipping routes and changing port conditions.

[0003] Specifically, existing scheduling methods often lack in-depth mining and utilization of historical scheduling data. Although basic information such as the starting location of the LCL, the target port, and the loaded cargo are recorded, most of this data is in a scattered and isolated state, and has failed to form an effective data set for further analysis. Due to the lack of extraction of path characteristics and environmental associations hidden in historical scheduling data, it is difficult to fully consider various dynamic factors in the actual transportation process, such as weather changes, port congestion, and the impact of cargo characteristics on the transportation path, when formulating scheduling strategies.

[0004] In addition, the traditional scheduling strategy matching method is relatively rigid, usually based on preset fixed rules, and cannot be flexibly adjusted according to the actual transportation environment and path conditions. This static strategy matching method often cannot respond effectively in time when faced with emergencies or complex and changing transportation needs, resulting in frequent problems such as low scheduling efficiency, untimely route adjustments, and unreasonable port resource allocation. At the same time, the existing scheduling methods lack an effective self-optimization mechanism. Once the scheduling strategy is formulated, it is often difficult to dynamically adjust and optimize it according to the actual implementation situation, and it is impossible to learn and improve from each scheduling practice, which limits the improvement of the scheduling level. Summary of the invention

[0005] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides an intelligent scheduling method for sea freight LCL, the method comprising: Acquire a historical LCL scheduling data set under a target shipping route, wherein the historical LCL scheduling data set includes a plurality of scheduling sequences, each scheduling sequence being composed of at least one LCL starting position feature, a target port feature, a loaded cargo attribute feature, and a historical path trajectory feature; Performing dynamic path feature extraction processing on the historical LCL scheduling data set to obtain real-time environment association features and path optimization features of each scheduling sequence; Based on a preset dynamic strategy matching network, dynamic strategy matching is performed on the real-time environment association feature and the path optimization feature to generate a path marking result of the scheduling sequence, wherein the path marking result is used to indicate the route adjustment direction and port resource allocation direction of the scheduling sequence; A set of optimized strategies for scheduling assembly containers is generated according to the path labeling results, and the dynamic strategy matching network is trained based on the set of optimized strategies for scheduling assembly containers to update network parameters.

[0006] In a possible implementation of the first aspect, the dynamic path feature extraction process is performed on the historical LCL scheduling data set to obtain the real-time environment association feature and path optimization feature of each scheduling sequence, including: Extracting a plurality of trajectory point data units corresponding to the historical path trajectory features from the scheduling sequence, each trajectory point data unit including a location coordinate, a timestamp and corresponding environmental monitoring data; Calling a pre-trained path feature encoder to perform spatiotemporal correlation encoding processing on the multiple trajectory point data units to generate a spatiotemporal fusion vector of the scheduling sequence, wherein the spatiotemporal fusion vector includes a path deviation degree and an environmental fluctuation correlation degree between adjacent trajectory points; Performing port resource matching processing on the starting position characteristics and the target port characteristics in the scheduling sequence to obtain port resource constraint characteristics, wherein the port resource constraint characteristics include at least one of the following: a berth idle period of the target port, an available number of loading and unloading equipment, and a cargo stacking density restriction; splicing the spatiotemporal fusion vector with the port resource constraint feature to generate the real-time environment association feature; The cargo loading attribute characteristics in the scheduling sequence are subjected to risk classification processing to obtain cargo transportation priority characteristics, and the cargo transportation priority characteristics are dynamically weighted fused with the spatiotemporal fusion vector to generate the path optimization characteristics.

[0007] In a possible implementation of the first aspect, calling a pre-trained path feature encoder to perform spatiotemporal correlation encoding processing on the multiple trajectory point data units to generate a spatiotemporal fusion vector of the scheduling sequence includes: For each trajectory point data unit, extract the wind speed, wave height and ocean current direction parameters in the environmental monitoring data to construct an environmental fluctuation matrix; Performing geographic grid encoding on the location coordinates to generate a standardized location encoding vector, and converting the timestamp into a periodic time encoding vector; Input the environmental fluctuation matrix, the standardized position encoding vector and the periodic time encoding vector into the path feature encoder for multi-head attention calculation to obtain a local environmental association vector for each trajectory point data unit; According to the time interval and distance interval between adjacent trajectory point data units, the trajectory point association weight is calculated, and the local environment association vector is subjected to sliding window aggregation based on the trajectory point association weight to generate the spatiotemporal fusion vector.

[0008] In a possible implementation of the first aspect, the performing risk classification processing on the attribute characteristics of the loaded cargo in the scheduling sequence to obtain the cargo transportation priority characteristics includes: Extracting cargo type, weight distribution and temperature control requirement parameters from the cargo attribute characteristics, and constructing a cargo attribute matrix; Calling a pre-trained risk assessment model to determine a cargo damage risk score, a transportation urgency score, and a priority coefficient according to the cargo attribute matrix; generating an initial priority feature according to a weighted sum of the cargo damage risk score and the transportation urgency score; The initial priority feature is nonlinearly transformed based on the priority coefficient to obtain the cargo transportation priority feature, wherein the nonlinear transformation includes mapping the initial priority feature to a preset priority interval through a Sigmoid function.

[0009] In a possible implementation of the first aspect, dynamically weighting and fusing the cargo transportation priority feature with the spatiotemporal fusion vector to generate the path optimization feature includes: Expanding the feature dimension of the cargo transportation priority feature to make it consistent with the dimension of the spatiotemporal fusion vector; Calculating the cosine similarity between the expanded cargo transportation priority feature and the spatiotemporal fusion vector to generate a dynamic weight coefficient; Performing a weighted summation of the cargo transportation priority feature and the spatiotemporal fusion vector according to the dynamic weight coefficient to generate a fusion intermediate feature; The fused intermediate features are subjected to dimensionality reduction processing to obtain the path optimization features, wherein the dimensionality reduction processing includes mapping the high-dimensional features to a low-dimensional space through a fully connected layer and retaining the feature components with the largest variance.

[0010] In a possible implementation of the first aspect, the performing dynamic strategy matching on the real-time environment association feature and the path optimization feature based on a preset dynamic strategy matching network to generate a path labeling result of the scheduling sequence includes: Inputting the real-time environment-related features into the environment perception branch of the dynamic strategy matching network to generate an environment constraint strategy vector; Inputting the path optimization feature into the path optimization branch of the dynamic strategy matching network to generate a path adjustment strategy vector; Performing cross-attention calculation on the environmental constraint strategy vector and the path adjustment strategy vector to generate a strategy interaction matrix; Filtering key strategy channels according to the value of each element in the strategy interaction matrix, and generating a strategy fusion vector through channel weighted pooling; The strategy fusion vector is input into a labeling classifier, and the path labeling result is output, wherein the labeling classifier includes a plurality of fully connected layers and a Softmax layer, which is used to map continuous features to discrete labeling categories.

[0011] In a possible implementation of the first aspect, screening key strategy channels according to the value of each element in the strategy interaction matrix, and generating a strategy fusion vector by channel weighted pooling, includes: Performing global average pooling on the strategy interaction matrix along the channel dimension to generate a channel importance score; According to a preset importance threshold, a target channel whose channel importance score is higher than the importance threshold is screened out, and a strategy interaction submatrix corresponding to the target channel is extracted; Performing a maximum pooling operation on the strategy interaction submatrix to obtain channel salient features; Multiplying the channel salient features by the strategy interaction submatrix element by element to generate a weighted strategy interaction feature; The weighted strategy interaction features are flattened to obtain the strategy fusion vector.

[0012] In a possible implementation of the first aspect, generating a set of LCL scheduling optimization strategies according to the path marking result includes: Analyzing the route adjustment direction in the path marking result to generate at least one candidate route adjustment plan, each candidate route adjustment plan including a route deviation angle, an adjustment distance, and an estimated time loss; Parsing the port resource allocation direction in the path marking result, generating port resource allocation constraint conditions, wherein the constraint conditions include berth occupancy time limit, the number of loading and unloading equipment allocated, and cargo storage area identification; Conduct feasibility verification on the candidate route adjustment plans and select effective adjustment plans that meet the port resource allocation constraints; A comprehensive optimization score is calculated according to the estimated time loss and path deviation in the effective adjustment scheme, and the set of optimization strategies for scheduling the assembled container is generated by sorting the comprehensive optimization scores.

[0013] In a possible implementation of the first aspect, the feasibility verification of the candidate route adjustment schemes to select effective adjustment schemes that meet the port resource allocation constraint conditions includes: Extracting the berth idle time period in the port resource allocation constraint condition and matching it with the estimated arrival time in the candidate route adjustment scheme, and eliminating the adjustment scheme whose arrival time exceeds the idle time period; Extract the assigned quantity of the loading and unloading equipment, calculate the quantity of equipment required for the candidate route adjustment plan, and eliminate the adjustment plan where the equipment demand exceeds the available quantity; The cargo storage area identifier is extracted, and it is verified whether the cargo type corresponding to the candidate route adjustment plan is allowed to enter the target storage area, and the adjustment plan that violates the storage rules is eliminated, thereby generating a valid adjustment plan that meets the port resource allocation constraint conditions.

