A method and system for optimizing the placement of goods in a cold storage refrigeration process simulation

Abstract

The present application relates to cold storage refrigeration simulation technical field, especially to a kind of cold storage refrigeration process simulation goods placement form optimization method and system.The method includes the following steps: obtaining cold storage three-dimensional model, refrigeration pipeline and goods basic data, sequentially carry out mixed mesh division, goods space mapping and multi-domain region division, and build the refrigeration physical model capable of representing fluid flow, pipe wall heat conduction and goods heat diffusion relationship;Boundary assignment and coupled solution are completed by combining equipment operation data and environmental monitoring data, and the refrigeration simulation data of the temperature field, velocity field and pressure field of cold storage changing with time are obtained, and the refrigeration space-time field and refrigeration state characteristics are extracted accordingly;Establish the goods-dock correlation between goods attributes, spatial location and refrigeration state, generate placement scheme using multi-head self-attention model, then perform simulation evaluation to obtain energy consumption prediction results.The present application can improve the uniformity of goods temperature distribution, energy consumption prediction accuracy and placement reliability.

Description

Technical Field

[0001] This invention relates to the field of cold storage refrigeration simulation technology, and in particular to a method and system for optimizing the placement of goods in a cold storage refrigeration process simulation. Background Technology

[0002] Cold storage facilities typically utilize refrigeration pipelines, evaporators, and fans to create a low-temperature circulating airflow, thereby reducing and maintaining the temperature of goods inside. The volume, shape, thermal properties, stacking method, and spatial location of goods can alter the effective circulation space and airflow channels within the cold storage. When goods are stacked too densely, the direction of the stacks does not match the airflow direction, or some passageways are blocked, it can easily lead to cold air bypassing, airflow stagnation, insufficient cold coverage, and slow cooling of the goods. Therefore, the way goods are arranged not only affects the utilization rate of storage space but also directly impacts the temperature distribution, cooling efficiency, and refrigeration energy consumption of the cold storage.

[0003] Current cargo placement methods are mostly determined based on cargo size, storage capacity, inbound / outbound sequence, and human experience, making it difficult to accurately consider factors such as refrigeration pipeline location, cold air propagation direction, cargo thermal inertia, and heat transfer at the fluid-solid interface. Although some technologies use 3D cold storage models and numerical simulation methods to calculate temperature and velocity fields, they require modifying the geometric model, re-meshing, and repeating the solution for different placement schemes. The calculation process is complex, and the simulation results are mostly used for post-event display or manual comparison, making it difficult to directly translate them into cargo placement schemes. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a method and system for optimizing the placement of goods in a cold storage refrigeration process simulation, thereby resolving at least one of the aforementioned technical issues.

[0005] This application provides a method for optimizing the placement of goods in a simulated cold storage refrigeration process, comprising the following steps: S1. Obtain the 3D model data of the cold storage, refrigeration pipeline data, and basic cargo data; perform mesh generation based on the 3D model data and refrigeration pipeline data to obtain mixed mesh data of the cold storage; perform spatial mapping on the mixed mesh data of the cold storage based on the basic cargo data to obtain cargo spatial mapping data. S2. Based on the cold storage hybrid grid data and cargo space mapping data, the regions are divided to obtain multi-domain calculation data; based on the multi-domain calculation data, a model is constructed to obtain the refrigeration physical model; S3. Acquire equipment operation data and environmental monitoring data; assign boundary values ​​to the refrigeration physical model based on the equipment operation data and environmental monitoring data to obtain initial model data; perform coupled solution based on the initial model data to obtain refrigeration simulation data; S4. Extract field quantities from the refrigeration simulation data to obtain refrigeration spatiotemporal field data; perform state identification based on the refrigeration spatiotemporal field data to obtain refrigeration state characteristic data. S5. Based on the refrigeration status characteristic data, cargo space mapping data and cargo basic data, the cargo location association data is constructed. S6: Perform multi-head self-attention relationship learning based on cargo location association data to obtain placement scheme data; perform back-substitution simulation based on placement scheme data and refrigeration physical model to obtain placement simulation data; evaluate the scheme based on placement simulation data to obtain energy consumption prediction data.

[0006] This invention employs hybrid mesh generation and multi-domain mapping to represent the three-dimensional space of the cold storage, refrigeration pipelines, and cargo-occupied areas, enabling accurate characterization of the spatial boundaries and heat transfer relationships between the fluid region, pipe wall region, and cargo region. By combining equipment operation data and environmental monitoring data to assign boundary values ​​and solve the fluid-structure-thermal coupling problem in the refrigeration physics model, the spatiotemporal variation characteristics of temperature, velocity, and cold energy propagation within the cold storage can be obtained. Refrigeration states such as cold energy coverage, temperature uniformity, and local heat accumulation are extracted through state recognition, and the correlation between cargo attributes, cargo location, and refrigeration states is established. This allows the multi-head self-attention model to identify the impact of different cargo placement methods on airflow organization and heat transfer. Substituting the generated placement scheme back into the refrigeration physics model for simulation verification reduces the scheme bias caused by relying solely on historical sample predictions, improving the physical rationality of the cargo placement scheme and the accuracy of energy consumption prediction.

[0007] Preferably, the region division is as follows: Grid labeling is performed based on the cold storage hybrid grid data and cargo space mapping data to obtain grid attribute labeling data; Perform domain occupancy determination on the grid attribute marker data to obtain domain occupancy determination data; Based on the domain determination data, fluid-structure partitioning is performed to obtain fluid-structure partitioning data; Interface identification data is obtained by performing interface identification based on fluid-structure partitioning data. Based on fluid-structure partitioning data and interface recognition data, regions are aggregated to obtain multi-domain computational data.

[0008] In this invention, by marking the cold storage hybrid grid and cargo space mapping results grid by grid, the fluid domain, cargo domain, refrigeration pipe wall domain, and storage structure domain corresponding to different grid units can be clearly identified, avoiding confusion in area attribution caused by cargo stacking and overlapping refrigeration pipelines. Through domain determination and fluid-structure partitioning, the grids involved in flow calculations can be separated from those involved in heat conduction calculations, while maintaining the topological continuity between each calculation region. Combined with interface identification, the heat transfer boundaries between fluid and cargo, fluid and pipe walls, and different material regions can be accurately extracted. The resulting multi-domain calculation data can provide a computational foundation for subsequent calculations, thereby reducing numerical divergence and local heat transfer distortion caused by misclassified grids, and improving the stability and accuracy of cold storage temperature field, velocity field, and cargo internal temperature simulation.

[0009] Preferably, the model construction is specifically as follows: Attribute configuration is performed based on multi-domain computational data to obtain multi-domain physical property data; Fluid and solid modeling are performed on multi-domain physical property data and multi-domain computational data to obtain fluid equation data and heat transfer equation data. Interface coupling is performed based on fluid equation data, heat transfer equation data, and multi-domain calculation data to obtain coupled equation data. Boundary configuration is performed based on coupled equation data and multi-domain computation data to obtain boundary condition data; The refrigeration physics model is obtained by integrating the coupled equation data and boundary condition data.

[0010] This invention configures physical properties such as density, viscosity, thermal conductivity, and specific heat capacity for the fluid region, pipe wall region, cargo region, and warehouse structure region respectively, enabling different computational domains to possess heat transfer and flow properties matching those of actual materials. By establishing separate fluid control equations and solid heat transfer equations, and coupling temperature continuity and heat flux conservation at the region boundaries, temperature jumps or heat transfer distortions at the fluid-solid interface can be reduced. Combined with the configuration of refrigeration inlet, outlet, warehouse wall heat exchange, and region boundary settings, the cold air circulation, pipe wall heat conduction, and internal heat diffusion processes of the cargo can be accurately described. The resulting refrigeration physics model has a unified equation system and boundary relationships, improving the accuracy and convergence stability of simulating the temperature field, velocity field, and cargo cooling process in the cold storage.

[0011] Preferably, the coupled solution is specifically as follows: Discretize the equations based on the initial model data to obtain the discrete model data; Momentum prediction is performed based on the discrete data of the model to obtain the predicted flow field data; Pressure velocity correction is performed on the predicted flow field data to obtain the corrected flow field data; Turbulent update is performed on the corrected flow field data to obtain turbulent flow field data; Heat transfer coupling is performed on turbulent flow field data and discrete model data to obtain fluid-structure interaction field data; The refrigeration simulation data is obtained by iterative convergence based on the fluid-structure interaction field data.

[0012] This invention unifies and discretizes the continuity equation, momentum equation, turbulence equation, and energy equation in the refrigeration physical model, transforming the continuous flow and heat transfer process into a set of iteratively solvable algebraic equations. By utilizing momentum prediction and pressure-velocity correction, the coupling relationship between the pressure and velocity fields can be coordinated, reducing mass non-conservation and local velocity deviations. Through turbulence updates, turbulent viscosity and cold air diffusion capacity can be dynamically corrected, improving the characterization accuracy of backflow, eddies, and attenuation phenomena at the air supply terminal. Coupled with the turbulent flow field and a multi-region discrete model for heat transfer, the fluid convection heat transfer, pipe wall heat conduction, and internal heat diffusion processes of the cargo can be solved simultaneously. The refrigeration simulation data obtained after iterative convergence exhibits good numerical stability and physical consistency.

[0013] Preferably, the state recognition specifically includes: Gradient calculations are performed based on the spatiotemporal field data of the cooling process to obtain the spatiotemporal field gradient data; Fluctuation extraction is performed on the spatiotemporal field data of cooling and the spatiotemporal field gradient data to obtain the temporal fluctuation data of field quantity. Based on the spatiotemporal field gradient data and the temporal fluctuation data of the field quantity, uniform evaluation is performed to obtain the uniform characteristic data of the field quantity. Cold energy is determined by analyzing the spatial-temporal field data and the uniformity characteristic data of the cooling field to obtain cold energy coverage characteristic data. Anomaly identification is performed based on spatiotemporal field gradient data, field temporal fluctuation data, and cold load coverage characteristic data to obtain refrigeration anomaly characteristic data. Based on the uniformity characteristic data of field quantity, the characteristic data of cold quantity coverage, and the characteristic data of refrigeration anomaly, the characteristic data of refrigeration status are collected to obtain the characteristic data of refrigeration status.

[0014] This invention utilizes gradient calculation and temporal fluctuation extraction of refrigeration spatiotemporal field data to simultaneously characterize the spatial abrupt changes and temporal trends of field quantities such as temperature, velocity, and pressure, avoiding judgment of refrigeration status based solely on a single moment or local measurement point. By combining gradient distribution and fluctuation characteristics for uniformity assessment, it can quantitatively identify uneven temperature distribution, airflow imbalance, and localized cooling capacity attenuation within the cold storage. Through cooling capacity coverage determination, it can clearly define effective refrigeration areas, insufficient cooling capacity areas, and cooling capacity propagation boundaries. The system performs anomaly identification, enabling early detection of abnormal states such as localized heat accumulation, airflow stagnation, sudden temperature rises, and cooling blind spots. The resulting refrigeration status characteristic data reflects the stability, uniformity, and anomaly risks of the refrigeration process.

[0015] Preferably, the multi-head self-attention relationship learning specifically includes: Stage identification is performed based on refrigeration status characteristic data to obtain refrigeration stage data; The cargo location association data and refrigeration stage data are segmented into stages to obtain stage placement sequence data; Based on the stage placement sequence data, correlation modeling is performed to obtain stage cargo correlation data; Cross-segment comparisons were performed based on the cargo correlation data and refrigeration stage data to obtain correlation mutation data; Head reconfiguration is performed based on associated mutation data to obtain attention head configuration data; Multi-head learning is performed based on attention head configuration data and stage placement sequence data to obtain placement scheme data.

