Container position distribution method based on large model driven intelligent agent network

Abstract

The invention provides a container position distribution method based on a large model driven intelligent agent network. The container position distribution method based on the large model driven intelligent agent network comprises the following steps: step 101, carrying out multi-view hybrid coding on a container by adopting a double-flow attention mechanism to obtain overall feature representation of the container; step 102, adopting an autoregression decoder to generate a candidate container set according to the overall feature representation of the container and the current slot position; step 103, based on the candidate container set, using a deterministic physical model to carry out constraint check and index calculation; and 104, carrying out structured evaluation on the container by adopting the large model, and carrying out intelligent sorting. By the adoption of the method, the accurate and reasonable loading sequence and position of the containers can be ensured, and the loading efficiency of the containers is greatly improved.

Description

Technical Field

[0001] This disclosure relates to the field of information technology, and in particular to a container space allocation method for large model-driven intelligent agent networks. Background Technology

[0002] With the rapid growth of global trade and transportation demand, maritime container shipping occupies a core position in the terminal operation system. Against the backdrop of deepening globalization, the scale of international trade continues to expand, with the breadth and frequency of commodity circulation reaching record highs. Maritime shipping, with its significant cost advantages and large-scale carrying capacity, has become the dominant mode of transportation supporting international trade. As the core operating mode of the modern maritime system, container shipping not only significantly improves cargo transshipment efficiency but also effectively reduces transportation costs and cargo damage risks. Its industry growth rate has consistently maintained a leading position, playing an irreplaceable supporting role in the stable operation of the global supply chain.

[0003] As a pillar of global trade, shipping relies heavily on container stowage planning. This problem requires allocating thousands of containers to specific locations on a ship while meeting complex constraints such as ship stability, structural strength, and turnover rate. However, current technologies have the following drawbacks: (1) Strong constraint coupling: stowage is a sequential decision-making process. Early misplacement may lead to serious stability problems or excessive structural stress. Furthermore, physical constraints (such as shear force and bending moment) are non-convex and highly coupled; (2) Lack of interpretability: Existing deep learning or reinforcement learning models are usually black-box models that can only output results but cannot explain the decision-making logic. In the field of maritime safety, this lack of transparency makes it difficult to gain the trust of human stowage operators; (3) Low manual efficiency: Currently, stowage mainly relies on manual stowage, which is time-consuming and prone to errors, making it difficult to cope with large ships and tight schedules. Summary of the Invention

[0004] This disclosure provides a container space allocation method based on a large model-driven agent network, including:

[0005] Step 101: Use a two-stream attention mechanism to perform multi-view hybrid encoding on the container to obtain the overall feature representation of the container;

[0006] Step 102: Use an autoregressive decoder to generate a set of candidate containers based on the overall feature representation of the container and the current slot.

[0007] Step 103: Based on the candidate container set, use a deterministic physical model to perform constraint checks and index calculations;

[0008] Step 104: Use a large model to perform a structured evaluation of the containers and perform intelligent sorting.

[0009] In some embodiments, in step 101:

[0010] For categorical attributes, a learnable embedding matrix is ​​used to map them to vectors, where for destination port attributes, they are mapped to ordinal numbers according to the shipping route order to inject prior knowledge of the unloading order; for numerical attributes, Fourier eigenmaps are used for projection; let v i,k For container c i For the k-th normalized numerical attribute, define a randomly initialized frequency basis vector f. k And the learnable linear projection matrix W num and bias vector b num Construct the sine-cosine eigenvector P i,k And calculate the numerical attribute embedding vector.