[0014] In a possible implementation of the first aspect, the training the dynamic strategy matching network based on the assembly box scheduling optimization strategy set to update network parameters includes: Extracting strategy execution result data from the LCL scheduling optimization strategy set, wherein the strategy execution result data includes time loss, resource utilization rate and cargo damage rate after actual route adjustment; Constructing a strategy effect evaluation function, generating a time error loss according to the difference between the time loss after the actual route adjustment and the estimated time loss, generating a resource loss according to the deviation between the resource utilization rate and a preset threshold, and generating a risk loss according to the cargo damage rate; The time error loss, resource loss and risk loss are weighted and summed to obtain the total training loss; A gradient descent algorithm is used to update the parameters of the dynamic strategy matching network based on the total training loss until the total training loss converges to a stable interval.

[0015] On the other hand, an embodiment of the present invention further provides a shipping Internet of Things management system, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0016] Based on the above aspects, the embodiments of the present invention realize the intelligent and efficient scheduling of LCL shipping, and significantly improve the accuracy and adaptability of scheduling strategies. Specifically, by obtaining the historical LCL scheduling data set under the target shipping route and performing dynamic path feature extraction and processing, it is possible to deeply explore the real-time environment association features and path optimization features of each scheduling sequence, and dynamically match the real-time environment association features with the path optimization features based on the preset dynamic strategy matching network, and generate a path marking result with clear route adjustment direction and port resource allocation direction, so that the scheduling strategy can be flexibly adjusted according to the actual environment and path conditions, greatly improving the flexibility and response speed of scheduling. Furthermore, a set of LCL scheduling optimization strategies is generated based on the path marking results, and a dynamic strategy matching network is trained based on the LCL scheduling optimization strategy set to update network parameters, which can not only continuously optimize the scheduling strategy and improve the scheduling efficiency, but also adapt to the ever-changing shipping environment and market demand. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 The present invention is a flowchart of an intelligent scheduling method for sea freight LCL provided in an embodiment of the present invention.

[0018] Figure 2 Schematic diagram of exemplary hardware and software components of a shipping IoT management system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0019] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 The figure is a flow chart of an intelligent scheduling method for LCL shipped by sea provided by an embodiment of the present invention. The intelligent scheduling method for LCL shipped by sea is introduced in detail below.

[0020] Step S110: Acquire a historical LCL scheduling data set under the target shipping route, wherein the historical LCL scheduling data set includes a plurality of scheduling sequences, each scheduling sequence being composed of at least one LCL starting position feature, target port feature, loaded cargo attribute feature, and historical path trajectory feature.

[0021] In order to realize the intelligent scheduling of LCL, it is necessary to first obtain the historical LCL scheduling data set under the target shipping line. For example, the historical LCL scheduling data can be collected with the help of the shipping company's operation database, port management system and related logistics information platform, which covers the LCL transportation situation on the shipping line in different time periods.

[0022] In detail, assuming that the target shipping route is a busy route connecting multiple large ports, the collected historical LCL scheduling data set contains 1,000 scheduling sequences, each of which represents a LCL transportation task and contains multiple key features.

[0023] Among them, in terms of the starting position characteristics, taking one of the scheduling sequences as an example, the starting port of the LCL is located at 123 degrees east longitude and 35 degrees north latitude. The port is a comprehensive port with a variety of loading and unloading facilities and service functions. The surrounding transportation network is well-developed, which is convenient for the distribution and transshipment of goods. The types of facilities at the starting port include container terminals, bulk cargo terminals, storage areas, etc. Different types of facilities have an important impact on the loading and unloading efficiency and transportation arrangements of the LCL. For example, the container terminal is equipped with advanced cranes and automation equipment, which can quickly complete the loading and unloading of LCL; while the bulk cargo terminal is more suitable for handling some bulk cargo loading and unloading.

[0024] In terms of the characteristics of the target port, the target port for the LCL is located at 130 degrees east longitude and 38 degrees north latitude. It is an important hub port with a large number of berths, advanced loading and unloading equipment and a complete logistics service system. The type of the target port is a dedicated container port with a maximum throughput of 5 million standard containers per year. The busyness and operating conditions of the port will affect the arrival time, loading and unloading arrangements and subsequent transportation connections of the LCL. For example, during busy periods at the port, the LCL may need to queue up for berths, resulting in delayed arrival time; when the port is operating efficiently, the loading and unloading operations can be completed quickly, shortening the transportation cycle.

[0025] The attribute characteristics of loaded goods include information about many aspects of the goods. For example, the types of goods loaded in the LCL are mainly electronic products, textiles and food. Electronic products are of high value and easy to damage, and have high requirements for the stability and safety of the transportation environment; textiles are relatively resistant to transportation, but attention should be paid to moisture and fire prevention; food goods have strict preservation and hygiene requirements, and some foods also need to be transported under specific temperature and humidity conditions. The weight distribution of the goods is 30 tons of electronic products, 40 tons of textiles, and 30 tons of food, with a total weight of 100 tons. Different weight distributions of goods will affect the center of gravity balance and transportation stability of the LCL, and the stacking position of the goods needs to be reasonably arranged during loading, unloading and transportation. In addition, food goods have strict requirements for temperature control and need to be transported in an environment of 2-8 degrees Celsius. This requires the LCL to be equipped with corresponding temperature control equipment and monitor temperature changes in real time during transportation.

[0026] The historical path trajectory feature records the information of each track point that the LCL passes through during transportation. For example, through the satellite positioning system and sensor equipment, the location coordinates, timestamp and corresponding environmental monitoring data of each track point can be obtained. Assume that the LCL passes through 20 track points during transportation, the location coordinates of track point 1 are 124 degrees east longitude and 36 degrees north latitude, and the timestamp is 10:00 on January 1, 2024. The environmental monitoring data at this time shows that the wind speed is 5 meters per second, the wave height is 1 meter, and the ocean current direction is due east. As the transportation progresses, the environmental conditions of each track point will change, and these changes will affect the transportation path and speed of the LCL. For example, when encountering strong winds or high waves, the ship may need to adjust the route to ensure navigation safety; while the downstream ocean current can increase the ship's sailing speed and shorten the transportation time.

[0027] Step S120: performing dynamic path feature extraction processing on the historical LCL scheduling data set to obtain real-time environment association features and path optimization features of each scheduling sequence.

[0028] In this embodiment, after obtaining the historical LCL scheduling data set, it is necessary to perform dynamic path feature extraction processing on it to mine the real-time environment association features and path optimization features of each scheduling sequence. The specific processing process is as follows: Step S121: extracting a plurality of trajectory point data units corresponding to the historical path trajectory features from the scheduling sequence, each trajectory point data unit including a location coordinate, a timestamp and corresponding environmental monitoring data.

[0029] For example, taking the previously mentioned dispatch sequence containing 20 trajectory points as an example, its historical path trajectory characteristics are analyzed in detail. Through data parsing and sorting, the relevant information of each trajectory point is extracted to form a trajectory point data unit.

[0030] For example, for track point 1, its location coordinates are 124 degrees east longitude and 36 degrees north latitude. This coordinate information accurately determines the location of the track point in geographic space. The timestamp is 10:00 on January 1, 2024, recording the specific time when the LCL arrived at the track point. Environmental monitoring data shows that the wind speed is 5 meters per second, the wave height is 1 meter, and the ocean current direction is due east. These environmental data reflect the marine environmental conditions at the time and have a certain impact on the transportation of LCL.

[0031] The coordinates of track point 2 are 125 degrees east longitude and 36.5 degrees north latitude, and the timestamp is 12:00 on January 1, 2024. At this time, the wind speed becomes 6 meters per second, the wave height is 1.2 meters, and the ocean current direction is still due east. As time goes by and the location changes, the environmental conditions are also constantly changing. By extracting and analyzing the data unit of each track point, we can understand the environmental changes that the LCL experiences during transportation.

[0032] Step S122: calling a pre-trained path feature encoder to perform spatiotemporal correlation encoding processing on the multiple trajectory point data units to generate a spatiotemporal fusion vector of the scheduling sequence, wherein the spatiotemporal fusion vector includes the path deviation degree and environmental fluctuation correlation degree between adjacent trajectory points.

[0033] In order to generate the spatiotemporal fusion vector of the scheduling sequence, it is necessary to call the pre-trained path feature encoder to perform spatiotemporal correlation encoding on multiple trajectory point data units. The pre-trained path feature encoder is trained based on a large amount of historical trajectory data and can effectively capture the spatiotemporal correlation information between trajectory points. The specific steps are as follows: Step S1221: For each trajectory point data unit, extract the wind speed, wave height and ocean current direction parameters in the environmental monitoring data to construct an environmental fluctuation matrix.