[0016] This invention identifies stages of refrigeration status characteristics and segments the cargo placement sequence according to different operating conditions such as refrigeration start-up, stable refrigeration, shutdown insulation, and defrosting. This avoids the distortion of relationships caused by using a fixed correlation matrix to uniformly represent the entire process. Through stage-based cargo correlation modeling and cross-segment comparison, it can identify abrupt changes in the relationships between cargo, such as cooling synergy and airflow obstruction, leading to heat interference and localized warming. Based on the correlation abrupt change results, attention points are dynamically reconfigured, allowing different attention points to focus on the dominant airflow propagation, surface heat transfer, internal thermal inertia, and heat accumulation relationships in the current stage. The resulting placement scheme can more accurately adapt to changes in the refrigeration stage, reduce the risk of insufficient local cooling and heat accumulation, and improve the uniformity of cargo temperature distribution, the physical rationality of the scheme, and the accuracy of refrigeration energy consumption prediction.

[0017] Preferably, the head rematch specifically involves: Based on the associated mutation data, local reconfiguration and time-delay reconfiguration are performed to obtain the first attention head configuration data and the second attention head configuration data, respectively. Head alignment is performed based on the first attention head configuration data and the second attention head configuration data to obtain attention head alignment data; Adaptation data is obtained by performing adaptation determination based on attention head alignment data and associated mutation data; Conflict resolution is performed based on head adaptation data to obtain attention head configuration data.

[0018] This invention limits the scope of goods, storage areas, and propagation links involved in associated mutations through localized reconfiguration, and determines the activation time and effective range of attention heads corresponding to different physical relationships through time-delay reconfiguration, thus avoiding the triggering of synchronous adjustments of all attention heads by localized cooling state changes. The configuration data of the first and second attention heads are aligned to ensure consistency in the learning objects, action areas, and start / stop sequences of each attention head. Adaptation judgment is performed based on associated mutation data to filter out configuration results that do not match in area range, association type, or response time delay. Conflict resolution addresses the problem of the same attention head being repeatedly occupied by multiple mutation relationships or having contradictory configurations, forming attention head configuration data that balances spatial locality and temporal lag. This improves the adaptability of multi-head learning to cooling stage switching, localized heat accumulation, and cold energy propagation mutations, and reduces the disruption of stable associations and distortion of placement schemes caused by global reconfiguration.

[0019] Preferably, the local reconfiguration specifically includes: Mutation location is performed based on associated mutation data to obtain mutation node data; The domain boundaries are expanded based on the mutation node data and associated mutation data to obtain mutation scope data; Link tracing is performed based on mutation scope data and associated mutation data to obtain mutation association chain data; Head domain mapping is performed based on mutation association chain data to obtain attention head data to be matched; Configuration inheritance data is obtained by performing configuration inheritance based on the attention head data to be configured and the mutation scope data. Incremental updates are performed based on the configuration inheritance data and associated mutation data to obtain the first attention head configuration data.

[0020] This invention, by locating mutations in associated mutation data, accurately identifies cargo nodes where refrigeration relationships change, preventing the misinterpretation of local anomalies to all cargo locations. Combining domain boundary expansion and link tracing, the actual impact range of mutation nodes and their associated links formed by airflow propagation and heat diffusion can be identified, thus establishing a local domain with spatial boundaries and propagation direction constraints. Through head domain mapping, mutation association chains can be accurately matched to attention heads responsible for the corresponding cargo areas and physical relationships. Subsequently, configuration inheritance preserves the original stable attention configuration of non-mutation areas, performing incremental updates only on the affected parts. This reduces invalid parameter adjustments and stable relationship disruptions caused by global reconfiguration, improves the local adaptability, computational efficiency, and placement stability of attention head configurations, enabling the model to more accurately handle complex refrigeration states such as cold energy propagation interruptions, local heat accumulation, and cargo-related mutations.

[0021] Preferably, the time-delay reconfiguration specifically includes: Response chain deconstruction is performed based on associated mutation data to obtain heterogeneous transfer chain data; Phase lag inversion is performed based on heterogeneous transmission chain data to obtain relational phase lag data; Based on the phase lag data and heterogeneous transmission chain data, head misalignment mapping is performed to obtain head time offset data; Based on the head temporal offset data, segmented start-stop reconstruction is performed to obtain the second attention head configuration data.

[0022] This invention deconstructs the response chain of correlated mutation data, breaking down physical effects at different time scales, such as airflow propagation, cargo surface heat transfer, cargo internal heat diffusion, and local heat accumulation, into heterogeneous transmission chains. This avoids the aliasing of multiple heat flow relationships caused by uniform time-series processing. Through phase lag inversion, the initial response, peak response, and decay range of each transmission chain relative to the switching moment of the refrigeration stage can be determined. Combined with inter-head misalignment mapping, the lag characteristics of different physical relationships are converted into the activation and deactivation offsets of the corresponding attention heads. Through segmented start-stop reconstruction, each attention head can sequentially enter delayed activation, parallel transition, stabilization, and gradual deactivation states according to the actual refrigeration process. This reduces stage mismatch caused by synchronous switching, improves the model's temporal adaptability to cold air inertia, cargo thermal inertia, and defrosting heat propagation processes, and enhances the stability and physical consistency of the generated placement scheme.

[0023] Preferably, this application also provides a goods placement optimization system for simulating the refrigeration process of a cold storage facility, used to execute the goods placement optimization method for simulating the refrigeration process of a cold storage facility as described above. The goods placement optimization system for simulating the refrigeration process of a cold storage facility includes: The 3D mesh mapping module is used to acquire 3D model data of cold storage, refrigeration pipeline data, and basic cargo data; it divides the cold storage into meshes based on the 3D model data and refrigeration pipeline data to obtain mixed mesh data of cold storage; and it performs spatial mapping on the mixed mesh data of cold storage based on the basic cargo data to obtain cargo spatial mapping data. The multi-domain model building module is used to divide the area into regions based on the cold storage hybrid grid data and cargo space mapping data to obtain multi-domain calculation data; and to build a model based on the multi-domain calculation data to obtain a refrigeration physical model. The refrigeration coupled solution module is used to acquire equipment operation data and environmental monitoring data; assign boundary values ​​to the refrigeration physical model based on the equipment operation data and environmental monitoring data to obtain initial model data; and perform coupled solution based on the initial model data to obtain refrigeration simulation data. The refrigeration state identification module is used to extract field quantities from refrigeration simulation data to obtain refrigeration spatiotemporal field data; and to identify the state based on the refrigeration spatiotemporal field data to obtain refrigeration state feature data. The cargo location association construction module is used to construct associations based on refrigeration status characteristic data, cargo space mapping data, and cargo basic data to obtain cargo location association data. The placement optimization evaluation module is used to learn multi-head self-attention relationships based on cargo location association data to obtain placement scheme data; to perform back-substitution simulation based on placement scheme data and refrigeration physical model to obtain placement simulation data; and to evaluate the scheme based on placement simulation data to obtain energy consumption prediction data.

[0024] The purpose of this invention is to: perform hybrid mesh generation based on 3D model data of cold storage and refrigeration pipeline data, and map basic cargo data to corresponding mesh regions, so that the internal fluid space, refrigeration pipelines, and cargo-occupied areas of the cold storage can be accurately represented under a unified spatial coordinate system. Through multi-domain partitioning and the construction of a refrigeration physics model, the coupling relationship between fluid flow, pipe wall heat conduction, cargo heat diffusion, and regional interface heat transfer is established, enabling the model to realistically reflect the cold air circulation and cargo cooling process. By using equipment operation data and environmental monitoring data for boundary assignment and coupled solution, refrigeration simulation results of temperature field, velocity field, pressure field, and cargo internal thermal state changes over time can be obtained. Through field quantity extraction and state identification, state characteristics such as insufficient cold coverage, airflow stagnation, local heat accumulation, and uneven temperature distribution can be identified. The refrigeration state, cargo spatial location, and cargo attributes are correlated and constructed, and complex refrigeration influence relationships between cargo and between cargo and cargo location are extracted through multi-head self-attention relationship learning, thereby generating a placement scheme adapted to the current refrigeration state. Substituting the placement scheme back into the refrigeration physical model for further simulation and evaluation allows for the verification of the rationality of the AI-generated scheme using physical calculation results. This improves the accuracy of energy consumption prediction, reduces localized cooling blind spots and ineffective refrigeration energy consumption, and creates a progressive closed loop in the goods placement optimization process. Attached Figure Description

[0025] Other features, objects, and advantages of this application will become more apparent from the following detailed description of the non-limiting embodiments, taken with reference to the accompanying drawings: Figure 1 A flowchart illustrating the steps of a method for optimizing the placement of goods in a cold storage refrigeration process simulation, according to one embodiment, is shown. Figure 2 A flowchart illustrating the steps of a region partitioning method according to an embodiment is shown; Figure 3 A flowchart illustrating the steps of a coupled solution method according to one embodiment is shown. Figure 4 A flowchart illustrating the steps of a state recognition method according to an embodiment is shown. Detailed Implementation

[0026] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0027] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. Functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0028] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0029] Using a low-temperature cold storage facility measuring 18m long, 12m wide, and 5m high as the model, the facility is equipped with 4 evaporators, 8 sets of refrigeration pipelines, and stores 320 pallets of frozen meat. The target storage temperature is -22℃ to -18℃. First, the 3D model of the cold storage, the refrigeration pipelines, and the dimensions, mass, location, and thermal properties of the goods were retrieved. A hexahedral-tetrahedral hybrid mesh was generated with a base step size of 0.10m, producing a total of 1.86 million mesh elements. Compared to the 3.25 million elements of a purely tetrahedral mesh, the number of mesh elements was reduced by 42.8%, and the simulation time was shortened from 9.6 hours to 5.8 hours.

[0030] The computational domain was divided into an air domain within the warehouse, a fluid domain within the pipes, a pipe wall domain, and a cargo domain. Flow, turbulence, and fluid-structure heat transfer models were established. Boundary values ​​were assigned based on operational data including a supply air temperature of -28℃, an average supply air velocity of 4.2 m / s, an external temperature of 30℃, and a warehouse door opening time of 90 seconds, and coupled solutions were completed. Under the original placement conditions, the average temperature of the cargo area was -18.6℃, the maximum temperature deviation was 3.4℃, the cooling coverage was 81.7%, and 11 areas with abnormal temperature gradients were identified.

[0031] Based on cargo size, temperature zone, load-bearing capacity, and isolation requirements, a cargo location association was constructed, and 10 candidate schemes were generated through multi-head self-attention learning. The optimal scheme splits dense stacks into 4 groups and increases the main ventilation spacing from 0.18m to 0.35m. Back-substitution simulation results show that after optimization, the maximum temperature deviation is reduced to 1.2℃, the cold energy coverage rate is increased to 95.3%, the abnormal areas are reduced to 3, and the time for cargo to enter the target temperature zone is shortened from 182min to 148min; the predicted energy consumption over 8 hours is reduced from 447.8kWh to 386.4kWh, a reduction of 13.7%.