[0011] P i,k =Concat[cos(2πf k v i ,k),sin(2πf k v i ,k)]

[0012]

[0013] In the formula, Concat[·] represents vector concatenation. For container c i The embedding representation corresponding to the j-th attribute, OneHot(·) is a one-hot encoding. For container c i The j-th attribute value;

[0014] The specific computational process for constructing the two-stream attention mechanism is as follows: Construct container c i The input sequence matrix X i By utilizing flow-based methods to capture the dependencies between attributes within a single container, the calculation formula is as follows:

[0015]

[0016] In the formula, MHSA row This indicates that attention is calculated in the row direction. The attention representation is used within the container attributes; flow two-dimensionality is used to capture the distribution characteristics of container groups on specific attributes, and the calculation formula is as follows:

[0017]

[0018] In the formula, MHSA col This indicates that attention is calculated in the column direction. This represents the slice of the j-th attribute in the Batch dimension. This represents the distribution characteristics or attention of containers across specific attributes; ultimately, the container c is obtained by fusing these features using a feedforward neural network. i semantic embedding

[0019] In some embodiments, step 102 includes a process of generating a query vector, which includes:

[0020] Loading container position l k Modeled as a virtual container, its feature set A slot Including inherent physical properties and spatial neighborhood-based properties Deduced contextual inference properties; generate query vector q k =StowFormer(A slot Construct the decoder input matrix Z; k :

[0021]

[0022] In the formula, Z k For the input of the decoder at step k, e start The initial marker vector, To determine the container vector to be loaded at each step, P pos It is a sinusoidal coding matrix;

[0023] The process of generating a candidate container set includes: Z k The input is fed into the L-layer Transformer decoder, and the last position of the output is selected as the context decision vector h. k ; Calculate unassigned container c j Choice probability distribution:

[0024]

[0025] In the formula, mask hard (·) represents a hard constraint, used to ensure that the hard constraint is not violated. k For the current container position, c j For candidate containers, s k,j Let S be a scalar representing the matching score between container slots and containers; select the N containers with the highest probabilities to form a candidate container set S. cand .

[0026] In some embodiments, in step 103, the deterministic physical model includes a ship stability assessment, specifically calculated as follows: assuming the current ship displacement is Δ... curr The sum of the vertical moments of the ship's center of gravity is If the added weight is w jFor candidate containers, calculate the updated center of gravity height (KG). new And the center radius KM new :

[0027]

[0028] KM new =f KM (Δ curr +w j )

[0029] Finally, calculate the ship's initial metacentric height:

[0030] GM new (c j ) = KM new -KG new .

[0031] In some embodiments, in step 104, the intelligent sorting process includes: constructing a compound prompt word P. k :

[0032]

[0033] In the formula, D state It describes the current state, including the current location of the loading containers and the overall stability of the ship. goal This is the priority strategy for this loading task. j The physical evaluation vector from the previous step is converted into a natural language description, C format The model is required to output a specific structure.

[0034] By adopting the method disclosed herein, the accurate and reasonable loading sequence and position of containers can be ensured, thereby significantly improving the loading efficiency of containers. Attached Figure Description

[0035] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and elements are not necessarily drawn to scale.

[0036] Figure 1 A schematic flowchart of a container space allocation method based on a large model-driven agent network, according to some embodiments, is shown.

[0037] Figure 2 A two-stream attention encoding process according to some embodiments is illustrated.

[0038] Figure 3The process of constructing large model prompts according to some embodiments is illustrated. Detailed Implementation

[0039] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0040] It should be understood that the various steps described in the method embodiments of this disclosure can be performed in sequence and / or in parallel. Furthermore, method embodiments may include additional steps and / or omit the steps shown. The scope of this disclosure is not limited in this respect.

[0041] The term "comprising" and its variations as used herein are open-ended inclusions, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". Definitions of other terms will be given in the description below.

[0042] It should be noted that the concepts of "first" and "second" mentioned in this disclosure are used only to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0043] It should be noted that the use of the word "a" in this disclosure is illustrative rather than restrictive, and those skilled in the art should understand that it should be understood as "one or more" unless otherwise expressly indicated in the context.

[0044] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

[0045] This disclosure provides a container location allocation method based on a large model-driven agent network. By determining the allocation position of containers within the ship based on the properties of containers at the terminal, the method can improve the performance and efficiency of container location allocation.

[0046] refer to Figure 1 This paper presents a container space allocation method based on a large model-driven agent network, the method including:

[0047] Step 101: Use a two-stream attention mechanism to perform multi-view hybrid encoding on the container to obtain the overall feature representation of the container;

[0048] Step 102: Use an autoregressive decoder to generate a set of candidate containers based on the overall feature representation of the container and the current slot.