[0034] Still taking trajectory point 1 as an example, the wind speed in its environmental monitoring data is 5 meters per second, the wave height is 1 meter, and the ocean current direction is due east. Therefore, these parameters can be sorted and combined to construct an environmental fluctuation matrix. Assuming that the wind speed, wave height and ocean current direction are respectively used as the three dimensions of the matrix, the environmental fluctuation matrix of trajectory point 1 can be expressed as [5, 1, due east]. The due east direction here can be further quantified. For example, if the due east direction is defined as an angle of 0 degrees, then the environmental fluctuation matrix can be more accurately expressed as [5, 1, 0].

[0035] For trajectory point 2, the wind speed is 6 meters per second, the wave height is 1.2 meters, and the ocean current direction is still due east. The corresponding environmental fluctuation matrix is ​​[6, 1.2, 0]. By processing the environmental monitoring data of each trajectory point data unit, the corresponding environmental fluctuation matrix is ​​constructed. These environmental fluctuation matrices can intuitively reflect the environmental fluctuation situation of each trajectory point.

[0036] Step S1222: Perform geographic grid encoding on the location coordinates to generate a standardized location encoding vector, and convert the timestamp into a periodic time encoding vector.

[0037] In this embodiment, the geographic grid coding is to map the geographic location coordinates into a regular grid system to facilitate unified processing and analysis of location information. The geographic grid coding system can divide the earth's surface into grid cells of equal size, and each grid cell has a unique code. For the location coordinates of trajectory point 1, 124 degrees east longitude and 36 degrees north latitude, by querying the geographic grid coding table, it is mapped to the corresponding grid cell, and the geographic grid code of the trajectory point is G123. The geographic grid code is converted into a standardized position coding vector. For example, the geographic grid code can be represented as a binary vector of length 10, and the binary vector corresponding to G123 is [0, 0, 1, 0, 0, 1, 1, 0, 0, 1].

[0038] The purpose of converting a timestamp into a periodic time coding vector is to capture the periodic characteristics of time. In detail, the timestamp can be divided into a certain period, for example, one day is a period, and 10:00 on January 1, 2024 is converted into a periodic time coding vector. First, the relative position of the timestamp in a day is calculated, and the proportion of 10:00 in 24 hours a day is 10 / 24≈0.42. This ratio is converted into a vector of length 10, and through the common predefined mapping rules in the prior art, the periodic time coding vector is obtained as [0, 0, 0, 0, 0, 1, 0, 0, 0, 0].

[0039] For trajectory point 2, the same geographic grid encoding and timestamp conversion are performed to obtain its standardized position encoding vector and periodic time encoding vector. In this way, the position coordinates and timestamps are converted into a unified encoding vector, which is convenient for subsequent multi-head attention calculation.

[0040] Step S1223: Input the environmental fluctuation matrix, the standardized position encoding vector and the periodic time encoding vector into the path feature encoder for multi-head attention calculation to obtain the local environmental association vector of each trajectory point data unit.

[0041] In this embodiment, multi-head attention calculation is the core operation in the path feature encoder, which can capture the correlation information between different features and form a comprehensive input vector by combining the environmental fluctuation matrix, standardized position encoding vector and periodic time encoding vector of each trajectory point.

[0042] Taking trajectory point 1 as an example, concatenate its environmental fluctuation matrix [5, 1, 0], standardized position encoding vector [0, 0, 1, 0, 0, 1, 1, 0, 0, 1] and periodic time encoding vector [0, 0, 0, 0, 0, 1, 0, 0, 0, 0] together to obtain the input vector [5, 1, 0, 0, 0, 1, 0, 0, 1, 1, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0].

[0043] Then, the input vector is input into the multi-head attention mechanism of the path feature encoder. The multi-head attention mechanism calculates the degree of association between different parts of the input vector in parallel through multiple attention heads. Each attention head generates a query vector, a key vector, and a value vector based on the input vector, and then obtains the attention weight by calculating the similarity between the query vector and the key vector. The attention weight is applied to the value vector to obtain the output result of each attention head. Finally, the output results of multiple attention heads are concatenated and linearly transformed to obtain the local environment association vector of trajectory point 1. Assume that after multi-head attention calculation, the local environment association vector of trajectory point 1 is [0.2, 0.3, 0.1, 0.4, 0.5, 0.2, 0.3, 0.1, 0.4, 0.5].

[0044] The above operation is also performed for trajectory point 2 and other trajectory points to obtain their respective local environment association vectors, which reflect the degree of association between each trajectory point and the surrounding environment.

[0045] Step S1224: Calculate the trajectory point association weight according to the time interval and distance interval between adjacent trajectory point data units, and perform sliding window aggregation on the local environment association vector based on the trajectory point association weight to generate the spatiotemporal fusion vector.

[0046] In this embodiment, the time interval and distance interval between adjacent track points are important indicators to measure the degree of association between them. Still taking track point 1 and track point 2 as an example, the timestamp of track point 1 is 10:00 on January 1, 2024, and the timestamp of track point 2 is 12:00 on January 1, 2024, and the time interval is 2 hours. The distance interval between track point 1 and track point 2 is calculated by geographic coordinates, assuming that the calculated distance is 100 nautical miles.

[0047] Then, the track point association weight is calculated based on the time interval and the distance interval. For example, a simple weighted calculation method can be used, such as normalizing the time interval and the distance interval, assigning the set weights to each, and then adding them to obtain the track point association weight. Assuming that the weight of the time interval is 0.6 and the weight of the distance interval is 0.4, the time interval of 2 hours is normalized to the interval [0, 1] as 0.2, and the distance interval of 100 nautical miles is normalized to the interval [0, 1] as 0.3, then the track point association weight between track point 1 and track point 2 is 0.6×0.2+0.4×0.3=0.24.

[0048] Then, the sliding window aggregation method is used to process the local environment association vector. Assume that the size of the sliding window is 3, that is, 3 adjacent trajectory points are considered each time. For trajectory point 2, its sliding window contains trajectory point 1, trajectory point 2 and trajectory point 3. According to the calculated trajectory point association weights, the local environment association vectors of these three trajectory points are weighted summed. Assume that the local environment association vectors of trajectory point 1, trajectory point 2 and trajectory point 3 are [0.2, 0.3, 0.1, 0.4, 0.5, 0.2, 0.3, 0.1, 0.4, 0.5], [0.3, 0.4, 0.2, 0.5, 0.6, 0.3, 0.4, 0.2, 0.5, 0.6] and [0.4, 0.5, 0.3, 0.6, 0.7, 0.4, 0.5, 0.3, 0.6, 0.7] respectively, and the association weight of trajectory point 1 and trajectory point 2 is 0.24, The association weight of trajectory point 2 and trajectory point 3 is 0.26, so the intermediate result after weighted summation is 0.24×[0.2, 0.3, 0.1, 0.4, 0.5, 0.2, 0.3, 0.1, 0.4, 0.5]+1×[0.3, 0.4, 0.2, 0.5, 0.6, 0.3, 0.4, 0.2, 0.5, 0.6]+0.26×[0.4, 0.5, 0.3, 0.6, 0.7, 0.4, 0.5, 0.3, 0.6, 0.7].

[0049] Next, the intermediate results are further processed, such as normalization, to obtain the final spatiotemporal fusion vector. Assume that after normalization, the spatiotemporal fusion vector corresponding to trajectory point 2 is [0.25, 0.35, 0.15, 0.45, 0.55, 0.25, 0.35, 0.15, 0.45, 0.55]. In this way, each trajectory point is processed to obtain the spatiotemporal fusion vector of the entire scheduling sequence, which contains important information such as the path deviation between adjacent trajectory points and the correlation of environmental fluctuations.

[0050] Step S123: performing port resource matching processing on the starting position characteristics and the target port characteristics in the scheduling sequence to obtain port resource constraint characteristics, wherein the port resource constraint characteristics include at least one of the following: a berth idle period of the target port, the available number of loading and unloading equipment, and a cargo stacking density limit.

[0051] In this embodiment, the starting position characteristics and the target port characteristics contain important information related to the port resources. The port resource constraint characteristics are determined by performing port resource matching processing on these information.

[0052] For the target port, data is exchanged with the port management system to obtain information such as berth idle time periods, available number of loading and unloading equipment, and cargo stacking density restrictions. Assume that the target port has a total of 20 berths, and 5 berths are idle from January 5 to January 7, 2024. This is the berth idle time period of the target port during this time period.

[0053] In terms of the available number of loading and unloading equipment, the target port is equipped with 10 cranes and 20 forklifts. Currently, 3 cranes and 5 forklifts are in use. Therefore, the current number of available cranes is 7 and the number of available forklifts is 15.