[0032] Please see Figures 1 to 4 This application provides a method for optimizing the placement of goods in a cold storage refrigeration process simulation, comprising the following steps: S1. Obtain the 3D model data of the cold storage, refrigeration pipeline data, and basic cargo data; perform mesh generation based on the 3D model data and refrigeration pipeline data to obtain mixed mesh data of the cold storage; perform spatial mapping on the mixed mesh data of the cold storage based on the basic cargo data to obtain cargo spatial mapping data. Specifically, the cold storage BIM or CAD model is read, and the cold storage body, doors, walls, floor, ceiling, and equipment openings are uniformly converted to the cold storage coordinate system to form the cold storage 3D model data M; the j-th segment of the refrigeration pipeline is represented as... ,in, , These are the start and end coordinates of the pipe segment centerline. The outer radius is (or Let M be the wall thickness. Traverse all vertices of M, expanding outwards by the minimum and maximum values ​​of each coordinate axis. , such as 2 Construct a closed bounding volume B. Within B, increment by the base step size. Radiation with a square cross section is emitted along the x-axis and y-axis. =0.05m; When the center of a candidate unit is located inside the cold storage and the corresponding ray segment does not cross the boundary of the cold storage or refrigeration pipelines, the candidate unit is set as a hexahedral mesh. For boundary transition areas and void areas that cannot be covered by hexahedrons, their boundary nodes are extracted and Delaunay tetrahedrals are performed, and tetrahedrons are grown layer by layer until the voids are filled. For each mesh unit With pipeline entity Calculate the overlap rate ,when When necessary, the corresponding overlapping elements are deleted, constraint nodes are added to the outer surface of the pipeline, and Delaunay meshing is performed again to ensure that the boundaries of the new elements fit the pipe wall. Odd / even rays are used to detect the center of the mesh elements; if the number of intersections between the ray and the cold storage boundary is even, the element is determined to be outside the cold storage and deleted. For mesh clusters connected to the main region only through narrow cross-sections, the cross-sectional ratio is calculated. ,in For the cross-sectional area of ​​the connection, The minimum cross-sectional area of ​​adjacent main elements is determined by the projected areas of the main mesh elements on both sides of the connecting section in the plane perpendicular to the connecting direction, and the smaller of the two values ​​is taken. The secondary mesh cluster is then deleted to obtain mixed mesh data for the cold storage, which includes nodes, cells, faces, adjacency relationships, and pipe wall boundary markers.

[0033] Read the dimensions, shape, placement coordinates, orientation matrix, category, density, specific heat capacity, and thermal conductivity of the q-th item, and establish the item's physical region. Calculate the cargo occupancy rate for each grid cell. ,when At that time, Marked as cargo-occupying unit; when When, mark it as a cargo interface unit and refine it locally; when The data is maintained as air or other area units. Cargo spatial mapping data is generated according to cargo number, grid cell number, occupancy rate, spatial coordinates, physical property parameters, and stacking relationship.

[0034] S2. Based on the cold storage hybrid grid data and cargo space mapping data, the regions are divided to obtain multi-domain calculation data; based on the multi-domain calculation data, a model is constructed to obtain the refrigeration physical model; Specifically, read the set of cells in the hybrid grid data of the cold storage. Unit center coordinates Identify adjacent surfaces and pipeline boundaries, and read the cargo number and unit occupancy rate from the cargo space mapping data. Mesh attribute labeling is performed in the following order: cargo domain—pipe wall domain—fluid domain inside the pipe—air domain inside the storage area. At that time, the unit The solid unit is marked as the qth item; for units not occupied by goods, calculate the shortest distance from its center to the centerline of the jth segment of the refrigeration pipeline. ,when The time mark is the internal cooling medium unit, when The units marked with time stamps are designated as pipe wall units, while the remaining units located inside the cold storage are designated as internal air units. The outer radius of the pipeline. The pipe wall thickness is determined by the actual pipe specifications in the refrigeration pipeline data. Units with the same attributes and connected by a common surface are grouped into independent computational regions; if two adjacent units have different attributes, their common surface is marked as a coupling interface, and the interface area is recorded. Unit normal vector And the unit numbers on both sides. This forms the cooling medium fluid domain inside the pipe, the air fluid domain inside the storage, the solid domain of the pipe wall, the solid domain of the cargo, and the interface set between each domain, resulting in multi-domain calculation data.

[0035] Configure physical property parameters based on multi-domain calculation data. Configure density for the fluid domain. Dynamic viscosity Specific heat capacity at constant pressure and thermal conductivity Establish the continuity equation, momentum equation, and energy equation: , For fluid velocity vector, Let be the fluid velocity gradient tensor. For turbulent dynamic viscosity, For Hamiltonian operators, Configure the divergence operator and RNG. Turbulence equations. Density configuration for the pipe wall region and cargo region. Specific heat capacity and thermal conductivity Establish the heat conduction equation: In the fluid-structure interface settings and heat flow continuity conditions , This refers to the fluid-side interface temperature. This refers to the solid-side interface temperature. The effective thermal conductivity on the fluid side, The thermal conductivity is the solid side. For the fluid-side temperature gradient, For the solid-side temperature gradient, This represents the unit normal vector of the interface pointing from the fluid region to the solid region. By integrating the various regions, governing equations, physical properties, coupling interfaces, and boundary types such as inlet, outlet, and storage wall, a refrigeration physical model is obtained.

[0036] S3. Acquire equipment operation data and environmental monitoring data; assign boundary values ​​to the refrigeration physical model based on the equipment operation data and environmental monitoring data to obtain initial model data; perform coupled solution based on the initial model data to obtain refrigeration simulation data; Specifically, the system collects data on compressor start / stop status and power, evaporator fan speed, refrigerant inlet temperature, outlet pressure, pipe flow rate, and valve opening to generate equipment operation data; it also collects data on temperature, humidity, pressure, outside temperature, and door opening / closing status at various measuring points inside the storage facility to generate environmental monitoring data. All data are then aligned to a unified calculation time according to timestamps. .

[0037] Values ​​are assigned based on the device number and the boundary identifiers in the multi-domain calculation data. For the evaporator supply air boundary... Based on the measured volumetric flow rate and entrance area Determine the inlet velocity: The measured supply air temperature was assigned as a value. , Let n be the evaporator supply air inlet velocity vector at the nth calculation time. This is the unit direction vector at the evaporator supply air inlet, used to represent the direction of the supply airflow entering the cold storage computational domain; for return air or outlet boundaries... This gives the measured export pressure The fluid inlet of the refrigeration pipes uses measured flow rate, temperature, and pressure, while the outlet uses a pressure boundary; the outer wall of the cold storage uses a third-class heat exchange boundary. ,in, The thermal conductivity of the storage wall is... The external convective heat transfer coefficient is... Let be the unit normal vector from the outer wall of the cold storage to the external environment. This refers to the wall surface temperature at the outer boundary of the cold storage. This refers to the external temperature. When the warehouse door is closed, it is treated as a warehouse wall boundary; when open, the door area is set as a time-varying opening boundary that is connected to the external temperature and pressure.

[0038] Initial velocity is generated by performing partitioned interpolation using the positions of each measuring point within the computational domain. Initial pressure Initial temperature Initial turbulent kinetic energy and initial dissipation rate The initial model data is formed together with boundary conditions, physical properties, and mesh data. The finite volume method is used to discretize the governing equations. First, the momentum equation is solved to obtain the predicted velocity. Then, the pressure and velocity are corrected using the pressure correction equation. The corrected flow field is then used to update the turbulent kinetic energy k and dissipation rate. And turbulent viscosity, and then solve the energy equations for the fluid inside the pipe, the pipe wall, the air inside the storage tank, and the cargo region in sequence, performing temperature continuity and heat flux conservation at the fluid-solid interface. Repeat the above process until velocity, pressure, k, The residuals are all less than 10 -4 And the energy residual is less than 10 -6The system outputs the velocity field, pressure field, temperature field, turbulence field, and cargo temperature changes at each moment, thus obtaining refrigeration simulation data.

[0039] S4. Extract field quantities from the refrigeration simulation data to obtain refrigeration spatiotemporal field data; perform state identification based on the refrigeration spatiotemporal field data to obtain refrigeration state characteristic data. Specifically, read the calculation time points from the refrigeration simulation data. and each grid unit Corresponding temperature ,pressure Velocity vector Turbulent kinetic energy Turbulent dissipation rate and synchronously read the cell volume Center coordinates Public area Surface normal vector and adjacent unit set By encapsulating the field quantity sequence according to the time step number and unit number, the spatiotemporal field data of cooling is obtained. For temperature, pressure, and velocity moduli Finite volume gradient calculation is used: ,in, The field quantity value at the common surface f at the nth calculation time can represent temperature T, pressure p, or velocity magnitude v, and is obtained by linear interpolation of the element values ​​on both sides of the common surface; , and Encapsulated as spatiotemporal field gradient data. Let be the temperature gradient magnitude of the i-th grid cell at the n-th calculation time. Let be the pressure gradient magnitude of the i-th grid cell at the n-th calculation time. Let be the velocity magnitude gradient magnitude of the i-th grid cell at the nth calculation time. Calculate the field fluctuations within a sliding window of length L time steps: We obtained the time series fluctuation data of the field quantity. Let represent the amplitude of the field fluctuation in the i-th grid cell at the nth calculation time within the sliding time window. Let be the field quantity value of the i-th grid cell at the s-th calculation time. Number the grid cells. Number the current calculation time. This is the index of the calculation time within the sliding time window. The number of time steps contained in the sliding time window.

[0040] According to the area identifiers in the refrigeration simulation data, the same area The cells within the area are aggregated to calculate the average temperature of the region. and maximum temperature difference , Number the grid cells. Let r be the set of grid cells for the r-th computational region. Let i be the volume of the i-th grid cell. Let be the temperature of the i-th grid cell at the nth calculation time. And temperature fluctuations within the region At that time, it was determined that the temperature in the area was uniform, forming uniform field characteristic data. The threshold for the maximum temperature deviation in the region, such as 1. This is the threshold for the temperature fluctuation amplitude of the grid cell, such as 0.5.

[0041] According to the preset cooling temperature range Determine the effective cooling capacity units and calculate the cooling capacity coverage: The preset cooling temperature range is determined based on the target storage temperature range in the cargo's basic data. Coverage, coordinates of uncovered units, and duration are encapsulated as cooling capacity coverage feature data. , or When M=3 time steps are reached consecutively, they are marked as temperature abrupt change, cooling fluctuation, or insufficient cooling coverage, respectively, to obtain cooling anomaly characteristic data. The threshold for determining temperature gradient anomalies is set, such as 1.5. A threshold for determining cooling capacity coverage is set, such as 0.9. Cooling status characteristic data is obtained by aggregating data according to region number, time interval, average temperature, maximum temperature difference, cooling capacity coverage, anomaly type, and anomaly location.

[0042] S5. Based on the refrigeration status characteristic data, cargo space mapping data and cargo basic data, the cargo location association data is constructed. Specifically, read the set of occupying units corresponding to the q-th item in the cargo space mapping data. Spatial centroid Outer packaging dimensions The system retrieves the cargo's position and current location number, and extracts the cargo category and quality from the basic cargo data. Permissible stacking layers, target storage temperature range And category isolation requirements. The cold storage space is divided into candidate storage locations according to shelf levels, aisles, and continuous available grids. Record the center coordinates of each storage location. Available sizes Maximum carrying capacity Occupancy status and the set of grid cells contained therein .

[0043] The refrigeration status characteristic data is mapped to each candidate storage location according to the grid number. For each storage location... Calculate the average temperature of the storage location: And extract the maximum temperature difference corresponding to that storage location. Cooling coverage Maximum temperature gradient Temperature fluctuation range and abnormal types When a storage location covers multiple state regions, the state of each region is retained based on the actual grid assignment result. An adaptation decision metric is established for cargo q and candidate storage location r: ,in, The value is 1 if the condition is met, and 0 otherwise. This is a sign indicating where the storage space is occupied. No abnormalities were found. The r-th candidate storage location is currently vacant. For different categories of goods that need to be isolated, if the shortest distance between the candidate storage location and the prohibited adjacent category of goods is less than the isolation distance... For example, 0.5m, then Set to 0; for goods that are allowed to be stacked, it is also necessary to verify the remaining load-bearing capacity of the lower layer of goods and the current stacking layer.

