A large model-based warehouse network planning method and simulation verification system

Abstract

The application relates to the technical field of logistics management, in particular to a warehouse network planning method based on a large model and a simulation verification system, which comprises the following steps: obtaining the number, coordinates, transportation routes and cargo demand of warehouses, inputting a graph neural network to deduce a warehouse network structure and learn logistics directions, generating a structure planning result, obtaining the inflow and outflow and in-transit time fed back by the transportation routes, calculating the net cargo flow difference and error analysis to identify abnormal line groups, extracting the associated warehouse in-out quantity and stock utilization rate to generate a residual time sequence, analyzing the trend in different time periods and marking abnormal trajectories, judging the number of abnormal warehouses and the connection relationship with abnormal lines, and outputting a simulation verification result.In the application, the graph neural network prediction and multi-source data residual analysis are fused, the logistics abnormal trend is identified, the warehouse network deviation is dynamically fed back, the structure planning precision and verification closed-loop capability are enhanced, and the system perception depth, trend identification strength and decision robustness are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of logistics management, in particular to a warehouse network planning method based on a large model and a simulation verification system. BACKGROUND

[0002] The technical field of logistics management includes a technical system for systematically organizing and scheduling the flow and storage of materials in the supply chain. The core content includes warehouse layout design, transportation path optimization, distribution center site selection, inventory control method, supply chain coordination management and information support system. The overall technical field not only covers the process arrangement of receiving, storing, sorting and distributing goods, but also involves the overall planning of warehouse network by using computer modeling, simulation and optimization means, thereby forming a full-process coverage of macro organization and micro execution of the logistics system.

[0003] Among them, the warehouse network planning method and simulation verification system based on a large model refer to a technical solution that applies deep neural networks and large-scale parameter models to generate schemes for warehouse node layout, transportation connection mode and distribution level during warehouse network planning, and verifies the consistency and feasibility of the generated schemes through computer simulation. It mainly covers warehouse node demand prediction, transportation line reachability modeling, warehouse facility site selection and level allocation strategy, as well as simulation scenario construction and operation.

[0004] Although the existing technology realizes the overall planning of the goods flow process under the information support system, it mainly relies on static data analysis and rule-driven methods in actual operation, lacking real-time feedback mechanism for the running state of warehouse nodes and transportation lines. In a multi-node system, the coupling relationship between warehouse inventory changes and transportation line cargo flow cannot be dynamically identified, resulting in that once a bottleneck or abnormality occurs in a transportation line, the system cannot quickly locate its influence range and root warehouse. For example, if the cargo inflow of a first-level distribution center suddenly decreases, the traditional system cannot calculate the residual error and judge the trend through the feedback data of upstream and downstream nodes, which easily leads to lagging identification of faults or significant deviation of planning schemes in actual implementation. The existing method fails to build a residual error identification and trend capturing path based on time sequence feedback, so that the cumulative effect of operation deviation at multiple time stamps is ignored, affecting the overall warehouse network operation efficiency and service level stability. SUMMARY

[0005] To solve the technical problems existing in the prior art, the embodiments of the present application provide a warehouse network planning method based on a large model, which includes the following steps:

[0006] To achieve the above-mentioned purpose, the present application adopts the following technical scheme: a warehouse network planning method based on a large model, including the following steps:

[0007] S1: Obtain the number of warehouses, coordinates, transportation routes and cargo demand, input the graph neural network to deduce the warehouse network spatial structure and learn the logistics flow, and construct a warehouse network planning result;

[0008] S2: Obtain the feedback cargo inflow, outflow and in-transit time of the transportation route, calculate the net cargo flow difference and feedback in-transit inventory reference value, perform error analysis on the warehouse network planning result and judge whether the difference threshold is exceeded, and generate an abnormal route group;