[0054] Cargo stacking density restrictions refer to certain requirements for the stacking density of different types of cargo at the port. For example, for container cargo, the port stipulates that the stacking density per square meter cannot exceed 5 tons; for bulk cargo, the stacking density per cubic meter cannot exceed 2 tons.

[0055] This information is collated and integrated to obtain the port resource constraint characteristics. These characteristics will have an important impact on the scheduling and transportation of LCLs. For example, when arranging the arrival time of LCLs, it is necessary to consider the idle time of berths; when carrying out loading and unloading operations, it is necessary to make reasonable arrangements based on the available number of loading and unloading equipment; when stacking goods, it is necessary to comply with the cargo stacking density restrictions.

[0056] Step S124: splicing the spatiotemporal fusion vector with the port resource constraint feature to generate the real-time environment association feature.

[0057] In this embodiment, the previously obtained spatiotemporal fusion vector and the port resource constraint feature can be spliced ​​to generate a real-time environment association feature. Assume that the spatiotemporal fusion vector is [0.25, 0.35, 0.15, 0.45, 0.55, 0.25, 0.35, 0.15, 0.45, 0.55], and the port resource constraint feature is expressed as [5 berth idle periods (January 5-January 7, 2024), 7 available cranes, 15 available forklifts, and a stacking density limit of no more than 5 tons per square meter for containers and a stacking density limit of no more than 2 tons per cubic meter for bulk cargo].

[0058] In order to perform splicing, the port resource constraint characteristics need to be quantified. For example, the berth idle period is converted into a time period coding vector, the number of available cranes and forklifts is converted into a numerical vector, and the cargo stacking density limit is converted into a constraint vector. Assume that after quantification, the vector corresponding to the port resource constraint characteristics is [0.5, 0.7, 0.15, 0.5, 0.2]. The spatiotemporal fusion vector and the quantified port resource constraint feature vector are spliced ​​to obtain the real-time environment association feature vector [0.25, 0.35, 0.15, 0.45, 0.55, 0.25, 0.35, 0.15, 0.45, 0.55, 0.5, 0.7, 0.15, 0.5, 0.2]. The real-time environment association feature vector integrates the spatiotemporal information and port resource constraint information in the LCL transportation process, and can more comprehensively reflect the current transportation environment.

[0059] Step S125: Perform risk classification processing on the attribute characteristics of the loaded cargo in the scheduling sequence to obtain cargo transportation priority characteristics, and dynamically weighted fuse the cargo transportation priority characteristics with the spatiotemporal fusion vector to generate the path optimization characteristics.

[0060] In order to achieve more reasonable path planning, it is necessary to perform risk classification on the attribute characteristics of the loaded goods in the scheduling sequence to determine the priority characteristics of cargo transportation, and then dynamically weighted fuse them with the spatiotemporal fusion vector to generate path optimization characteristics. The specific steps are as follows: Step S1251: extract the cargo type, weight distribution and temperature control requirement parameters from the cargo attribute characteristics, and construct a cargo attribute matrix.

[0061] Taking the LCL mentioned above as an example, the cargo attributes of the LCL include cargo type (electronic products, textiles, food), weight distribution (30 tons of electronic products, 40 tons of textiles, 30 tons of food), and temperature control requirements (food needs to be kept at 2-8 degrees Celsius). These parameters are sorted and combined to construct a cargo attribute matrix. The cargo type, weight distribution, and temperature control requirements can be used as different dimensions of the matrix.

[0062] Assume that electronic products, textiles, and food are coded as 1, 2, and 3 respectively, the weight distribution of goods is represented by the actual weight value, and the temperature control requirements are coded by whether temperature control is required and the temperature control range. Then the cargo attribute matrix can be expressed as: |Cargo type code|Weight (tons)|Temperature control requirement code| |----|----|----| |1|30|0 (electronic products do not require special temperature control)| |2|40|0 (textiles do not require special temperature control)| |3|30|1 (food needs 2-8 degrees Celsius temperature control)| Therefore, the above cargo attribute matrix comprehensively reflects the basic attribute information of the cargo in the assembled container.

[0063] Step S1252: calling a pre-trained risk assessment model to determine a cargo damage risk score, a transportation urgency score, and a priority coefficient according to the cargo attribute matrix.

[0064] In this embodiment, the pre-trained risk assessment model is trained based on a large amount of historical cargo transportation data, and can accurately assess the damage risk, transportation urgency, and priority coefficient of the cargo according to the cargo attribute matrix.

[0065] Therefore, the above cargo attribute matrix can be input into the pre-trained risk assessment model. For electronic products, due to their high value and fragility, the model gives a cargo damage risk score of 8 points (out of 10 points) based on their information in the cargo attribute matrix and comprehensive consideration of factors such as their transportation environment requirements. Because the delivery time of electronic products may have a significant impact on subsequent production or sales, the transportation urgency score is 7 points.

[0066] Textiles are relatively durable in transportation and have a low risk of damage. The model gives a cargo damage risk score of 3 points. Its transportation urgency is assessed as 4 points based on factors such as market demand and delivery time.

[0067] Since food has strict preservation requirements, the cargo damage risk score is 7 points. If this batch of food is supplied to some markets with extremely high requirements for freshness, the transportation urgency score is 8 points.

[0068] Based on the above scores, the risk assessment model further calculates the priority coefficient. The priority coefficient can be calculated by a weighted comprehensive method, for example, considering that the cargo damage risk score accounts for 0.6 and the transportation urgency score accounts for 0.4. For electronic products, the priority coefficient = 0.6 × 8 + 0.4 × 7 = 7.6; for textiles, the priority coefficient = 0.6 × 3 + 0.4 × 4 = 3.4; for food, the priority coefficient = 0.6 × 7 + 0.4 × 8 = 7.4.

[0069] The training of the risk assessment model can refer to the common training process in the conventional prior art.

[0070] Step S1253: Generate an initial priority feature according to the weighted sum of the cargo damage risk score and the transportation urgency score.

[0071] In this embodiment, the weighted sum of the cargo damage risk score and the transportation urgency score is calculated to generate the initial priority feature. Assume that the weight of the cargo damage risk score is 0.7 and the weight of the transportation urgency score is 0.3.

[0072] For electronic products, the initial priority feature = 0.7×8+0.3×7=7.7; for textiles, the initial priority feature = 0.7×3+0.3×4=3.3; for food, the initial priority feature = 0.7×7+0.3×8=7.3.

[0073] Next, these initial priority features are combined into a vector. Assuming that they are in the order of electronic products, textiles, and food, the initial priority feature vector is [7.7, 3.3, 7.3].

[0074] Step S1254: Perform a nonlinear transformation on the initial priority feature based on the priority coefficient to obtain the cargo transportation priority feature, wherein the nonlinear transformation includes mapping the initial priority feature to a preset priority interval through a Sigmoid function.

[0075] In order to map the initial priority feature to a suitable priority interval, the Sigmoid function is used for nonlinear transformation. The formula of the Sigmoid function is S(x)=1 / (1+e^(-x)), which can map the input value to the (0, 1) interval.

[0076] For the initial priority feature vector [7.7, 3.3, 7.3], substitute each value into the Sigmoid function for calculation.

[0077] For 7.7, S(7.7)=1 / (1+e^(-7.7)), which is calculated to be about 0.999. For 3.3, S(3.3)=1 / (1+e^(-3.3)), which is about 0.964. For 7.3, S(7.3)=1 / (1+e^(-7.3)), which is about 0.998.

[0078] Finally, these calculation results are combined into a cargo transportation priority feature vector [0.999, 0.964, 0.998], which represents the transportation priority of different types of cargo. The closer the value is to 1, the higher the priority.

[0079] Step S1255: Dynamically weight the cargo transportation priority feature and the spatiotemporal fusion vector to generate the path optimization feature.

[0080] In this embodiment, in order to generate the route optimization feature, it is necessary to dynamically weight the cargo transportation priority feature and the spatiotemporal fusion vector. The specific steps are as follows: Step S1255-1: Expand the feature dimension of the cargo transportation priority feature to make it consistent with the dimension of the space-time fusion vector.

[0081] For example, the cargo transportation priority feature vector is [0.999, 0.964, 0.998], and the spatiotemporal fusion vector is [0.25, 0.35, 0.15, 0.45, 0.55, 0.25, 0.35, 0.15, 0.45, 0.55]. In order to make the dimensions of the two consistent, the cargo transportation priority feature needs to be expanded.

[0082] Exemplarily, the dimension expansion can be performed by repeated padding. Assume that the dimension of the spatiotemporal fusion vector is 10 and the dimension of the cargo transportation priority feature vector is 3. Repeat the padding of the cargo transportation priority feature vector according to certain rules to make its dimension reach 10. For example, [0.999, 0.964, 0.998] is expanded to [0.999, 0.964, 0.998, 0.999, 0.964, 0.998, 0.999, 0.964, 0.998, 0.999].