[0044] Only for Establish an association edge between the goods and storage locations, and record the goods number, candidate storage location number, and spatial distance. The system includes data on average temperature of the storage location, cold air coverage, temperature gradient, fluctuation range, abnormal conditions, and stacking relationships. A cargo placement sequence is generated according to the aisle, layer, and column order of the storage location, and this sequence is used as a component of the cargo location association data.

[0045] S6. Based on the cargo location association data, perform multi-head self-attention relationship learning to obtain placement scheme data; based on the placement scheme data and the refrigeration physical model, perform back-substitution simulation to obtain placement simulation data; based on the placement simulation data, evaluate the scheme to obtain energy consumption prediction data.

[0046] Specifically, the system reads the cargo location association data, including cargo number, available candidate locations, placement sequence, cargo size, category, stacking relationship, average location temperature, cold storage coverage, temperature gradient, fluctuation range, and anomaly flags. The relationship between the q-th cargo and the r-th candidate location is then represented as a relational feature. And only retain the aforementioned adaptation determination quantity. =1 combination. Based on the temperature change rate, cold coverage rate and anomaly indicator carried by the relationship characteristics, the continuous simulation period is divided into cooling stage, stable stage and anomaly stage; adjacent time stages with the same indicator are grouped into the same stage, resulting in the stage placement sequence.

[0047] Construct a query matrix, a key matrix, and a value matrix (query projection matrix) for the stage placement sequence of stage s. Key projection matrix Value projection matrix Initialization is performed during the model training phase, and updates are made through backpropagation based on historical placement schemes and their back-substitution simulation evaluation results. , , , ,in Let be the query matrix for the h-th attention head in the s-th stage, used to represent the placement relationships that each cargo node currently needs to match. The corresponding key matrix represents the location and constraint features of other cargo nodes that can be compared. Let h be the value matrix of the attention head in the s-th cooling stage. Assign matrices to the corresponding attention levels. Let h be the feature matrix of the stage placement sequence nodes for the s-th cooling stage, and h be the attention head number. For single-head dimension; when the storage space is occupied, the size is incompatible, the load capacity is insufficient, the temperature zone is incompatible, or the isolation and stacking requirements are violated, the mask matrix... The corresponding position is set to negative infinity, so that the relationship is not involved in learning. The attention matrices of adjacent stages are compared, and the calculation is performed. The scope of attention heads whose mutations are concentrated in local locations is limited, and the key value sequence is moved according to the response time difference for attention heads with temperature response lag, thereby completing the head reconfiguration.

[0048] After head rematching is completed, calculate the output of the h-th attention head: The outputs of each attention point are then concatenated and output mapping is performed: ,in, To output the projection matrix. For the q-th item and its r-th valid candidate storage location, Output of the node corresponding to the qth item Characteristics of the cargo-location relationship Combine the data and calculate the scores for the candidate storage locations: ,in, and The output layer parameters are the storage location parameters. The scores of all valid candidate storage locations for the q-th item are normalized: ,in, Let q be the set of valid candidate storage locations for the q-th item. The probability of selecting the r-th candidate storage location for the q-th item. According to... Generate a sorting result of storage locations from largest to smallest, and select candidate storage locations that are not yet occupied and do not have spatial conflicts in turn.

[0049] The outputs of each attention point are concatenated to generate a cargo location ranking result, in order of cargo selection probability and absence of spatial conflict. After each allocation, the cargo location occupancy status and remaining carrying capacity are updated until all cargo is allocated, forming placement scheme data containing cargo number, target cargo location, placement orientation, and stacking order. Each candidate scheme is rewritten into the refrigeration physics model, updating the cargo solid domain, fluid channels, and fluid-solid interface. Using the same boundary conditions and coupling solution rules, the temperature field, velocity field, and cargo temperature changes are recalculated within the same simulation duration to obtain placement simulation data. When any cargo temperature exceeds its allowable upper limit, the equipment operation flag is activated. =1, otherwise 0, and calculate the predicted cooling energy consumption corresponding to the placement scheme: ,in, The power of the refrigeration equipment in the aforementioned equipment operation data. This represents the calculation time interval between two adjacent calculation time steps. First, schemes with excessive temperature range, insufficient coverage, or persistent abnormalities are deleted. Then, they are sorted by E from smallest to largest to obtain the predicted energy consumption, runtime, and optimal order for each placement scheme, forming the energy consumption prediction data.

[0050] Preferably, the region division is as follows: S21. Based on the cold storage mixed grid data and cargo space mapping data, perform grid marking to obtain grid attribute marking data; Specifically, reading the grid cells from the cold storage hybrid grid data. Unit center coordinates Unit volume The system retrieves the adjacency relationships of public surfaces and the boundary markers of refrigeration pipelines, and reads the cargo number and occupancy rate from the cargo space mapping data. ,in, Let q be the volume percentage of the q-th item in the i-th grid cell. For each grid cell, mark the items, pipelines, and warehouse space: when there are... At that time, record the corresponding cargo number and ; Calculate the center coordinates of the unit The shortest distance to the centerline of the j-th segment of the refrigeration pipeline And read the outer radius of the j-th segment of the refrigeration pipeline. and pipe wall thickness .when Set candidate markers for the area within the pipe when Candidate markers are set for the pipe wall area, and candidate markers for the remaining grid cells inside the cold storage are set for the internal air. The grid cell number, candidate marker, and cargo number are then used. , In addition to encapsulating the original boundary markers, we obtain the grid attribute marker data.

[0051] S22. Perform domain occupancy determination on the grid attribute marker data to obtain domain occupancy determination data; Specifically, the unique domain category of each grid cell is determined in the following order: cargo domain first, pipe wall domain second, pipe fluid domain third, and warehouse air domain last. When it exists season ,in, This represents the cargo domain corresponding to the q-th item; if the same grid cell corresponds to multiple items, then select... The largest item is determined by ascending item number when values ​​are the same. For grid cells not occupied by items, when... season ,in, This represents the pipe wall region corresponding to the j-th segment of the refrigeration pipeline; when season ,in, Let j-th segment of the refrigeration pipeline be the fluid domain within the pipeline; the remaining mesh elements are defined as follows: ,in, This represents the air domain within the storage area. The grid cell number and domain category are specified. The corresponding cargo number or refrigeration pipeline number is packaged to obtain the domain occupancy determination data.

[0052] S23. Perform fluid-structure partitioning based on the domain determination data to obtain fluid-structure partitioning data; Specifically, according to the category of territory occupied The mesh is divided into sets of fluid elements and sets of solid elements. This will satisfy... or The mesh elements are incorporated into the fluid element set, which will satisfy... or The grid cells are categorized into the solid cell set; at the same time, the category distinction between the air domain in the warehouse and the fluid domain in each pipe, as well as the object distinction between each pipe wall domain and each cargo domain, are preserved. The adjacency relationship of the common surface is traversed, and the adjacent grid cell numbers, the domain categories on both sides, and the common surface number are recorded, so that both the fluid cell set and the solid cell set have complete internal connectivity and cross-domain adjacency relationships, thus obtaining the fluid-solid partitioning data.

[0053] S24. Perform interface identification based on fluid-structure partitioning data to obtain interface identification data; Specifically, we iterate through each common surface f in the fluid-structure partitioning data. Let the mesh cells on both sides of the common surface f be... and ,when At that time, the common surface f is defined as the region interface, and the area of ​​the common surface is read. and by point to unit normal vector When both sides belong to a fluid element set and a solid element set respectively, it is marked as a fluid-solid heat transfer interface; when both sides belong to a solid element set and have different domain categories, it is marked as a solid-solid contact interface. Further interfaces are formed according to domain category: in-pipe fluid-pipe wall interface, pipe wall-in-storage air interface, cargo-in-storage air interface, and cargo-pipe wall interface. For exposed surfaces adjacent to only one grid element, the boundary markers are used to determine the storage wall, storage door, inlet, or outlet boundary. The common surface number, the grid element numbers on both sides, the interface type, and... and Encapsulation yields interface recognition data.

[0054] S25. Based on the fluid-structure partitioning data and interface recognition data, the regions are aggregated to obtain multi-domain calculation data.

[0055] Specifically, a connectivity graph is constructed using grid cells with the same domain category and interconnected through common surfaces. Each connected component is then aggregated into an independent computational region. Air cells within the warehouse are aggregated into an air fluid domain; fluid cells within pipes and pipe wall cells belonging to the same refrigeration pipeline number are aggregated into an fluid domain and a solid domain, respectively; cargo cells belonging to the same cargo number are aggregated into the corresponding cargo solid domain. For each computational region, the region number, domain category, grid cell set, total region volume, internal common surface set, and outer boundary set are recorded, and the interface correspondence between computational regions is established based on interface identification data. Finally, the fluid domain within pipes, air fluid domain within the warehouse, solid domain within pipes, cargo solid domain, region connectivity relationships, interface geometric parameters, and boundary markers are uniformly encapsulated to obtain multi-domain computational data.

[0056] Preferably, the model construction is specifically as follows: S26. Configure attributes based on multi-domain calculation data to obtain multi-domain physical property data; Specifically, the computational regions in the aforementioned multi-domain computational data are read. Regional types Grid cell set Interface set and boundary markers, among which, This includes airflow within the storage area, fluids within pipes, solids on pipe walls, and solid cargo. Corresponding material parameters are matched based on the media type, pipeline number, or cargo number retained within the calculation area. For fluid areas, the area density is configured. Dynamic viscosity Specific heat capacity at constant pressure thermal conductivity and Prandtl number of turbulence For solid regions, configure the regional density. Specific heat capacity at constant pressure and thermal conductivity Simultaneously configure the turbulence model constants, including the dissipation rate generation term model constants. Dissipation rate decay term model constant turbulent viscosity model constants Prandtl number of turbulent kinetic energy diffusion and turbulent kinetic energy dissipation rate diffusion Prandtl number Each parameter is then bound to a calculation region number to obtain multi-domain physical property data.

[0057] S27. Perform fluid modeling and solid modeling on multi-domain physical property data and multi-domain computational data to obtain fluid equation data and heat transfer equation data. Specifically, continuity and momentum equations are established for the air fluid region inside the reservoir and the fluid region inside the pipe: , ,in, For the region velocity vector, For regional pressure, For turbulent viscosity, This is the gravitational acceleration vector. According to RNG... The model establishes the turbulent kinetic energy equation and the dissipation rate equation: , ,in, For turbulent kinetic energy, For dissipation rate, For turbulent kinetic energy generation, This is the RNG correction term, from which the fluid equation data is obtained. An energy equation is then established for the fluid region: Establish heat conduction equations for the solid regions of the pipe wall and the solid regions of the cargo: ,in, The temperature is the region. The equation type, the region of effect, and the variables to be solved are stored accordingly to obtain the heat transfer equation data.

[0058] S28. Based on the fluid equation data, heat transfer equation data, and multi-domain calculation data, perform interface coupling to obtain coupled equation data. Specifically, traversing the region interfaces in multi-domain computational data. .when When the interface is a fluid-solid heat transfer interface, a no-slip condition is set on the fluid side. =0, and set the temperature continuity condition. and the heat flow conservation condition: ,in, For the region Pointing to area The unit normal vector, Let be the temperature of the m-th computational region at the region interface. Let be the temperature at the interface of the nth computational region. It is a solid-solid contact interface with contact thermal resistance. At that time, set the interface heat flow And make the heat flow on both sides equal to When there is no contact thermal resistance, the conditions of temperature continuity and heat flux conservation are adopted. The interface number, the numbers of the two side regions, and the corresponding coupling conditions are encapsulated to obtain the coupling equation data.