[0049] Step 103: Based on the candidate container set, use a deterministic physical model to perform constraint checks and index calculations;

[0050] Step 104: Use a large model to perform a structured evaluation of the containers and perform intelligent sorting.

[0051] In some instances, step 101 specifically involves differentiated mapping and feature decoupling of heterogeneous data. For the destination port category attribute, this embodiment does not directly perform one-hot encoding, but instead implements ordinal mapping based on the shipping plan. For continuous attributes, this embodiment uses Fourier feature mapping to address the problem of neural networks being insensitive to high-frequency information. This implementation allows the model to identify both macroscopic differences between light and heavy containers and subtle differences, such as between "20.1 tons" and "20.5 tons," thereby precisely controlling torque during loading. In the feature extraction stage, the system performs parallel attention calculations in two dimensions, focusing not only on identifying mutually exclusive or strongly correlated features within the container but also on statistical distribution features.

[0052] Specifically, assume the container set is C = {c1, c2, ..., c} M For each container c in the terminal yard i ∈C has two main categories of attributes: categorical attributes and numerical attributes. Categorical attributes include container type, size, and destination port; numerical attributes include container weight.

[0053] For categorical attributes, a learnable embedding matrix E is used. cat Map it to a vector:

[0054]

[0055] In the formula, For container c i The embedding representation corresponding to the j-th attribute, OneHot(·) is a one-hot encoding. For container c i The j-th attribute value. In particular, for the destination port in the category attribute, it is mapped to the ordinal z∈{1,...,Z} according to the specific route order, thereby explicitly injecting prior knowledge of the unloading order.

[0056] For numerical attributes, Fourier eigenmaps are used for projection to capture their nonlinear relationships and high-frequency details in high-dimensional space. Assume v i,k For container c i For the k-th normalized numerical attribute, define a randomly initialized frequency basis vector. And a learnable linear projection matrix W num and bias vector b num Then the scalar attribute value v i,k Expanded into sine-cosine eigenvectors:

[0057] P i,k =Concat[cos(2πf k v i ,k),sin(2πf k v i ,k)]

[0058] In the formula, Concat[·] denotes vector concatenation. Subsequently, a linear transformation is applied to map it to the model dimension, yielding the numerical attribute embedding vector:

[0059]

[0060] In the formula, W is the embedded representation of the numerical attribute. num b num All of these are learnable parameters.

[0061] The processed categorical attribute embeddings and numerical attribute embeddings are then concatenated, and a learnable global feature identifier, denoted as e, is introduced. [CLS] Ultimately targeting container c i The complete input sequence matrix is ​​obtained:

[0062]

[0063] in For container c i The characteristics are represented.

[0064] To decouple the attribute constraints within containers from the group distribution patterns among containers, a two-stream attention mechanism is designed. Let there be M containers in the current batch, and its input tensor be...

[0065] The purpose of flow one is to capture the dependencies between various attributes within a single container, and its formula is expressed as:

[0066]

[0067] In the formula, MHSA row This indicates that attention is calculated in the row direction, and the output is... This represents the attention representation within the container attributes.

[0068] The purpose of Flow 2 is to capture the distribution characteristics of container groups based on specific attributes, such as comparing the "port of discharge" attribute of all containers to identify which containers belong to the same destination group. The formula is expressed as:

[0069]

[0070] In the formula, MHSA col This indicates that attention is calculated in the column direction. This represents the slice of the j-th attribute in the Batch dimension.

[0071] Finally, the features processed by the two streams are fused and passed through a feedforward neural network to obtain e. [CLS] The vector at the corresponding position is used as container c. i semantic embedding

[0072]

[0073] In the formula, FFN is a feedforward neural network, representing the overall feature representation of the output.

[0074] Figure 2 The diagram illustrates a two-stream attention encoding process according to some embodiments, in which the input matrix is ​​first processed, intra-attribute attention is calculated in one stream, inter-group attention is calculated in another stream, and finally the two streams are fused together using a feedforward neural network to obtain the final container semantic embedding or overall feature representation of the container.