[0083] Step S1255-2: Calculate the cosine similarity between the expanded cargo transportation priority feature and the spatiotemporal fusion vector to generate a dynamic weight coefficient.

[0084] In this embodiment, cosine similarity is an indicator for measuring the cosine value of the angle between two vectors, which can reflect the similarity between the two vectors. The calculation formula is: cosθ=(A·B) / (||A||||B||), where A and B are two vectors, A·B represents the dot product of the vectors, and ||A|| and ||B|| represent the modulus of the vectors.

[0085] The expanded cargo transportation priority feature vector A=[0.999, 0.964, 0.998, 0.999, 0.964, 0.998, 0.999, 0.964, 0.998, 0.999], and the spatiotemporal fusion vector B=[0.25, 0.35, 0.15, 0.45, 0.55, 0.25, 0.35, 0.15, 0.45, 0.55].

[0086] First calculate the dot product of the vectors A·B, that is, multiply the corresponding elements and then sum them: A·B=0.999×0.25+0.964×0.35+0.998×0.15+0.999×0.45+0.964×0.55+0.998×0.25+0.999×0.35+0.964×0.15+0.998×0.45+0.999×0.55 The value of A·B is obtained by calculation.

[0087] Then calculate the vector's magnitude ||A|| and ||B||.

[0088] ||A||=√(0.999²+0.964²+0.998²+0.999²+0.964²+0.998²+0.999²+0.964²+0.998²+0.999²) ||B||=√(0.25²+0.35²+0.15²+0.45²+0.55²+0.25²+0.35²+0.15²+0.45²+0.55²) The values ​​of ||A|| and ||B|| are obtained by calculation.

[0089] Finally, the cosine similarity cosθ=(A·B) / (||A||||B||) is calculated. Assuming that the calculated cosine similarity is 0.6, it is used as the dynamic weight coefficient.

[0090] Step S1255-3: Perform weighted summation on the cargo transportation priority feature and the spatiotemporal fusion vector according to the dynamic weight coefficient to generate a fusion intermediate feature.

[0091] For example, if the dynamic weight coefficient is 0.6, then the weight of the cargo transportation priority feature is 0.6, and the weight of the spatiotemporal fusion vector is 1-0.6=0.4.

[0092] The calculation method of each element of the fusion intermediate feature is: fusion intermediate feature element = 0.6 × cargo transportation priority feature element + 0.4 × space-time fusion vector element.

[0093] For example, the first element of the fused intermediate feature = 0.6×0.999+0.4×0.25, and each element is calculated in turn to obtain the fused intermediate feature vector.

[0094] Step S1255-4: Perform dimensionality reduction processing on the fused intermediate features to obtain the path optimization features, wherein the dimensionality reduction processing includes mapping the high-dimensional features to a low-dimensional space through a fully connected layer and retaining the feature components with the largest variance.

[0095] In this embodiment, the fully connected layer is a commonly used neural network layer that can map the input high-dimensional features to a low-dimensional space. Assume that the dimension of the fused intermediate feature vector is 10, and it is desired to reduce the dimension to 5 dimensions.

[0096] The fully connected layer contains multiple neurons, each of which is connected to each element of the input vector. Through a series of weight parameters and bias parameters, the input vector is linearly transformed and nonlinearly activated to obtain the output vector.

[0097] In the process of dimensionality reduction, the feature component with the largest variance is retained. Variance reflects the degree of discreteness of the data, and feature components with large variance contain more information. By calculating the variance of the output of the fully connected layer, the five feature components with the largest variance are selected as the path optimization feature vector. Assume that after dimensionality reduction, the path optimization feature vector is [0.3, 0.5, 0.2, 0.6, 0.4], which integrates the cargo transportation priority and spatiotemporal information.

[0098] Step S130: Based on a preset dynamic strategy matching network, dynamic strategy matching is performed on the real-time environment association characteristics and the path optimization characteristics to generate a path marking result of the scheduling sequence, and the path marking result is used to indicate the route adjustment direction and port resource allocation direction of the scheduling sequence.

[0099] In this embodiment, the preset dynamic strategy matching network is a trained neural network model that can perform dynamic strategy matching based on real-time environment association features and path optimization features to generate path annotation results. The specific steps are as follows: Step S131: inputting the real-time environment-related features into the environment perception branch of the dynamic strategy matching network to generate an environment constraint strategy vector.

[0100] The real-time environment association feature vector is [0.25, 0.35, 0.15, 0.45, 0.55, 0.25, 0.35, 0.15, 0.45, 0.55, 0.5, 0.7, 0.15, 0.5, 0.2]. It is input into the environment perception branch of the dynamic policy matching network.

[0101] The environmental perception branch is composed of multiple neural network layers, including convolutional layers, pooling layers, and fully connected layers. The convolutional layer is used to extract local features from real-time environmental correlation features, the pooling layer is used to reduce the dimension of features and aggregate information, and the fully connected layer is used to map high-dimensional features to a vector of fixed dimension.

[0102] Assume that after a series of calculations and transformations, the environmental perception branch maps the real-time environmental correlation features to a vector of dimension 8, and obtains the environmental constraint strategy vector [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8]. The environmental constraint strategy vector represents the constraints of the current transportation environment on the allocation of routes and port resources.

[0103] Step S132: inputting the path optimization feature into the path optimization branch of the dynamic strategy matching network to generate a path adjustment strategy vector.

[0104] In this embodiment, the path optimization feature vector is [0.3, 0.5, 0.2, 0.6, 0.4], which is input into the path optimization branch of the dynamic strategy matching network. The path optimization branch is also composed of multiple neural network layers, and its structure and function are similar to those of the environmental perception branch, but it focuses more on the processing and analysis of path optimization features. After calculation and transformation by the path optimization branch, the path optimization feature is mapped to a vector with a dimension of 8, and the path adjustment strategy vector is obtained as [0.2, 0.4, 0.1, 0.5, 0.3, 0.6, 0.2, 0.7]. The path adjustment strategy vector represents the adjustment strategy required to optimize the path.

[0105] Step S133: Perform cross-attention calculation on the environmental constraint strategy vector and the path adjustment strategy vector to generate a strategy interaction matrix.

[0106] In this embodiment, the cross attention calculation is used to capture the correlation information between the environmental constraint strategy vector and the path adjustment strategy vector. The specific calculation process is as follows: First, the environment constraint strategy vector and the path adjustment strategy vector are linearly transformed to obtain the query vector Q, key vector K and value vector V. Assume that the environment constraint strategy vector A = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8] and the path adjustment strategy vector B = [0.2, 0.4, 0.1, 0.5, 0.3, 0.6, 0.2, 0.7].

[0107] The query vector Q, key vector K and value vector V are obtained through linear transformation, and then the similarity between the query vector Q and the key vector K is calculated, usually using the dot product method. After obtaining the similarity matrix, softmax normalization is performed to obtain the attention weight matrix.

[0108] Multiply the attention weight matrix by the value vector V to get the policy interaction matrix. Assume that the calculated policy interaction matrix is ​​an 8×8 matrix: |0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8| |0.2 0.4 0.1 0.5 0.3 0.6 0.2 0.7| |...| |...| |...| |...| |...| |...| Therefore, the strategy interaction matrix reflects the interaction relationship between the environmental constraint strategy and the path adjustment strategy.

[0109] Step S134: Filter key strategy channels according to the value of each element in the strategy interaction matrix, and generate a strategy fusion vector through channel weighted pooling.

[0110] In order to extract key information from the strategy interaction matrix, it is necessary to screen key strategy channels and perform channel weighted pooling. The specific steps are as follows: Step S1341: Perform global average pooling on the strategy interaction matrix along the channel dimension to generate a channel importance score.

[0111] In this embodiment, global average pooling is to average the strategy interaction matrix in the channel dimension. For each column of the strategy interaction matrix, all elements of the column are added up and then divided by the number of elements to obtain the average value of the column, that is, the channel importance score.

[0112] Assuming that the policy interaction matrix is ​​an 8×8 matrix, after global average pooling, we get a channel importance score vector of length 8 [0.3, 0.4, 0.2, 0.5, 0.3, 0.6, 0.2, 0.7].

[0113] Step S1342: Filter out target channels whose channel importance scores are higher than the importance threshold according to a preset importance threshold, and extract the strategy interaction submatrix corresponding to the target channels.

[0114] The preset importance threshold is 0.4. Based on this importance threshold, target channels with channel importance scores higher than 0.4 are selected, namely, channels 2, 4, 6, and 8.

[0115] Extract the policy interaction submatrices corresponding to these target channels. Assume that the extracted policy interaction matrix is ​​an 8×4 matrix, as shown below: |0.2 0.4 0.6 0.8| |0.4 0.5 0.6 0.7| |...| |...| |...| |...| |...| |...| Step S1343: performing a maximum pooling operation on the strategy interaction sub-matrix to obtain channel salient features.