[0059] S29. Based on the coupled equation data and multi-domain calculation data, perform boundary configuration to obtain boundary condition data; Specifically, boundary types are set based on the inlet, outlet, wall, and door boundary identifiers in the multi-domain calculation data. The inlet boundary reserves inlet velocity. Inlet temperature Inlet turbulent kinetic energy and inlet dissipation rate Export border reserves allow for export pressure. Zero normal gradients were set for velocity, temperature, turbulent kinetic energy, and dissipation rate. No-slip conditions were set for the reservoir wall boundaries, and the following configurations were made: ,in, The external convective heat transfer coefficient is... The external temperature is used. The warehouse door boundary is configured with the warehouse wall boundary type in the closed state and the pressure opening boundary type in the open state. The above time-varying parameters only establish an assignment interface; their specific values ​​are assigned by subsequent equipment operation data and environmental monitoring data to obtain the boundary condition data.

[0060] S2a. Based on the coupled equation data and boundary condition data, the model is integrated to obtain the refrigeration physical model.

[0061] Specifically, multi-domain property data, fluid equation data, and heat transfer equation data are bound to corresponding mesh cells according to the computational domain number; coupling equation data is bound to the corresponding domain interface; and boundary condition data is bound to the corresponding outer boundary surface. It is checked that each mesh cell belongs to only one computational domain, each domain interface has two side regions, and each outer boundary surface has a unique boundary type. The mesh topology, domain properties, governing equations, variables to be solved, interface coupling relationships, and boundary assignment interfaces are uniformly encapsulated to obtain the refrigeration physical model.

[0062] Preferably, the coupled solution is specifically as follows: S31. Discretize the equations based on the initial model data to obtain the discrete model data; Specifically, read the grid cells from the initial model data. Unit volume Public area Unit normal vector Regional physical parameters, governing equations, interface coupling conditions, boundary values, and initial velocities. Initial pressure Initial temperature Initial turbulent kinetic energy and initial dissipation rate The finite volume method is used to integrate the governing equations. The time term is discretized using a first-order implicit scheme, the convection term using a second-order upwind scheme, and the diffusion term is finitely differentiald based on the center-to-center distance between the grid cells on both sides of the common surface. The unknown field quantities are then... ( For speed, For pressure, For turbulent kinetic energy, For dissipation rate, For temperature) uniformly discretized as: ,in, For grid cells The set of adjacent units, The coefficients of dispersion are centered. The adjacency dispersion coefficient is . This is a discrete source term that includes the field quantities, boundary conditions, and source terms from the previous time step. The discrete coefficients of each field quantity, the discrete source term, the common surface flux relationship, and the interface discrete relationship are encapsulated to obtain the model's discrete data.

[0063] S32. Momentum prediction is performed based on the discrete data of the model to obtain the predicted flow field data; Specifically, in the r-th iteration, the current pressure is used. Turbulent viscosity Construct the momentum discrete equation based on the velocity at the previous moment: ,in Let be the central discretization coefficient of the momentum equation for the i-th grid cell. To predict speed, Let be the adjacency discretization coefficient corresponding to the l-th adjacent grid cell in the momentum equation. Let i be the discrete source term of the momentum equation for the i-th grid cell. Let be the pressure gradient of the i-th mesh element in the r-th iteration. Solve for the predicted velocity of each fluid mesh element, and interpolate the predicted velocities on both sides of the common surface to obtain the predicted velocity at the common surface. Then, the predicted mass flow rate of the common surface is calculated: ,in, The fluid density at the common surface is used. The current pressure, predicted velocity, and predicted mass flow rate at the common surface are correlated and stored to obtain the predicted flow field data.

[0064] S33. Perform pressure velocity correction on the predicted flow field data to obtain the corrected flow field data; Specifically, based on the mass conservation condition of each grid cell A pressure correction equation is constructed using the momentum discrepancy coefficient. Let i be the set of boundary surfaces of the i-th mesh cell. Let be the corrected mass flow rate at the f-th common surface. Let the pressure correction be... The speed correction relationship is as follows: , , The velocity-pressure correction factor for the i-th grid cell is... For spatial gradient operators, Let be the center discrete coefficient of the momentum equation for the i-th grid element. Substituting this relationship into the expression for the mass flow rate at the common surface, we get: ,in The central discrete coefficients of the pressure correction equation are... The pressure correction is determined by the sum of the predicted mass flow rates of all common faces of the grid cells. After solving for the pressure correction, it is calculated according to... The pressure is corrected, and the velocity and common surface mass flow rate are simultaneously corrected to ensure that each fluid mesh element satisfies mass conservation. The corrected pressure, corrected velocity, and corrected mass flow rate are then encapsulated to obtain the corrected flow field data.

[0065] S34. Perform turbulence update on the corrected flow field data to obtain turbulent flow field data; Specifically, the average strain rate tensor is calculated based on the correction rate: And calculate the turbulent kinetic energy generation term. Substituting the correction velocity and turbulent kinetic energy generation terms into the turbulent kinetic energy equation and dissipation rate equation in the discrete data of the model, the following results are obtained: and To avoid non-physical interpretation, when the calculated result is less than the corresponding positive lower limit (a preset value, taking values ​​of 10 to the power of -8), it is corrected to that positive lower limit. Update turbulent viscosity: ,in, The constants for the aforementioned turbulence model are set to, for example, 0.0845. The corrected pressure, corrected velocity, turbulent kinetic energy, dissipation rate, and turbulent viscosity are then encapsulated to obtain the turbulent flow field data.

[0066] S35. Perform heat transfer coupling on the turbulent flow field data and the discrete model data to obtain fluid-structure interaction field data; Specifically, for the air fluid domain inside the reservoir and the fluid domain inside the pipe, the convection term in the energy equation is updated based on the corrected mass flow rate, and... As the effective thermal conductivity of a fluid, Let be the thermal conductivity of the medium belonging to the i-th grid cell. Let i be the isobaric specific heat capacity of the i-th fluid grid cell. The Prandtl number represents the turbulent flow; the corresponding thermal conductivity is used for the solid domain of the pipe wall and the solid domain of the cargo. Solve the heat conduction equation. For the common surface f of the fluid-structure interface, based on the distance from the center of the mesh elements on both sides of the interface to the common surface... and Calculate the interface heat flux: ,in Let i be the temperature of the i-th fluid-side grid cell. The temperature of the l-th solid-side grid cell is... Let be the normal distance from the center of the i-th mesh cell to the f-th fluid-structure interface. Let be the effective thermal conductivity of the i-th fluid grid element. Let be the normal distance from the center of the l-th grid cell to the f-th fluid-structure interface. Let be the thermal conductivity of the l-th solid-side grid element. For interface contact thermal resistance; when there is no contact thermal resistance, it is considered as follows: =0. (The rest of the text appears to be incomplete and requires further context.) Discrete source terms are written into the grid cells on both sides in the form of equal values ​​but opposite directions to ensure that the fluid-solid interface satisfies heat flux conservation. The temperatures of the fluid region, pipe wall region, and cargo region are solved sequentially to obtain the temperature of each grid cell and the heat flux of each interface, which are then combined with the turbulent flow field data to form fluid-solid coupled field data.

[0067] S36. Iterative convergence is performed based on the fluid-structure interaction field data to obtain refrigeration simulation data.

[0068] Specifically, normalized discrete residuals are calculated for velocity, pressure, turbulent kinetic energy, dissipation rate, and temperature respectively: ,in, To prevent positive numbers with a denominator of zero. If the velocity residual... Pressure residual Turbulent kinetic energy residual and dissipation rate residual All less than 10 -4 And the energy residual corresponding to temperature Less than 10 -6 If convergence is achieved at the current calculation time, the calculation is considered complete. Otherwise, the pressure, velocity, turbulent kinetic energy, dissipation rate, turbulent viscosity, and temperature from the fluid-structure interaction field data are used as the values ​​for the next iteration, and the process returns to S32 momentum prediction / continues solving. After convergence at the current calculation time, the results are used as the initial field quantities for the next calculation time and the process continues. The velocity field, pressure field, temperature field, turbulent kinetic energy field, dissipation rate field, and interface heat flux at each calculation time are collected to obtain the refrigeration simulation data.

[0069] Preferably, the state recognition specifically includes: S41. Perform gradient calculation based on the refrigeration spatiotemporal field data to obtain the spatiotemporal field gradient data; Specifically, read the calculation times from the aforementioned cooling spatiotemporal field data. Each grid cell Corresponding temperature ,pressure Velocity vector Unit volume Public area Common surface unit normal vector and adjacent unit set And calculate the velocity modulus. The field quantities to be analyzed are uniformly denoted as... ,in, The field gradient is calculated for each grid cell using a finite volume scheme: ,in, The field value at the common surface f is used. When mesh elements exist on both sides of the common surface, the linear interpolation result of the field values ​​on both sides is taken. When the common surface is an outer boundary, the field value given by the corresponding boundary conditions in the refrigeration physics model is taken. The temperature gradient magnitude is calculated separately. Pressure gradient magnitude and velocity gradient magnitude The data is then stored in conjunction with the grid cell number, spatial coordinates, and calculation time to obtain the spatiotemporal field gradient data.

[0070] S42. Extract fluctuations from the spatiotemporal field data of the cooling system and the spatiotemporal field gradient data to obtain the temporal fluctuation data of the field quantity. Specifically, at the current calculation time Using the endpoint of the window as the starting point, select L consecutive calculation times (e.g., 3) to form a sliding time window. For each grid cell and each type of field quantity, calculate the field quantity fluctuation amplitude: And calculate the gradient fluctuation amplitude based on the spatiotemporal field gradient data: ,in, Let be the amplitude of the field fluctuation in the i-th grid cell within the current sliding time window. This represents the corresponding gradient fluctuation amplitude. When the number of available computation time points is less than L, a sliding time window is constructed using all currently available computation time points. The field quantity fluctuation amplitudes, gradient fluctuation amplitudes, window start and end times, and grid cell numbers corresponding to temperature, pressure, and velocity are encapsulated to obtain the field quantity time-series fluctuation data.

[0071] S43. Based on the spatiotemporal field gradient data and the temporal fluctuation data of the field quantity, a uniform evaluation is performed to obtain the uniform characteristic data of the field quantity. Specifically, based on the area numbers retained in the refrigeration spatiotemporal field data, areas belonging to the same storage evaluation area are classified. The grid cells are aggregated. The maximum gradient magnitude, maximum field fluctuation amplitude, and maximum gradient fluctuation amplitude of each field quantity are extracted separately. , , ,when , and At that time, the uniformity of the field quantity is marked. Set to 1 otherwise set to 0. , and These represent the upper limits for the gradient, fluctuation, and gradient fluctuation of the corresponding field quantities. For example, the upper limits for temperature are 1.5, 0.5, and 0.5, respectively; for pressure, they are 20, 30, and 10; and for velocity, they are 0.8, 0.3, and 0.3. The maximum gradient magnitude, maximum field quantity fluctuation amplitude, maximum gradient fluctuation amplitude, uniformity indicator, and out-of-limit grid cell set for each region are encapsulated to obtain the field quantity uniformity characteristic data.

[0072] S44. Determine the cooling capacity based on the refrigeration time-space field data and the field quantity uniformity characteristic data to obtain the cooling capacity coverage characteristic data. Specifically, for each storage evaluation area Calculate the average temperature of the region: ,in, This refers to the number of grid cells within the storage evaluation area. Based on the preset cooling temperature range. , will satisfy The grid cells are determined as effective cold coverage cells, and the cold coverage rate is calculated: When the temperature is uniformly marked and When this occurs, the area is determined to be a stable cooling capacity coverage area; when but When, it is determined to be a non-uniform cold energy coverage area; when At that time, it was determined to be an area with insufficient cooling coverage. The value is 0.9. The average temperature of the recorded area, the cold coverage rate, the cold determination result, and the set of grid cells not effectively covered by cold are used to obtain cold coverage characteristic data.