[0009] S3: Obtain the feedback cargo in and out of the warehouse and inventory utilization rate of the associated warehouse of the abnormal route group, calculate the net inventory change and warehouse capacity change and analyze the warehouse operation feedback residual under multiple time stamps, and generate a residual time series;

[0010] S4: Based on the residual time series, the residual is summarized by single warehouse and time period, and the residual change trend is analyzed, if the residual direction is the same and the residual change trend is increasing, the warehouse is marked as abnormal, and a residual trend abnormal trajectory set is generated;

[0011] S5: Obtain the residual trajectory of all abnormal warehouses in the residual trend abnormal trajectory set and accumulate, when the number of abnormal warehouses exceeds the half threshold and there is a connection relationship with the abnormal transportation route, it is judged that the planning result has deviation, and a warehouse network simulation verification result is generated.

[0012] As a further scheme of the present application, the warehouse network planning result includes the number of warehouses, warehouse coordinates, and transportation routes, the abnormal route group specifically includes the net cargo flow difference, feedback in-transit inventory reference value, and difference threshold, the residual time series includes the net inventory change, warehouse capacity change, and warehouse operation feedback residual, the residual trend abnormal trajectory set specifically refers to the residual direction, residual change trend, and abnormal warehouse marking, and the warehouse network simulation verification result specifically includes the number of abnormal warehouses, abnormal warehouse residual trajectory cumulative value, and abnormal transportation route connection relationship.

[0013] As a further scheme of the present application, the specific steps of S1 are:

[0014] S101: Obtain the number of warehouses, warehouse coordinates and transportation route parameters, and perform weighted accumulation on the number of warehouses and the length of multiple paths in the transportation route, then compare the accumulated value with the cargo demand, select the path set that meets the cargo demand, and generate a path matching set;

[0015] S102: Based on the path matching set, construct a corresponding graph structure data frame, vectorize the warehouse coordinates and cargo demand, and perform splicing operation with the graph structure data frame to obtain a warehouse network input feature set;

[0016] S103: According to the warehouse network input feature set, performing layer-by-layer propagation calculation in the graph neural network, and performing normalized comparison on the state values and node demand of the multi-propagation layer output, screening the stable convergence propagation output result, and obtaining the warehouse network planning result.

[0017] As a further scheme of the present application, the specific steps of S2 are:

[0018] S201: Obtain the feedback goods inflow quantity, feedback goods outflow quantity and in-transit time parameter of the transportation line, perform difference calculation on the inflow quantity and outflow quantity, and then weight the difference value and the in-transit time parameter to generate a net goods flow difference;

[0019] S202: Based on the net goods flow difference, call the in-transit time parameter and the feedback goods inflow quantity to perform multiplication operation, and compare the result with the outflow quantity to calculate a feedback in-transit stock reference value;

[0020] S203: According to the feedback in-transit stock reference value and the warehouse network planning result, compare each item, judge the comparison value and the preset difference threshold, select the line exceeding the difference threshold, and obtain an abnormal line group.

[0021] As a further scheme of the present application, the difference threshold is obtained by statistically sampling the feedback goods inflow quantity, feedback goods outflow quantity and in-transit time parameter of the transportation line, calculating the mean and variance of the net goods flow difference of multiple lines, and then linearly weighting the mean and variance results to set.

[0022] As a further scheme of the present application, the specific steps of S3 are:

[0023] S301: Obtain the feedback goods storage quantity, feedback goods out-of-storage quantity and inventory utilization rate data of the associated warehouse of the abnormal line group, perform difference calculation on the storage quantity and the out-of-storage quantity, weight the difference result and the inventory utilization rate parameter to generate a net inventory change quantity;

[0024] S302: Based on the net inventory change quantity, obtain the warehouse's accommodable parameter and inventory utilization rate data, calculate the remaining accommodable quantity of multiple warehouses, and perform warehouse-by-warehouse comparison between the calculation result and the net inventory change quantity to generate a warehouse accommodable change quantity;

[0025] S303: According to the warehouse accommodable change quantity and the net inventory change quantity, perform time-by-time difference operation on the warehouse feedback data under multiple time stamps, compare the difference residual of each time stamp with a preset residual threshold, extract the residual information of all time stamps, and obtain a residual time sequence.