[0075] In some instances, step 102, when implemented, hinges on how to translate physical vacancies into query intents that the model can understand. This involves constructing the query vector q. k In addition to the physical limitations of the input slot itself, this embodiment also includes dynamic perception of the surrounding environment. For example, when the slot to be loaded... k When located at the top deck, the system retrieves the attributes of the containers already loaded directly below it. If the destination port of the containers below is P5, the system deduces that the ideal attributes for that slot should be either "destination port ≥ P5" or "weight ≤ weight below". These derived soft constraints are encoded into q. k This naturally leads to a subsequent generation process that tends to satisfy the principles of "last in, first out" and "heavy on the bottom, light on the top." When calculating the selection probability P(c... j In this embodiment, a dynamic hard constraint mask is introduced. hardThe system maintains a real-time table of available container statuses. Before the Softmax calculation, for all candidate containers with size mismatches (e.g., a 40-foot container placed in a 20-foot slot) or special type conflicts (e.g., a refrigerated container placed in a regular slot), their Logit values ​​are forcibly set to -∞. This ensures that the Top-N candidate set S output by the generated model is accurate. cand It is perfectly legal in terms of geometry and device compatibility, avoiding unnecessary waste of subsequent physics calculations.

[0076] Specifically, in order to measure the matching degree between container slots and containers within the same semantic space, the container slots to be loaded are... k Modeled as a virtual shipping container, and the query vector q is generated using a two-stream encoder with shared weights. k Assume there are 1 container slot available for the current container to be allocated. k Its feature set A slot It consists of two parts: the first is the inherent physical properties, representing the limitations of the container space inside the ship, including size restrictions and whether it has a refrigeration socket, etc.; the second is the context-derived properties, based on l k Containers have been deployed within the space neighborhood. The derived desired properties, such as the ideal unloading port for this location based on the unloading port of the container below, are used to satisfy the last-in-first-out (LIFO) and first-in-last-out (FIFO) principles.

[0077] First, using the two-stream encoder trained in step 1, the virtual attribute set is mapped to a query vector.

[0078] q k =StowFormer(A slot )

[0079] In the formula, StowFormer(·) is the two-stream encoder based on the two-stream attention mechanism in step 1, and A slot For box position attributes, q k In semantic space, it represents the characteristics that the ideal container for this location should possess.

[0080] The decoder's input is a sequence containing historical loading information and the current query intent. Let the sequence of containers loaded in the first k-1 steps be denoted as . Construct the decoder input matrix at step k:

[0081]

[0082] In the formula, Z k For the input of the decoder at step k, e start The initial marker vector, The vector of containers to be loaded is determined in each step, q, calculated in step 1. k P is the query vector for the current box location.pos It is a sinusoidal coding matrix used to inject sequence order information.

[0083] The completed Z will then be k The input is fed into a Transformer decoder consisting of L stacked layers:

[0084]

[0085] In the formula, LayerNorm() is the layer normalization, Z (l) This represents the output of the L-layer decoder. MaskedMHSA() is a multi-head self-attention mechanism for masking, and FFN is a feedforward neural network. Finally, the last position of the output is selected as the context decision vector for the current step.

[0086]

[0087] h k The aggregation of historical configuration strategies and current container space requirements forms the core basis for generating candidate proposals.

[0088] Next, the context decision vector is matched with the embedding vectors of all available containers (i.e., unassigned containers) from step one:

[0089]

[0090] In the formula, h k For the context decision variables of the current step, For the unassigned container embedding vectors, this formula essentially calculates the cosine similarity between "historical context requirements" and "candidate container features" in the semantic space. The final container selection probability distribution is calculated using the softmax function.

[0091]

[0092] In the formula, mask hard (·) represents a hard constraint, used to ensure that the hard constraint is not violated. k For the current container position, c j For candidate containers, s k,j The scalar represents the matching score between container slots and containers.

[0093] Finally, based on the calculated probability distribution, the N containers with the highest probabilities are selected to form a candidate set, which serves as the input for subsequent physical evaluation and agent decision-making.

[0094]

[0095] In the formula, S candLet N be the top N candidate containers with the highest scores, and C be the set of candidate containers.