[0073] S45. Based on the spatiotemporal field gradient data, field temporal fluctuation data, and cold load coverage characteristic data, anomaly identification is performed to obtain refrigeration anomaly characteristic data. Specifically, anomaly rule determination is performed for each grid cell and reservoir evaluation area. When When, the corresponding grid cell is marked as a field gradient anomaly; when or When, the corresponding grid cell is marked as having an abnormal field fluctuation; when At that time, the storage evaluation area will be... The anomaly is marked as a cooling capacity coverage anomaly. A valid anomaly is confirmed only when the same anomaly condition is met consecutively for M calculation times, thus eliminating instantaneous numerical disturbances caused by a single calculation time. Adjacent mesh cells with the same anomaly type are connected and aggregated according to their common surface adjacency relationship to form an anomaly region. The anomaly type, start and end times, duration, set of anomaly mesh cells, maximum gradient magnitude, and maximum field quantity fluctuation amplitude are recorded to obtain cooling anomaly characteristic data.

[0074] S46. Based on the field uniformity characteristic data, cold energy coverage characteristic data, and refrigeration anomaly characteristic data, feature aggregation is performed to obtain refrigeration state characteristic data.

[0075] Specifically, using the storage evaluation area number and calculation time as a unified index, the various field quantity uniformity identifiers, regional maximum gradient modulus, and regional maximum field quantity fluctuation amplitude in the field quantity uniformity characteristic data are correlated with the regional average temperature, cold quantity coverage rate, cold quantity judgment result, and uncovered grid cell set in the cold quantity coverage characteristic data. Then, the anomaly type, anomaly location, anomaly duration, and anomaly start and end time in the refrigeration anomaly characteristic data are written into the same index. If multiple anomaly types exist simultaneously in the same area, each anomaly record is retained separately without merging. Finally, the refrigeration state characteristic data is uniformly encapsulated according to the area number, calculation time, regional average temperature, temperature uniformity identifier, pressure uniformity identifier, velocity uniformity identifier, regional maximum temperature gradient, regional maximum temperature fluctuation amplitude, cold quantity coverage rate, cold quantity judgment result, and anomaly record.

[0076] Preferably, the multi-head self-attention relationship learning specifically includes: S61. Based on the refrigeration status characteristic data, perform stage identification to obtain refrigeration stage data; Specifically, the associated refrigeration status characteristic data is retrieved from the aforementioned cargo location association data, and the data for each calculation time point is read. Lower storage evaluation area Corresponding regional average temperature Temperature uniformity indicator Cooling coverage 1. Cold load determination results and anomaly records. 2. Calculate the temperature change between adjacent calculation times: When any reservoir evaluation area has a valid anomaly record, or When the current calculation time is identified as an abnormal phase, and no valid abnormal record exists, the current calculation time is identified as an abnormal phase; when no valid abnormal record exists, and there is an abnormal phase, the current calculation time is identified as an abnormal phase. , When the temperature is low, it is identified as the cooling stage; when all storage evaluation areas meet the requirements... , and When the condition is met, it is identified as a stable phase; the rest satisfy the following conditions. The calculation time is identified as the warming phase. Among them, , Using the aforementioned parameters for determining cooling capacity, A preset threshold for stage changes, such as 0.1, is used. Calculation times of consecutive stages with the same type are merged into a stage interval, and the stage number, stage type, start and end times, and corresponding storage evaluation area are recorded to obtain refrigeration stage data.

[0077] S62. Perform stage segmentation on the cargo location association data and refrigeration stage data to obtain stage placement sequence data; Specifically, basic cargo data corresponding to the cargo number is extracted from the cargo location association data, including the volume, outer packaging dimensions, category, shape, and target storage temperature range of the q-th cargo. The system includes the allowed stacking layers, isolation requirements, current storage location number, and candidate storage location numbers. Using the stage intervals in the refrigeration stage data as segment boundaries, the storage location status, target storage temperature requirements, and candidate storage location status of each item within the corresponding stage are written into the stage item node. For each stage, the system first calculates the target storage temperature limit. Sort by low to high; If the quantities are the same, sort the goods by volume from largest to smallest; if they are still the same, sort them by item number from smallest to largest. The sorting result is represented as: ,in, This is the placement sequence for stage s. Each cargo node contains basic cargo fields, current cargo location number, candidate cargo location number, and stage type, with Q representing the cargo quantity. Each stage forms a corresponding placement sequence, resulting in stage placement sequence data.

[0078] S63. Perform correlation modeling based on the stage placement sequence data to obtain stage cargo correlation data; Specifically, the cargo nodes in the stage placement sequence are used as graph nodes, and for any two cargo nodes within the same stage... and Perform relationship determination. If the shortest distance between the current or candidate storage locations of two locations is not greater than the preset adjacency distance... At that time, establish spatial adjacency relationships , =0.3m; when the target storage temperature ranges of the two overlap, a temperature compatibility relationship is established. When the qth item can be stacked on top of the lth item, and the stacking layer number and load-bearing conditions are both met, a stacking relationship is established. When both categories belong to an isolation combination and the distance is less than the corresponding isolation distance, such as 0.5m, an isolation conflict relationship is established. Represent the relation type as: Each relation component takes the value 1 when the corresponding condition is met, and 0 otherwise. The stage number, cargo node, relation type, associated cargo number, and corresponding storage location number are encapsulated to obtain the stage cargo association data.

[0079] S64. Based on the stage cargo correlation data and refrigeration stage data, cross-segment comparison is performed to obtain correlation mutation data; Specifically, based on the chronological order of the refrigeration stage data, the relationships between the same cargo combinations in adjacent stages s-1 and s are compared. The relationship change indicators are calculated. ,in, The value is 1 if the condition is true, and 0 otherwise. When a change occurs, the changing relational components are recorded. Similarly, when a cargo's candidate location set is added to, deleted from, or its location status changes in an adjacent stage, it is also identified as a node-related mutation. The mutation occurrence time is calculated based on the stage start and end times, and the cargo nodes experiencing mutations are continuously expanded through relational edges to form the scope of the related mutation. The stage number, the previous stage relation, the current stage relation, the mutated cargo number, the mutation relation type, the mutation occurrence time, and the scope of the related mutation are encapsulated to obtain the related mutation data.

[0080] S65. Perform head reconfiguration based on the associated mutation data to obtain attention head configuration data; Specifically, read the preset attention head set. Furthermore, the feature channels, scope of action, and time offset of each attention head are adjusted according to the associated mutation type. There are 1 to H attention heads. When the associated mutation is only located in the same aisle, the same shelf level, or consecutive adjacent storage locations, the corresponding attention head is set as a local attention head, and its scope is limited to the scope of the associated mutation. When the mutation is caused by a temperature change, and there is a response time lag between the change in cargo status and the stage switch. At that time, the corresponding attention head is set to a time-delay attention head, and its time offset steps are: ,in, This represents the calculation time interval for the refrigeration space-time field data. Isolation attention heads, stacking attention heads, and anomaly attention heads are set for isolation conflict, stacking change, and anomaly stages, respectively. If the same attention head corresponds to multiple configurations, the configurations are retained in the following order: anomaly, isolation conflict, stacking, temperature zone compatibility, and spatial adjacency. The attention head configuration data is obtained by encapsulating the attention head number, attention head type, affected cargo set, affected location set, time offset steps, and activation stage.

[0081] S66. Perform multi-head learning based on attention head configuration data and stage placement sequence data to obtain placement scheme data.

[0082] Specifically, the stage placement sequence data of stage s is converted into a node feature matrix. Calculate the following for the h-th attention head: , , . ,in, , and These are the model parameters obtained through pre-training. Let h be the feature dimension of the h-th attention head. This is a mask matrix generated according to the attention head configuration data. For a local attention head, positions outside its effective range are set as non-participatory; for a time-delay attention head, the key matrix and the value matrix are set according to the time offset steps. Perform displacement; for cargo-location combinations that violate size, load capacity, temperature zone, isolation, or stacking conditions, set the corresponding mask value to negative infinity.

[0083] The outputs of each attention point are concatenated to generate the association results for each candidate storage location corresponding to each item. Based on the aforementioned matching determination results, only the unoccupied storage location with the largest association result is selected from the valid candidate storage locations; if the values ​​are the same, the storage location with the smaller storage location number is selected. After each item is allocated, the storage location occupancy status, remaining carrying capacity, and stacking layers are updated until all items are allocated. Finally, the item number, target storage location number, placement orientation, stacking order, applicable refrigeration stage, and updated storage location occupancy results are recorded to obtain the placement scheme data.

[0084] Preferably, the head rematch specifically involves: Based on the associated mutation data, local reconfiguration and time-delay reconfiguration are performed to obtain the first attention head configuration data and the second attention head configuration data, respectively. Specifically, read the stage number and the set of mutated goods from the associated mutation data. Mutation storage location set Mutation relationship types Timing of mutation The relationships before and after the mutation, and the scope of the associated mutation, are defined, where e is the associated mutation event number. When performing localized reassortment, the following is used: Starting with the mutated cargo in the data, the system traces layer by layer along the spatial adjacency, temperature compatibility, stacking, or isolation / conflict relationships in the cargo association data of each stage; it then moves to the next layer of cargo and examines the mutation relationship type with the currently associated mutation event. If the differences are not observed, or if the corresponding relationships remain unchanged in adjacent refrigeration stages, the expansion is stopped. The expanded set of goods and its candidate location set are defined as the local scope of attention. For each local attention head, the attention head number, activation stage, mutation relationship type, local scope of attention, and local mask are recorded; the local mask is 0 within the local scope of attention and negative infinity outside the local scope of attention, thus obtaining the first attention head configuration data.

[0085] During time-delayed reconfiguration, the mutated goods are sequentially connected into a response chain according to the direction of relationship propagation within the scope of the associated mutation. Let the mutation time of the starting goods in the response chain be t. The relationship between downstream goods changes at any given time. The corresponding time offset steps are: ,in This is the rounding function. This represents the calculation time interval for the space-time field data during cooling. If the number of computation time steps exceeds the limit of the stage, the response relationship is marked as invalid; otherwise, the attention head number, activation stage, response chain, affected goods set, and time offset steps are recorded for the corresponding time-delay attention head. The second attention head configuration data is obtained.

[0086] Head alignment is performed based on the first attention head configuration data and the second attention head configuration data to obtain attention head alignment data; Specifically, the configuration data of the first and second attention heads are arranged in the order of activation phase—mutation relationship type—minimum cargo number. For two attention heads that have the same activation phase and mutation relationship type, and whose localized scope intersects with the response chain's cargo set, they are placed in the same attention head position. For attention heads without a corresponding configuration, an empty configuration is set at the corresponding position, without changing the original configuration on the other side. Let the localized scope corresponding to the h-th attention head position be... The response chain acts as a collection of goods. Then through: Determine the collective goods that work together Attention should be paid to the head location, activation phase, mutation relationship type, and localization scope. Response chain function of goods collection A collection of goods that work together The local mask and time offset steps are aligned item by item to obtain attention head aligned data.

[0087] Adaptation data is obtained by performing adaptation determination based on attention head alignment data and associated mutation data; Specifically, for each attention head location, four adaptation criteria are determined: scope of action, relationship type, stage, and time sequence. This applies to the set of mutated goods corresponding to associated mutation events. Completely contained within the local scope If the local adaptation condition is met, then the response chain affects the set of goods. If the corresponding upstream and downstream goods are included, and the time offset steps do not exceed the current stage length, then the time delay adaptation condition is met; if the mutation relationship type of the attention head is consistent with the mutation relationship type of the associated mutation event... Consistent, and the activation phase includes the time when the mutation occurs. If so, then the relationship stage adaptation condition is met.