[0116] In this embodiment, the maximum pooling operation is to select the maximum value in each local area of ​​the strategy interaction submatrix as the representative value of the area. Assume that a maximum pooling operation is performed on the strategy interaction submatrix with a 2×2 window.

[0117] Starting from the upper left corner, take a 2×2 sub-region, such as the sub-region consisting of the first two columns of the first row and the first two columns of the second row: |0.2 0.4| |0.4 0.5| The maximum value 0.5 is selected in this sub-region. In this way, the maximum pooling operation is performed on each 2×2 sub-region of the strategy interaction sub-matrix in turn.

[0118] After the maximum pooling operation, a new matrix is ​​obtained, and the elements of the matrix are the channel salient features. Assume that the channel salient feature matrix obtained is a 4×2 matrix: |0.5 0.6| |0.7 0.8| |...| |...| Step S1344: multiply the channel salient features by the strategy interaction submatrix element by element to generate a weighted strategy interaction feature.

[0119] In this embodiment, the channel significant feature matrix is ​​multiplied by the elements at the corresponding positions of the strategy interaction submatrix. For example, the first element 0.5 of the channel significant feature matrix is ​​multiplied by the element at the corresponding position of the strategy interaction submatrix, that is, 0.5 multiplied by the element 0.2 in the first row and first column of the strategy interaction submatrix to obtain 0.1.

[0120] In this way, each corresponding element of the channel significant feature matrix and the strategy interaction submatrix is ​​multiplied to obtain a weighted strategy interaction feature matrix. Assume that the obtained weighted strategy interaction feature matrix is ​​an 8×4 matrix: |0.1 0.24 0.36 0.48| |0.28 0.35 0.42 0.49| |...| |...| |...| |...| |...| |...| Step S1345: Flatten the weighted strategy interaction features to obtain the strategy fusion vector.

[0121] Flattening is to convert the weighted strategy interaction feature matrix into a one-dimensional vector. Each row element of the weighted strategy interaction feature matrix is ​​connected in sequence to form a long vector.

[0122] Assuming that the weighted policy interaction feature matrix is ​​an 8×4 matrix, after flattening, we get a policy fusion vector of length 32: [0.1, 0.24, 0.36, 0.48, 0.28, 0.35, 0.42, 0.49, ...] Step S135: inputting the strategy fusion vector into a labeling classifier, and outputting the path labeling result, wherein the labeling classifier includes a plurality of fully connected layers and a Softmax layer, and is used to map continuous features to discrete labeling categories.

[0123] In this embodiment, the main function of the annotation classifier is to map the continuous strategy fusion vector to a discrete annotation category, thereby generating a path annotation result.

[0124] First, the strategy fusion vector of length 32 is input to the first fully connected layer of the annotation classifier. Each neuron in the fully connected layer is connected to each element of the input vector, and the input vector is linearly transformed through a series of weight parameters and bias parameters. Assuming that the first fully connected layer has 16 neurons, there will be 32×16 weight parameters and 16 bias parameters. After linear transformation and activation function (such as ReLU function), a vector of length 16 is obtained.

[0125] Next, this vector of length 16 is input to the second fully connected layer. Similarly, the second fully connected layer will also perform linear transformation and activation processing on the input vector. Assuming that the second fully connected layer has 8 neurons, a vector of length 8 is obtained after processing.

[0126] Finally, this vector of length 8 is input to the Softmax layer. The function of the Softmax layer is to convert the input vector into a probability distribution so that each element of the input vector is between 0 and 1, and the sum of all elements is 1. Through the calculation of the Softmax function, a probability vector of length 8 is obtained.

[0127] Assume that the probability vector is [0.1, 0.05, 0.2, 0.05, 0.3, 0.1, 0.1, 0.1]. According to the correspondence between the preset annotation categories and the probability vector elements, the annotation category corresponding to the element with the largest probability is selected as the path annotation result. In this example, the element with the largest probability is 0.3, and its corresponding annotation category may represent a specific route adjustment direction and port resource allocation direction, such as "10 degrees eastward deviation route, priority allocation of berth 3".

[0128] Step S140: generating a set of optimization strategies for scheduling assembly containers according to the path labeling results.

[0129] The path marking results indicate the route adjustment direction and port resource allocation direction of the scheduling sequence, based on which a set of optimization strategies for LCL scheduling can be generated. The specific steps are as follows: Step S141: parsing the route adjustment direction in the path marking result, generating at least one candidate route adjustment plan, each candidate route adjustment plan including a route deviation angle, an adjustment distance and an estimated time loss.

[0130] Assume that the path marking result indicates "10 degrees eastward deviation from the route". Based on this information, multiple candidate route adjustment plans can be generated.

[0131] Solution 1: The route deviation angle is 10 degrees to the east, and the adjustment distance is 50 nautical miles. In order to estimate the time loss, it is necessary to consider the current sailing speed of the ship. Assuming that the average sailing speed of the ship is 20 nautical miles per hour, the time required to adjust 50 nautical miles is 50÷20=2.5 ​​hours, that is, the estimated time loss is 2.5 hours.

[0132] Option 2: The route deviation angle is 10 degrees to the east, and the adjustment distance is 80 nautical miles. Also calculated based on the sailing speed of 20 nautical miles per hour, the time required to adjust 80 nautical miles is 80÷20=4 hours, and the estimated time loss is 4 hours.

[0133] Option 3: The route deviation angle is 10 degrees to the east, and the adjustment distance is 100 nautical miles. The time required to adjust 100 nautical miles is 100÷20=5 hours, and the estimated time loss is 5 hours.

[0134] Step S142: parsing the port resource allocation direction in the path marking result, generating port resource allocation constraint conditions, wherein the constraint conditions include berth occupancy time limit, the number of loading and unloading equipment allocations, and cargo storage area identification.

[0135] Assume that the path marking result indicates "Give priority to berth 3". By further interacting with the port management system, relevant information of berth 3 is obtained.

[0136] Regarding the berth occupation time limit, it is known that Berth No. 3 will be occupied by other ships from 12:00 to 14:00 on January 10, 2024, and will be vacant during other time periods. Therefore, the berth occupation time limit is limited to 12:00 to 14:00 on January 10, 2024.

[0137] In terms of the number of loading and unloading equipment allocated, based on the size of the LCL and the type of cargo, the port management system recommends allocating 2 cranes and 5 forklifts for the loading and unloading operations of the LCL.

[0138] Regarding the identification of the cargo storage area, it was found through inquiry that the cargo in the LCL is suitable for stacking in Area A of the port, so the cargo storage area is identified as Area A.

[0139] Step S143: verifying the feasibility of the candidate route adjustment schemes, and selecting effective adjustment schemes that meet the port resource allocation constraint conditions.

[0140] Verify the feasibility of the candidate route adjustment plan generated above. The specific steps are as follows: Step S1431: extract the berth idle time period in the port resource allocation constraint condition, and match it with the estimated arrival time in the candidate route adjustment scheme, and eliminate the adjustment scheme whose arrival time exceeds the idle time period.

[0141] Taking Plan 1 as an example, assuming that the LCL was originally expected to arrive at the port at 11:00 on January 10, 2024, since the estimated time loss of this plan is 2.5 hours, the adjusted estimated arrival time is 13:30 on January 10, 2024. However, Berth No. 3 is occupied by other ships from 12:00 to 14:00 on January 10, 2024, and the arrival time of this plan is within the occupied period, so this plan does not meet the requirements and is eliminated.

[0142] For Option 2, assuming that the original estimated arrival time remains unchanged, the adjusted estimated arrival time is 15:00 on January 10, 2024, which is within the idle period of berth 3. This option is temporarily retained.

[0143] For Option 3, the adjusted estimated arrival time is 16:00 on January 10, 2024, which is also within the idle period of berth 3. This option is also temporarily retained.

[0144] Step S1432: extract the assigned quantity of the loading and unloading equipment, calculate the quantity of equipment required for the candidate route adjustment scheme, and eliminate the adjustment scheme whose equipment demand exceeds the available quantity.

[0145] The port resource allocation constraint condition recommends the allocation of 2 cranes and 5 forklifts. The loading and unloading equipment required for Plan 2 and Plan 3 after arriving at the port is related to the cargo situation of the LCL. Assume that after evaluation, the number of cranes required for Plan 2 and Plan 3 is 2 and the number of forklifts is 5, which is consistent with the allocation quantity recommended by the port. The number of cranes and forklifts currently available in the port is 3 and 6, respectively. The equipment demand does not exceed the available quantity, so Plan 2 and Plan 3 are retained.

[0146] Step S1433: extract the cargo storage area identifier, verify whether the cargo type corresponding to the candidate route adjustment plan is allowed to enter the target storage area, eliminate the adjustment plan that violates the storage rules, and thus generate a valid adjustment plan that meets the port resource allocation constraint conditions.