[0096] In some instances, step 103 serves as a bridge between large-scale model predictions and physical safety, focusing on transforming complex mechanical calculations into standardized numerical indicators. In evaluating the initial stability height GM, this embodiment not only calculates the final value but also considers the incremental impact. For each candidate container, the system simulates the overall ship moment change ΔM after loading. vert =w j ·z(l k ), and then update the center of gravity KG new If a candidate bin causes the GM value to deviate from the ideal range (e.g., from 1.0m to 0.8m), this index will be affected in the evaluation vector r. k,j This is reflected as a high-risk value. For operational efficiency assessment, the system implements strict sequence checks. For the current slot l k The system compares candidate boxes c j Port(c) of the unloading port j ) and the container c below it below The ordinal number of the port of discharge. If Port(c j ) <Port(c below This means that if the first unloaded box presses down on the second unloaded box, the system will adjust the box-turning index v. acc Set it to 1 if it is not set to 0 otherwise. This discretized index provides a clear "penalty signal" for subsequent decisions in larger models.

[0097] Specifically, hard constraint verification is performed using a deterministic physical model. For each container in the obtained candidate container set, it is placed in the current target container position to construct a temporary ship stowage state.

[0098] Next, a ship stability assessment is performed based on the current loading status, specifically calculating the ship's initial metacentric height. Assume the current ship displacement is Δ. c urr, the sum of the vertical moments of the ship's center of gravity is If candidate containers are added to the stowage, the updated ship center of gravity height will be:

[0099]

[0100] In the formula, KG new To determine the center height after adding the new container, w j To include the weight of the container, z(l k () represents the vertical center of gravity height. This is the sum of the vertical moments of the ship's center of gravity at the moment of impact.

[0101] The current geocentric radius can be obtained by consulting the ship's hydrostatic parameters table:

[0102] KM new =f KM (Δ curr +w j )

[0103] In the formula, f KM (·) represents the displacement polynomial function, and finally, the initial metacentric height of the ship is obtained:

[0104] GM new (c j ) = KM new -KG new

[0105] The current container handling rate and operational efficiency are then evaluated. Specifically, it is checked whether placing containers into the designated spaces violates the last-in, first-out (LIFO) principle, i.e., whether it obstructs the unloading of containers below. Let c bellow Calculate the penalty for tilting the container directly below the container bay:

[0106]

[0107] In the formula, Port(·) represents the destination port of the container. When the container c above... j Compared to container c below below If unloading occurs earlier, the containers below will be "stuck at the port," resulting in a container repositioning operation.

[0108] Finally, for each candidate container c j Generate a structured physical evaluation vector:

[0109] r k,j =[GM new (c j ),v acc (c j ),I valid (c j )]

[0110] In the formula, GM new (c j ) represents the initial metacentric height of the ship, v acc (c j ) is the penalty indicator for overturning boxes, I valid (c j ) is a binary compliance indicator, which is 1 only when all hard constraints are satisfied and 0 otherwise. k,j The complexity of the physical world is compressed into numerical indicators for subsequent understanding of large-scale intelligent models.

[0111] In some embodiments, numerical tradeoffs are transformed into semantic reasoning tasks through cue engineering. In one embodiment, the cue word Pk The construction is dynamic. Figure 3 The process of constructing large model cue words according to some embodiments is illustrated, including current task description, strategic objectives, candidate box analysis, and output format.

[0112] Specifically, to enable large models to handle specialized loading tasks, a composite cue is constructed that includes structured data and unstructured instructions:

[0113]

[0114] In the formula, D state It describes the current state, including the current location of the loading containers and the overall stability of the ship. goal This is the priority strategy for this loading task. j The physical evaluation vector from the previous step is converted into a natural language description, C format The model is required to output a specific structure.

[0115] Then the constructed prompt word P k Input to large model M LLM Generate through reasoning:

[0116] Y k =M LLM (P k ;θ)

[0117] In the formula, θ represents the model parameters, and Y... k The decision index is used as the output of the large model, which ultimately yields the container sequence.

[0118] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features disclosed in this disclosure that have similar functions.

[0119] Furthermore, while the operations are described in a specific order, this should not be construed as requiring these operations to be performed in the specific order shown or in a sequential order. In certain environments, multitasking and parallel processing may be advantageous. Similarly, while several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of this disclosure. Certain features described in the context of individual embodiments may also be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may also be implemented individually or in any suitable sub-combination in multiple embodiments.