[0088] When the local fit condition and the relational stage fit condition are met, but the time delay fit condition is not met, the attention head is marked as local fit; when the time delay fit condition and the relational stage fit condition are met, but the local fit condition is not met, it is marked as time delay fit; when all three conditions are met, it is marked as multi-fit fit; otherwise, it is marked as misfit. For multi-fit attention heads, the actual scope of action is determined as follows: The time offset steps are retained; for unsuitable attention heads, a deactivation flag is set. The attention head position, adaptation type, actual range of effect, time offset steps, and deactivation flag are encapsulated to obtain the head adaptation data.

[0089] Conflict resolution is performed based on head adaptation data to obtain attention head configuration data.

[0090] Specifically, the head adaptation data is traversed to identify attention heads that act on the same goods or the same candidate cargo locations within the same activation phase. When the actual effective ranges of two attention heads overlap but their mutation relationship types differ, the attention head with higher priority is retained in the order of anomaly relationship, isolation conflict relationship, stacking relationship, temperature zone compatibility relationship, and spatial adjacency relationship; the attention head with lower priority has overlapping goods removed, and only the non-overlapping effective range is retained. When two attention heads have the same mutation relationship type but different time offset steps, their response chains are checked separately; the configuration with the smaller time offset step is retained if it can completely cover the time offset steps of the starting goods and downstream goods in the associated mutation event.

[0091] When a local attention head and a time-delay attention head apply to the same goods and are both in a compatible state, they are merged into a composite attention head. The local scope of action is used to limit the goods nodes participating in the calculation, and the corresponding key and value matrices are shifted according to the time offset steps. Attention heads whose scope is empty after conflict resolution, whose response chains are broken, or who are marked as incompatible are deactivated. Finally, attention head configuration data is recorded according to the attention head number, including attention head type, activation stage, mutation relationship type, actual scope of action, local mask, time offset steps, and activation / deactivation status. This data serves as the direct input for subsequent stages of multi-head learning when placing sequence data.

[0092] Preferably, the local reconfiguration specifically includes: Mutation location is performed based on associated mutation data to obtain mutation node data; Specifically, read the stage number s and the set of mutated goods from the associated mutation data. Mutation storage location set Mutation relationship types Timing of mutation Pre-mutation relationship and the relationship at the current stage Where e is the associated mutation event number. Calculate the number of relationship changes for each cargo node q: .when When the set of candidate storage locations for cargo q changes, or the status of the storage location corresponding to cargo q changes from normal to abnormal, cargo q is identified as a mutated cargo node. For each changed relationship, if the relationship change first occurs on cargo q, then cargo q is marked as the starting node; if the relationship change is passed from other cargoes to cargo q, then cargo q is marked as the responding node. The associated mutation event number, stage number, mutated cargo node number, node type, mutation relationship type, mutation occurrence time, and associated storage location number are encapsulated to obtain the mutated node data.

[0093] The domain boundaries are expanded based on the mutation node data and associated mutation data to obtain mutation scope data; Specifically, the set of starting nodes corresponding to the same associated mutation event. As the zeroth level scope, it expands layer by layer according to the relation edges in the stage cargo association data: Where k is the number of expansion layers. If the mutation relationship type between candidate node l and the current node is not... Candidate nodes do not belong to the set of mutated goods. Or, the candidate node corresponds to a storage location that does not belong to the set of mutation storage locations. If so, then the expansion along that relation edge stops. The expansion is terminated at the specified time, and the final set of cargo nodes is denoted as... .collection The candidate storage locations corresponding to each type of cargo are used to obtain the set of available storage locations. The associated mutation event number, stage number, mutation relationship type, and affected goods set will be included. Functional storage location collection The mutation scope data is obtained by encapsulating the set of domain boundary nodes.

[0094] Link tracing is performed based on mutation scope data and associated mutation data to obtain mutation association chain data; Specifically, in the action of goods collection Within, starting from the initial node, follow the chain along the mutation relationship type. The direction of corresponding relationship changes is tracked. For adjacent cargo nodes q and l, a directed mutation edge is established only when the relationship between them changes from unestablished to established, from established to disappeared, or the relationship attribute changes. When both nodes can serve as preceding or following nodes, the direction is determined from earliest to latest time of the mutation; when the mutation times are the same, the direction is determined from smallest to largest cargo number to avoid forming a circular link. The path from the starting node to each response node is represented as follows: Where b is the mutation association chain number, The number of link nodes. For the first to the first mutation in the b-th mutation association chain corresponding to the e-th associated mutation event. There are several cargo nodes. If there are multiple arrival paths for the same node, the path with the fewest nodes is retained; if the number of nodes is the same, the path with the smaller lexicographical order of the cargo number sequence is retained. The mutation association chain number, node order, corresponding cargo location order, mutation relationship type, and link start and end nodes are encapsulated to obtain the mutation association chain data.

[0095] Head domain mapping is performed based on mutation association chain data to obtain attention head data to be matched; Specifically, the set of attention heads pre-set by the multi-head self-attention model is read. Each attention head is pre-bound to a relation type. Based on the mutation relation type, spatial adjacency, temperature zone compatibility, stacking, isolation conflict, and anomalous relations are mapped to the corresponding attention heads. If multiple attention heads correspond to the same relation type, they are arranged in ascending order of the mutation association chain's starting node number, and attention heads are assigned sequentially in a cyclical manner. For mutation association chains... Set its node set: The scope of the link corresponding to the attention head is determined, and this scope is associated with the set of action locations in the mutation scope data. Simultaneously, retrieve the configuration number of the assigned attention head in the previous stage; if no previous stage configuration exists, set the previous stage configuration number to empty. Encapsulate the attention head number, associated mutation event number, mutation association chain number, mutation relationship type, chain scope, set of affected storage locations, and previous stage configuration number to obtain the attention head data to be assigned.

[0096] Configuration inheritance data is obtained by performing configuration inheritance based on the attention head data to be configured and the mutation scope data. Specifically, for each attention head to be assigned, its previous stage configuration number is read. This is done when the previous stage configuration number is not empty, and the mutation relationship type of the previous stage is the same as the current mutation relationship type, and the set of goods acted upon in the previous stage is the same as the set of goods acted upon in the current stage. When there is an intersection, the attention head type, local mask structure, start / stop state, and feature channel configuration of the previous stage are inherited; the specific scope of action of the previous stage is not inherited, but the set of action goods in the current mutation scope data is inherited. and function of cargo location collection Write the inheritance configuration. When no inheritable configuration exists, a new local attention head configuration is created according to the current mutation relationship type, with the initial value of the local mask defined as: The attention head number, inheritance source, mutation relationship type, affected goods set, affected location set, local mask, feature channel configuration and start / stop status are encapsulated to obtain configuration inheritance data.

[0097] Incremental updates are performed based on the configuration inheritance data and associated mutation data to obtain the first attention head configuration data.

[0098] Specifically, the set of goods that play a role in the configuration inheritance data is denoted as... The set of goods whose current role is reconfirmed by the associated mutation data is denoted as Determine the sets of goods to be added and the sets of goods to be removed, respectively: , For newly added cargo collections In the local mask, open the relationship between it and the currently active goods set, and the mutation type is... The calculation location for the removed goods set; Set the local mask value of the corresponding row and column to negative infinity. If the associated mutation data indicates that the mutation relationship has been restored to the pre-mutation relationship, delete the corresponding relationship edge; if a new mutation relationship appears, add the corresponding cargo node and cargo location to the scope of effect. After the update is complete, check that the set of affected cargo for each attention head is not empty and that all affected cargo locations belong to the mutated cargo location set. Furthermore, the local mask only exposes valid relational locations within the affected cargo set. Finally, the attention head number, activation stage, mutation relation type, affected cargo set, affected storage location set, local mask, feature channel configuration, and start / stop status are recorded to obtain the first attention head configuration data.

[0099] Preferably, the time-delay reconfiguration specifically includes: Response chain deconstruction is performed based on associated mutation data to obtain heterogeneous transfer chain data; Specifically, read the associated mutation event number e, stage number s, and mutation cargo set from the associated mutation data. Mutation storage location set Mutation relationship types Timing of mutation The relationships before and during the current stage are considered. Related mutation events are connected based on the same cargo, the same candidate cargo location, or direct relationship edges: if related mutation events... Associated mutation events It must contain at least one identical item or the same storage location, and meet the following requirements: Then establish by point to Candidate passing edge, For the first The timing of mutations in a series of related mutation events. For the first The mutation occurrence time of each associated mutation event is considered. When two mutation relationships differ but belong to spatial adjacency, temperature zone compatibility, stacking, isolation conflict, or anomalous relationships respectively, the relationship type difference is retained, and no merging of similar relationships occurs. If candidate transition edges form a closed loop, the reverse edge with the latest mutation occurrence time is deleted. The directed path from an associated mutation event without predecessors to an associated mutation event without successors is represented as: ,in, For the th heterogeneous transmission chain, This represents the number of associated mutation events in the chain. For the first to the second in the b-th heterogeneous transmission chain Each associated mutation event is encapsulated, along with the sequence of these events, the mutated cargo, the mutated cargo location, the mutation relationship type, and the time of occurrence, to obtain heterogeneous transmission chain data.

[0100] Phase lag inversion is performed based on heterogeneous transmission chain data to obtain relational phase lag data; Specifically, for heterogeneous transmission chains Adjacent association mutation events and Extract the time of mutation occurrence. If both occur only once within the current phase, calculate the time lag: , Let be the mutation occurrence time of the (j+1)th associated mutation event in the b-th heterogeneous transmission chain. Let be the mutation time of the j-th associated mutation event in the b-th heterogeneous mutation chain. When the value is less than 0, it is determined that the current event sequence does not satisfy the response propagation relationship, and the corresponding propagation edge is deleted. When the same relationship changes multiple times within the current stage, it is determined according to the calculation time interval. Constructing a sequence of relationship changes The correspondence is at the calculation time. Set the value to 1 if a change occurs, otherwise set it to 0. Calculate the candidate lag number. The following time-shift matching quantity: ,in The candidate lag step number represents the number of computational time steps required to shift the sequence of relational changes of subsequent associated mutation events relative to preceding associated mutation events. This is used to identify the relationship change of the j-th associated mutation event in the b-th heterogeneous transfer chain at the n-th computational time step. This represents the number of computation time steps included in the current stage. Select within the range Maximum minimum lag steps And order: , ,in, The duration of the current phase. For the relation phase lag, The calculation time interval between adjacent calculation moments of the cooling space-time field data. This represents the time lag between the j-th associated mutation event and the (j+1)-th associated mutation event in the b-th heterogeneous transfer chain. The heterogeneous transfer chain number, adjacent associated mutation event numbers, time lag, lag steps, and relational phase lag are encapsulated to obtain relational phase lag data.

[0101] Based on the phase lag data and heterogeneous transmission chain data, head misalignment mapping is performed to obtain head time offset data; Specifically, based on the mutation relationship types of each associated mutation event in the heterogeneous transmission chain, spatial adjacency, temperature compatibility, stacking, isolation conflict, and anomalous relationships are mapped to attention heads of corresponding relationship types. Let associated mutation events... Corresponding attention head Its downstream related mutation events Corresponding attention head Then based on the number of lag steps Establish the time offset relationship between the heads: When learning with multiple heads, focus your attention on one head. Preserve the query sequence at the current computation time step, attention head The key sequence and value sequence are shifted backward. Each computation time step. If =0, then the two attention heads establish a relationship at the same computation time step; if If so, the time offset relationship between the heads is marked as invalid. For cases where the same attention head continuously receives multiple levels of response, the hysteresis steps are accumulated along the heterogeneous transfer chain: This is used as the total offset steps from the first attention head to the j-th attention head. The attention head number, predecessor attention head number, successor attention head number, affected cargo set, affected location set, lag steps, and total offset steps are encapsulated to obtain the head time-series offset data.