[0147] The cargo storage area is marked as Area A. By querying the port's storage rules, it is known that Area A allows the storage of electronic products, textiles, and food in the LCL. The cargo types corresponding to Plan 2 and Plan 3 are all these goods, so both meet the storage rules. Plan 2 and Plan 3 are effective adjustment plans that meet the port resource allocation constraints.

[0148] Step S144: Calculate a comprehensive optimization score according to the estimated time loss and path deviation in the effective adjustment scheme, and generate the assembly container scheduling optimization strategy set by sorting the comprehensive optimization scores.

[0149] For Option 2, the estimated time loss is 4 hours, and the path deviation can be measured by adjusting the distance to 80 nautical miles. For Option 3, the estimated time loss is 5 hours, and the path deviation is 100 nautical miles.

[0150] The comprehensive optimization score can be calculated using a weighted sum method, assuming that the weight of the estimated time loss is 0.6 and the weight of the path deviation is 0.4.

[0151] The comprehensive optimization score of solution 2 = 0.6 × 4 + 0.4 × 80 ÷ 100 (normalizing the path deviation to the range of 0-1) = 2.4 + 0.32 = 2.72 The comprehensive optimization score of plan 3 = 0.6 × 5 + 0.4 × 100 ÷ 100 = 3 + 0.4 = 3.4 According to the comprehensive optimization score sorting from high to low, we get a set of optimization strategies for LCL scheduling, among which Scheme 3 is ranked ahead of Scheme 2.

[0152] Step S150: training the dynamic strategy matching network based on the assembly container scheduling optimization strategy set to update network parameters.

[0153] In order to continuously improve the performance of the dynamic strategy matching network, it is necessary to train it based on the set of optimization strategies for container scheduling and update the network parameters. The specific steps are as follows: Step S151: extracting strategy execution result data from the LCL scheduling optimization strategy set, wherein the strategy execution result data includes time loss, resource utilization rate and cargo damage rate after actual route adjustment.

[0154] When actually executing the solutions in the LCL scheduling optimization strategy set, collect relevant execution result data. Taking Solution 3 as an example, after executing this solution, the actual time loss after the route adjustment is 5.5 hours (due to factors such as the actual ocean environment, it may be different from the estimated situation).

[0155] In terms of resource utilization, the usage time of berth 3 and the usage of loading and unloading equipment are recorded. Assuming that berth 3 is actually used for 3 hours and the port stipulates that the standard usage time of berth 3 is 2.5 hours each time, then the resource utilization rate of berth 3 is 3÷2.5=1.2. For loading and unloading equipment, the two cranes are actually used for 2.5 hours, and the port stipulates that the standard usage time of each crane is 2 hours each time. The resource utilization rate of the cranes is (2×2.5)÷(2×2)=1.25; the five forklifts are actually used for 3 hours, and the port stipulates that the standard usage time of each forklift is 2.5 hours each time. The resource utilization rate of the forklifts is (5×3)÷(5×2.5)=1.2. Taking into account the resource utilization rates of berths and loading and unloading equipment, the average value can be taken as the overall resource utilization rate, that is, (1.2+1.25+1.2)÷3≈1.22.

[0156] In terms of the damage rate of goods, after checking the goods in the container, it was found that one electronic product was slightly damaged. The total number of electronic products in the container was 100, so the damage rate of electronic products was 1÷100=0.01; textiles and food were not damaged. Taking all the goods into consideration, assuming that the total number of goods was 1,000 (including electronic products, textiles and food), the damage rate of goods was 1÷1000=0.001.

[0157] Step S152: construct a strategy effect evaluation function, generate time error loss according to the difference between the time loss after the actual route adjustment and the estimated time loss, generate resource loss according to the deviation between the resource utilization rate and the preset threshold, and generate risk loss according to the cargo damage rate.

[0158] The strategy effectiveness evaluation function is used to measure the quality of strategy execution results by calculating time error loss, resource loss and risk loss for comprehensive evaluation.

[0159] Calculation of time error loss: The estimated time loss is 5 hours, the actual time loss is 5.5 hours, and the time error is 5.5-5=0.5 hours. A simple linear function can be used to calculate the time error loss. Assuming that the time error loss coefficient is 10, then the time error loss = 10×0.5=5.

[0160] Calculation of resource loss: The preset resource utilization threshold is 1. The actual resource utilization is 1.22, and the deviation between the resource utilization and the preset threshold is 1.22-1=0.22. Assuming the resource loss coefficient is 20, then the resource loss = 20×0.22=4.4.

[0161] Calculation of risk loss: The cargo damage rate is 0.001. Assuming the risk loss coefficient is 1000, then the risk loss = 1000 × 0.001 = 1.

[0162] Step S153: weighted sum of the time error loss, resource loss and risk loss to obtain the total training loss.

[0163] Assume that the weight of time error loss is 0.5, the weight of resource loss is 0.3, and the weight of risk loss is 0.2.

[0164] Total training loss = 0.5×5+0.3×4.4+0.2×1=2.5+1.32+0.2=4.02 Step S154: Adopting a gradient descent algorithm to update the parameters of the dynamic strategy matching network based on the total training loss until the total training loss converges to a stable interval.

[0165] The gradient descent algorithm is a commonly used optimization algorithm for finding the minimum value of a function. When training a dynamic policy matching network, the goal is to minimize the total training loss.

[0166] First, the gradient of the total training loss with respect to each parameter in the dynamic policy matching network is calculated. The gradient represents the rate of change of the total training loss in the parameter space. The gradient of each parameter can be calculated through the back-propagation algorithm.

[0167] Then, according to the direction and size of the gradient, the parameters are updated. For example, suppose the parameter update formula is: new parameter = old parameter - learning rate × gradient. The learning rate is a hyperparameter that controls the step size of the parameter update.

[0168] The above process of calculating gradients and updating parameters is repeated until the total training loss converges to a stable range. For example, when the change in the total training loss is less than 0.01 in 10 consecutive iterations, the total training loss is considered to have converged to a stable range. At this time, the training process ends, the parameters of the dynamic policy matching network are updated, and its performance is also improved.

[0169] Through the above steps, the embodiment of the present invention realizes the intelligent and efficient scheduling of LCL shipping, and significantly improves the accuracy and adaptability of the scheduling strategy. Specifically, by obtaining the historical LCL scheduling data set under the target shipping route and performing dynamic path feature extraction processing on it, it is possible to deeply explore the real-time environment association features and path optimization features of each scheduling sequence, and dynamically match the real-time environment association features with the path optimization features based on the preset dynamic strategy matching network, and generate a path marking result with clear route adjustment direction and port resource allocation direction, so that the scheduling strategy can be flexibly adjusted according to the actual environment and path conditions, greatly improving the flexibility and response speed of scheduling. Furthermore, a set of LCL scheduling optimization strategies is generated based on the path marking results, and a dynamic strategy matching network is trained based on the LCL scheduling optimization strategy set to update network parameters, which can not only continuously optimize the scheduling strategy and improve the scheduling efficiency, but also adapt to the ever-changing shipping environment and market demand.

[0170] Figure 2 The schematic diagram shows exemplary hardware and software components of a shipping IoT management system 100 that can implement the concept of the present invention according to some embodiments of the present invention. For example, the processor 120 can be used in the shipping IoT management system 100 and used to perform the functions of the present invention.

[0171] The shipping IoT management system 100 can be a general-purpose server or a special-purpose server, both of which can be used to implement the intelligent scheduling method for shipping LCL of the present invention. Although the present invention only shows one server, for convenience, the functions described in the present invention can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0172] For example, the shipping IoT management system 100 may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and storage media 140 in different forms, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the shipping IoT management system 100 may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The method of the present invention can be implemented according to these program instructions. The shipping IoT management system 100 also includes an input / output (I / O) interface 150 between the computer and other input / output devices.

[0173] For ease of description, only one processor is described in the shipping IoT management system 100. However, it should be noted that the shipping IoT management system 100 in the present invention may also include multiple processors, so the steps performed by one processor described in the present invention may also be performed jointly or individually by multiple processors. For example, if the processor of the shipping IoT management system 100 performs step A and step B, it should be understood that step A and step B may also be performed jointly by two different processors or individually in one processor. For example, the first processor performs step A, the second processor performs step B, or the first processor and the second processor perform steps A and B together.

[0174] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer executable instructions are preset. When a processor executes the computer executable instructions, the above-mentioned intelligent scheduling method for sea freight LCL is implemented.

[0175] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, various features are sometimes combined into one embodiment, drawing or description thereof.