[0120] Although the subject matter has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely illustrative examples of implementing the claims.

Claims

1. A container space allocation method based on a large model-driven agent network, characterized in that, include: Step 101: Use a two-stream attention mechanism to perform multi-view hybrid encoding on the container to obtain the overall feature representation of the container; Step 102: Use an autoregressive decoder to generate a set of candidate containers based on the overall feature representation of the container and the current slot. Step 103: Based on the candidate container set, use a deterministic physical model to perform constraint checks and index calculations; Step 104: Use a large model to perform a structured evaluation of the containers and perform intelligent sorting.

2. The container space allocation method based on a large model-driven agent network according to claim 1, characterized in that, In step 101: For categorical attributes, a learnable embedding matrix is ​​used to map them to vectors, where for destination port attributes, they are mapped to ordinal numbers according to the shipping route order to inject prior knowledge of the unloading order; for numerical attributes, Fourier eigenmaps are used for projection; let v i,k For container c i For the k-th normalized numerical attribute, define a randomly initialized frequency basis vector f. k And the learnable linear projection matrix W num and bias vector b num Construct the sine-cosine eigenvector P i,k And calculate the numerical attribute embedding vector. P i,k =Concat[cos(2πf k v i ,k),sin(2πf k v i ,k)] In the formula, Concat[·] represents vector concatenation. For container c i The embedding representation corresponding to the j-th attribute, OneHot(·) is a one-hot encoding. For container c i The j-th attribute value; The specific computational process for constructing the two-stream attention mechanism is as follows: Construct container c i The input sequence matrix X i By utilizing flow-based methods to capture the dependencies between attributes within a single container, the calculation formula is as follows: In the formula, MHSA row This indicates that attention is calculated in the row direction. Attention representation within container attributes; The formula for calculating the distribution characteristics of container groups on specific attributes using flow-second detection is as follows: In the formula, MHSA col This indicates that attention is calculated in the column direction. This represents the slice of the j-th attribute along the Batch dimension. This represents the distribution characteristics or attention of containers across specific attributes; ultimately, the container c is obtained by fusing these features using a feedforward neural network. i semantic embedding 3. The container space allocation method based on a large model-driven agent network according to claim 2, characterized in that, Step 102 includes the process of generating a query vector, which includes: Loading container position l k Modeled as a virtual container, its feature set A slot Including inherent physical properties and spatial neighborhood-based properties Deduced contextual inference properties; generate query vector q k =StowFormer(A slot Construct the decoder input matrix Z; k : In the formula, Z k For the input of the decoder at step k, e start The initial marker vector, To determine the container vector to be loaded at each step, P pos It is a sinusoidal coding matrix; The process of generating a candidate container set includes: Z k The input is fed into the L-layer Transformer decoder, and the last position of the output is selected as the context decision vector h. k ; Calculate unassigned container c j Choice probability distribution: In the formula, mask hard (·) represents a hard constraint, used to ensure that the hard constraint is not violated. k For the current container position, c j For candidate containers, s k,j Let S be a scalar representing the matching score between container slots and containers; select the N containers with the highest probabilities to form a candidate container set S. cand .

4. The container space allocation method based on a large model-driven agent network according to claim 3, characterized in that, In step 103, the deterministic physical model includes a ship stability assessment, the specific calculation process of which is as follows: assuming the current ship displacement is Δ curr The sum of the vertical moments of the ship's center of gravity is If the added weight is w j For candidate containers, calculate the updated center of gravity height (KG). new And the center radius KM new : KM new =f KM (Δ curr +w j ) Finally, calculate the ship's initial metacentric height: GM new (c j )6KM new -KG new 。 。 5. The container space allocation method based on a large model-driven agent network according to claim 4, characterized in that, In step 104, the intelligent sorting process includes: constructing a compound prompt word P k : In the formula, D state It describes the current state, including the current location of the loading containers and the overall stability of the ship. goal This is the priority strategy for this loading task. j The physical evaluation vector from the previous step is converted into a natural language description, C format The model is required to output a specific structure.

Images