[0102] Based on the head temporal offset data, segmented start-stop reconstruction is performed to obtain the second attention head configuration data.

[0103] Specifically, the current stage is divided into several continuous time segments using the start and end times of the cooling stage data, the occurrence times of each associated abrupt change event in the heterogeneous transfer chain, and their hysteresis response times as segmentation points. (For attention heads...) Its activation start time is determined as follows: Its activation and termination times The hysteresis response time is determined as either the last valid associated mutation event in the current heterogeneous transfer chain, or the termination time of the current stage, whichever is earlier. When the calculation time is... Within the current stage, if the corresponding mutation relationship remains the same, the attention head is set to enabled; if the mutation relationship reverts to the pre-mutation relationship, the affected cargo is removed from the mutation cargo set, or the total offset steps exceed the current stage length, the attention head is set to disabled.

[0104] For an attention head that is enabled, record the total offset steps of its key and value sequences; set a waiting flag for attention heads that have not yet reached their enable start time; set a termination flag for attention heads that have exceeded their enable end time. When the same attention head has different start and stop states in different segments, retain the corresponding segment configurations separately and do not merge them in advance. Encapsulate the second attention head configuration data according to the attention head number, enable stage, time sequence segment, affected cargo set, affected location set, predecessor attention head number, total offset steps, and start and stop state.

[0105] Preferably, this application also provides a goods placement optimization system for simulating the refrigeration process of a cold storage facility, used to execute the goods placement optimization method for simulating the refrigeration process of a cold storage facility as described above. The goods placement optimization system for simulating the refrigeration process of a cold storage facility includes: The 3D mesh mapping module is used to acquire 3D model data of cold storage, refrigeration pipeline data, and basic cargo data; it divides the cold storage into meshes based on the 3D model data and refrigeration pipeline data to obtain mixed mesh data of cold storage; and it performs spatial mapping on the mixed mesh data of cold storage based on the basic cargo data to obtain cargo spatial mapping data. The multi-domain model building module is used to divide the area into regions based on the cold storage hybrid grid data and cargo space mapping data to obtain multi-domain calculation data; and to build a model based on the multi-domain calculation data to obtain a refrigeration physical model. The refrigeration coupled solution module is used to acquire equipment operation data and environmental monitoring data; assign boundary values ​​to the refrigeration physical model based on the equipment operation data and environmental monitoring data to obtain initial model data; and perform coupled solution based on the initial model data to obtain refrigeration simulation data. The refrigeration state identification module is used to extract field quantities from refrigeration simulation data to obtain refrigeration spatiotemporal field data; and to identify the state based on the refrigeration spatiotemporal field data to obtain refrigeration state feature data. The cargo location association construction module is used to construct associations based on refrigeration status characteristic data, cargo space mapping data, and cargo basic data to obtain cargo location association data. The placement optimization evaluation module is used to learn multi-head self-attention relationships based on cargo location association data to obtain placement scheme data; to perform back-substitution simulation based on placement scheme data and refrigeration physical model to obtain placement simulation data; and to evaluate the scheme based on placement simulation data to obtain energy consumption prediction data.

Claims

1. A method for optimizing the placement of goods in a cold storage refrigeration process simulation, characterized in that, Includes the following steps: S1. Obtain the 3D model data of the cold storage, refrigeration pipeline data, and basic cargo data; perform mesh generation based on the 3D model data and refrigeration pipeline data to obtain mixed mesh data of the cold storage; perform spatial mapping on the mixed mesh data of the cold storage based on the basic cargo data to obtain cargo spatial mapping data. S2. Based on the cold storage hybrid grid data and cargo space mapping data, the regions are divided to obtain multi-domain calculation data; based on the multi-domain calculation data, a model is constructed to obtain the refrigeration physical model; S3. Acquire equipment operation data and environmental monitoring data; The refrigeration physical model is assigned boundary values ​​based on equipment operation data and environmental monitoring data to obtain initial model data; coupled solution is then performed based on the initial model data to obtain refrigeration simulation data; S4. Extract field quantities from the refrigeration simulation data to obtain refrigeration spatiotemporal field data; perform state identification based on the refrigeration spatiotemporal field data to obtain refrigeration state characteristic data. S5. Based on the refrigeration status characteristic data, cargo space mapping data and cargo basic data, the cargo location association data is constructed. S6: Perform multi-head self-attention relationship learning based on cargo location association data to obtain placement scheme data; perform back-substitution simulation based on placement scheme data and refrigeration physical model to obtain placement simulation data; evaluate the scheme based on placement simulation data to obtain energy consumption prediction data.

2. The method according to claim 1, characterized in that, The specific regional divisions are as follows: Grid labeling is performed based on the cold storage hybrid grid data and cargo space mapping data to obtain grid attribute labeling data; Perform domain occupancy determination on the grid attribute marker data to obtain domain occupancy determination data; Based on the domain determination data, fluid-structure partitioning is performed to obtain fluid-structure partitioning data; Interface identification data is obtained by performing interface identification based on fluid-structure partitioning data. Based on fluid-structure partitioning data and interface recognition data, regions are aggregated to obtain multi-domain computational data.

3. The method according to claim 1, characterized in that, The model construction is as follows: Attribute configuration is performed based on multi-domain computational data to obtain multi-domain physical property data; Fluid and solid modeling are performed on multi-domain physical property data and multi-domain computational data to obtain fluid equation data and heat transfer equation data. Interface coupling is performed based on fluid equation data, heat transfer equation data, and multi-domain calculation data to obtain coupled equation data. Boundary configuration is performed based on coupled equation data and multi-domain computation data to obtain boundary condition data; The refrigeration physics model is obtained by integrating the coupled equation data and boundary condition data.

4. The method according to claim 1, characterized in that, The coupling solution is specifically as follows: Discretize the equations based on the initial model data to obtain the discrete model data; Momentum prediction is performed based on the discrete data of the model to obtain the predicted flow field data; Pressure velocity correction is performed on the predicted flow field data to obtain the corrected flow field data; Turbulent update is performed on the corrected flow field data to obtain turbulent flow field data; Heat transfer coupling is performed on turbulent flow field data and discrete model data to obtain fluid-structure interaction field data; The refrigeration simulation data is obtained by iterative convergence based on the fluid-structure interaction field data.

5. The method according to claim 1, characterized in that, State recognition specifically refers to: Gradient calculations are performed based on the spatiotemporal field data of the cooling process to obtain the spatiotemporal field gradient data; Fluctuation extraction is performed on the spatiotemporal field data of cooling and the spatiotemporal field gradient data to obtain the temporal fluctuation data of field quantity. Based on the spatiotemporal field gradient data and the temporal fluctuation data of the field quantity, uniform evaluation is performed to obtain the uniform characteristic data of the field quantity. Cold energy is determined by analyzing the spatial-temporal field data and the uniformity characteristic data of the cooling field to obtain cold energy coverage characteristic data. Anomaly identification is performed based on spatiotemporal field gradient data, field temporal fluctuation data, and cold load coverage characteristic data to obtain refrigeration anomaly characteristic data. Based on the uniformity characteristic data of field quantity, the characteristic data of cold quantity coverage, and the characteristic data of refrigeration anomaly, the characteristic data of refrigeration status are collected to obtain the characteristic data of refrigeration status.

6. The method according to claim 1, characterized in that, Multi-head self-attention relationship learning specifically includes: Stage identification is performed based on refrigeration status characteristic data to obtain refrigeration stage data; The cargo location association data and refrigeration stage data are segmented into stages to obtain stage placement sequence data; Based on the stage placement sequence data, correlation modeling is performed to obtain stage cargo correlation data; Cross-segment comparisons were performed based on the cargo correlation data and refrigeration stage data to obtain correlation mutation data; Head reconfiguration is performed based on associated mutation data to obtain attention head configuration data; Multi-head learning is performed based on attention head configuration data and stage placement sequence data to obtain placement scheme data.

7. The method according to claim 6, characterized in that, The head rematch specifically refers to: Based on the associated mutation data, local reconfiguration and time-delay reconfiguration are performed to obtain the first attention head configuration data and the second attention head configuration data, respectively. Head alignment is performed based on the first attention head configuration data and the second attention head configuration data to obtain attention head alignment data; Adaptation data is obtained by performing adaptation determination based on attention head alignment data and associated mutation data; Conflict resolution is performed based on head adaptation data to obtain attention head configuration data.

8. The method according to claim 7, characterized in that, Local reconfiguration specifically refers to: Mutation location is performed based on associated mutation data to obtain mutation node data; The domain boundaries are expanded based on the mutation node data and associated mutation data to obtain mutation scope data; Link tracing is performed based on mutation scope data and associated mutation data to obtain mutation association chain data; Head domain mapping is performed based on mutation association chain data to obtain attention head data to be matched; Configuration inheritance data is obtained by performing configuration inheritance based on the attention head data to be configured and the mutation scope data. Incremental updates are performed based on the configuration inheritance data and associated mutation data to obtain the first attention head configuration data.

9. The method according to claim 7, characterized in that, The time-delay reconfiguration is specifically as follows: Response chain deconstruction is performed based on associated mutation data to obtain heterogeneous transfer chain data; Phase lag inversion is performed based on heterogeneous transmission chain data to obtain relational phase lag data; Based on the phase lag data and heterogeneous transmission chain data, head misalignment mapping is performed to obtain head time offset data; Based on the head temporal offset data, segmented start-stop reconstruction is performed to obtain the second attention head configuration data.

10. A system for optimizing the placement of goods in a cold storage refrigeration process, characterized in that, The method for optimizing the placement of goods in a cold storage refrigeration process simulation as described in claim 1, wherein the system for optimizing the placement of goods in a cold storage refrigeration process simulation comprises: The 3D mesh mapping module is used to acquire 3D model data of cold storage, refrigeration pipeline data, and basic cargo data; it divides the cold storage into meshes based on the 3D model data and refrigeration pipeline data to obtain mixed mesh data of cold storage; and it performs spatial mapping on the mixed mesh data of cold storage based on the basic cargo data to obtain cargo spatial mapping data. The multi-domain model building module is used to divide the area into regions based on the cold storage hybrid grid data and cargo space mapping data to obtain multi-domain calculation data; and to build a model based on the multi-domain calculation data to obtain a refrigeration physical model. The refrigeration coupled solution module is used to acquire equipment operation data and environmental monitoring data; assign boundary values ​​to the refrigeration physical model based on the equipment operation data and environmental monitoring data to obtain initial model data; and perform coupled solution based on the initial model data to obtain refrigeration simulation data. The refrigeration state identification module is used to extract field quantities from refrigeration simulation data to obtain refrigeration spatiotemporal field data; and to identify the state based on the refrigeration spatiotemporal field data to obtain refrigeration state feature data. The cargo location association construction module is used to construct associations based on refrigeration status characteristic data, cargo space mapping data, and cargo basic data to obtain cargo location association data. The placement optimization evaluation module is used to learn multi-head self-attention relationships based on cargo location association data to obtain placement scheme data; to perform back-substitution simulation based on placement scheme data and refrigeration physical model to obtain placement simulation data; and to evaluate the scheme based on placement simulation data to obtain energy consumption prediction data.

Images