Claims

1. An intelligent dispatching method for sea freight LCL, characterized in that: The method comprises: Acquire a historical LCL scheduling data set under a target shipping route, wherein the historical LCL scheduling data set includes a plurality of scheduling sequences, each scheduling sequence being composed of at least one LCL starting position feature, a target port feature, a loaded cargo attribute feature, and a historical path trajectory feature; Performing dynamic path feature extraction processing on the historical LCL scheduling data set to obtain real-time environment association features and path optimization features of each scheduling sequence; Based on a preset dynamic strategy matching network, dynamic strategy matching is performed on the real-time environment association feature and the path optimization feature to generate a path marking result of the scheduling sequence, wherein the path marking result is used to indicate the route adjustment direction and port resource allocation direction of the scheduling sequence; A set of optimized strategies for scheduling assembly containers is generated according to the path labeling results, and the dynamic strategy matching network is trained based on the set of optimized strategies for scheduling assembly containers to update network parameters.

2. The intelligent dispatching method for LCL shipping according to claim 1 is characterized in that: The dynamic path feature extraction process is performed on the historical LCL scheduling data set to obtain the real-time environment association features and path optimization features of each scheduling sequence, including: Extracting a plurality of trajectory point data units corresponding to the historical path trajectory features from the scheduling sequence, each trajectory point data unit including a location coordinate, a timestamp and corresponding environmental monitoring data; Calling a pre-trained path feature encoder to perform spatiotemporal correlation encoding processing on the multiple trajectory point data units to generate a spatiotemporal fusion vector of the scheduling sequence, wherein the spatiotemporal fusion vector includes a path deviation degree and an environmental fluctuation correlation degree between adjacent trajectory points; Performing port resource matching processing on the starting position characteristics and the target port characteristics in the scheduling sequence to obtain port resource constraint characteristics, wherein the port resource constraint characteristics include at least one of the following: a berth idle period of the target port, an available number of loading and unloading equipment, and a cargo stacking density restriction; splicing the spatiotemporal fusion vector with the port resource constraint feature to generate the real-time environment association feature; The cargo loading attribute characteristics in the scheduling sequence are subjected to risk classification processing to obtain cargo transportation priority characteristics, and the cargo transportation priority characteristics are dynamically weighted fused with the spatiotemporal fusion vector to generate the path optimization characteristics.

3. The intelligent dispatching method for LCL shipping according to claim 2 is characterized in that: The calling of the pre-trained path feature encoder to perform spatiotemporal correlation encoding processing on the plurality of trajectory point data units to generate a spatiotemporal fusion vector of the scheduling sequence includes: For each trajectory point data unit, extract the wind speed, wave height and ocean current direction parameters in the environmental monitoring data to construct an environmental fluctuation matrix; Performing geographic grid encoding on the location coordinates to generate a standardized location encoding vector, and converting the timestamp into a periodic time encoding vector; Input the environmental fluctuation matrix, the standardized position encoding vector and the periodic time encoding vector into the path feature encoder for multi-head attention calculation to obtain a local environmental association vector for each trajectory point data unit; According to the time interval and distance interval between adjacent trajectory point data units, the trajectory point association weight is calculated, and the local environment association vector is subjected to sliding window aggregation based on the trajectory point association weight to generate the spatiotemporal fusion vector.

4. The intelligent dispatching method for sea freight LCL according to claim 2 is characterized in that: The risk classification process of the cargo attribute characteristics in the scheduling sequence to obtain cargo transportation priority characteristics includes: Extracting cargo type, weight distribution and temperature control requirement parameters from the cargo attribute characteristics, and constructing a cargo attribute matrix; Calling a pre-trained risk assessment model to determine a cargo damage risk score, a transportation urgency score, and a priority coefficient according to the cargo attribute matrix; generating an initial priority feature according to a weighted sum of the cargo damage risk score and the transportation urgency score; The initial priority feature is nonlinearly transformed based on the priority coefficient to obtain the cargo transportation priority feature, wherein the nonlinear transformation includes mapping the initial priority feature to a preset priority interval through a Sigmoid function.

5. The intelligent dispatching method for sea freight LCL according to claim 2 is characterized in that: The step of dynamically weighting and fusing the cargo transportation priority feature with the spatiotemporal fusion vector to generate the path optimization feature includes: Expanding the feature dimension of the cargo transportation priority feature to make it consistent with the dimension of the spatiotemporal fusion vector; Calculating the cosine similarity between the expanded cargo transportation priority feature and the spatiotemporal fusion vector to generate a dynamic weight coefficient; Performing a weighted summation of the cargo transportation priority feature and the spatiotemporal fusion vector according to the dynamic weight coefficient to generate a fusion intermediate feature; The fused intermediate features are subjected to dimensionality reduction processing to obtain the path optimization features, wherein the dimensionality reduction processing includes mapping the high-dimensional features to a low-dimensional space through a fully connected layer and retaining the feature components with the largest variance.

6. The intelligent dispatching method for LCL shipping according to claim 1 is characterized in that: The method of performing dynamic strategy matching on the real-time environment association feature and the path optimization feature based on a preset dynamic strategy matching network to generate a path labeling result of the scheduling sequence includes: Inputting the real-time environment-related features into the environment perception branch of the dynamic strategy matching network to generate an environment constraint strategy vector; Inputting the path optimization feature into the path optimization branch of the dynamic strategy matching network to generate a path adjustment strategy vector; Performing cross-attention calculation on the environmental constraint strategy vector and the path adjustment strategy vector to generate a strategy interaction matrix; Filtering key strategy channels according to the value of each element in the strategy interaction matrix, and generating a strategy fusion vector through channel weighted pooling; The strategy fusion vector is input into a labeling classifier, and the path labeling result is output, wherein the labeling classifier includes a plurality of fully connected layers and a Softmax layer, which is used to map continuous features to discrete labeling categories.

7. The intelligent dispatching method for LCL shipping according to claim 6 is characterized in that: The step of screening key strategy channels according to the value of each element in the strategy interaction matrix and generating a strategy fusion vector by channel weighted pooling includes: Performing global average pooling on the strategy interaction matrix along the channel dimension to generate a channel importance score; According to a preset importance threshold, a target channel whose channel importance score is higher than the importance threshold is screened out, and a strategy interaction submatrix corresponding to the target channel is extracted; Performing a maximum pooling operation on the strategy interaction submatrix to obtain channel salient features; Multiplying the channel salient features by the strategy interaction submatrix element by element to generate a weighted strategy interaction feature; The weighted strategy interaction features are flattened to obtain the strategy fusion vector.

8. The intelligent dispatching method for LCL shipping according to claim 1 is characterized in that: The step of generating a set of optimization strategies for LCL scheduling according to the path marking results includes: Analyzing the route adjustment direction in the path marking result to generate at least one candidate route adjustment plan, each candidate route adjustment plan including a route deviation angle, an adjustment distance, and an estimated time loss; Parsing the port resource allocation direction in the path marking result, generating port resource allocation constraint conditions, wherein the constraint conditions include berth occupancy time limit, the number of loading and unloading equipment allocated, and cargo storage area identification; Conduct feasibility verification on the candidate route adjustment plans and select effective adjustment plans that meet the port resource allocation constraints; Calculating a comprehensive optimization score according to the estimated time loss and path deviation in the effective adjustment scheme, and generating the LCL scheduling optimization strategy set by sorting the comprehensive optimization scores; The feasibility verification of the candidate route adjustment schemes to select effective adjustment schemes that meet the port resource allocation constraint conditions includes: Extracting the berth idle time period in the port resource allocation constraint condition and matching it with the estimated arrival time in the candidate route adjustment scheme, and eliminating the adjustment scheme whose arrival time exceeds the idle time period; Extract the assigned quantity of the loading and unloading equipment, calculate the quantity of equipment required for the candidate route adjustment plan, and eliminate the adjustment plan where the equipment demand exceeds the available quantity; The cargo storage area identifier is extracted, and it is verified whether the cargo type corresponding to the candidate route adjustment plan is allowed to enter the target storage area, and the adjustment plan that violates the storage rules is eliminated, thereby generating a valid adjustment plan that meets the port resource allocation constraint conditions.

9. The intelligent dispatching method for LCL shipping according to claim 1, characterized in that: The training of the dynamic strategy matching network based on the assembly box scheduling optimization strategy set to update network parameters includes: Extracting strategy execution result data from the LCL scheduling optimization strategy set, wherein the strategy execution result data includes time loss, resource utilization rate and cargo damage rate after actual route adjustment; Constructing a strategy effect evaluation function, generating a time error loss according to the difference between the time loss after the actual route adjustment and the estimated time loss, generating a resource loss according to the deviation between the resource utilization rate and a preset threshold, and generating a risk loss according to the cargo damage rate; The time error loss, resource loss and risk loss are weighted and summed to obtain the total training loss; A gradient descent algorithm is used to update the parameters of the dynamic strategy matching network based on the total training loss until the total training loss converges to a stable interval.

10. A shipping IoT management system, characterized in that: The shipping IoT management system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the intelligent scheduling method for sea freight LCL as described in any one of claims 1 to 9.

Images