[0026] As a further scheme of the present application, the residual threshold is calculated by statistically sampling the difference between the net inventory change and the warehouse change capacity in the warehouse feedback data under multiple timestamps, calculating the mean and variance of the residual value, and then performing nonlinear weighting on the mean and variance.

[0027] As a further scheme of the present application, the specific steps of S4 are:

[0028] S401: Based on the residual time series, the residual values of multiple time periods are accumulated and summed, and the residuals of multiple time periods for the same warehouse are arranged and summarized to generate a warehouse-specific residual sequence;

[0029] S402: The warehouse-specific residual sequence is called to extract the sign for the residual value change direction of multiple time periods, and the residual difference values of adjacent time periods are sorted and judged. If the signs are consistent, it is determined that the directions are the same and recorded to obtain a residual direction sequence;

[0030] S403: According to the residual direction sequence, the difference result of the residual value is subjected to monotonicity determination. If the residual directions are the same and the difference result is in an increasing relationship, the corresponding warehouse index position is marked as abnormal, and the time sequence trajectories of all abnormal warehouses are aggregated to generate a residual trend abnormal trajectory set.

[0031] As a further scheme of the present application, the specific steps of S5 are:

[0032] S501: Based on the residual trend abnormal trajectory set, the residual trajectories corresponding to all abnormal warehouses are retrieved, and the residual trajectories of the same warehouse are subjected to cumulative operation in the time dimension to generate an abnormal warehouse residual accumulation sequence;

[0033] S502: The abnormal warehouse residual accumulation sequence is called to statistically count the number of abnormal warehouses and calculate the ratio with the total number of warehouses. If the ratio is greater than the half threshold, the abnormal warehouse set that meets the condition is recorded to obtain an abnormal warehouse set;

[0034] S503: According to the abnormal warehouse set, the connection relationship between multiple abnormal warehouses and abnormal transportation lines is retrieved, and the warehouse index positions that exist are logically judged. If the judgment is correct, the associated warehouse and line mapping are marked to establish a warehouse network simulation verification result;

[0035] The half threshold is set by statistically counting the total number of warehouses and multiplying the fixed proportion factor 0.5.

[0036] A warehouse network simulation verification system based on a large model, comprising:

[0037] The structural deduction module obtains the number of warehouses, coordinates, transportation routes, and cargo requirements. It then inputs these data into a graph neural network to perform warehouse network spatial structure deduction and logistics flow direction learning. Finally, it constructs the warehouse network planning results and transmits them to the error analysis module.

[0038] The error analysis module obtains the feedback cargo inflow, outflow and transit time of the transportation route, calculates the net cargo flow difference and the feedback in-transit inventory reference value, performs error analysis with the warehouse network planning results and determines whether the difference threshold is exceeded, generates abnormal route groups and transmits them to the residual generation module.

[0039] The residual generation module obtains the feedback cargo inbound and outbound volume and inventory utilization rate of the warehouse associated with the abnormal route group, calculates the net inventory change and warehouse capacity change, analyzes the warehouse operation feedback residual under multiple time stamps, generates a residual time series, and transmits it to the trend recognition module.

[0040] The trend identification module, based on the residual time series, summarizes the residuals by time period for each warehouse and analyzes the residual change trend. If the residual direction is the same and the residual change trend is increasing, the warehouse is marked as abnormal, and a residual trend abnormal trajectory set is generated and transmitted to the simulation verification module.

[0041] The simulation verification module acquires and accumulates the residual trajectories of all abnormal warehouses in the residual trend abnormal trajectory set. When the number of abnormal warehouses exceeds half the threshold and there is a connection relationship with abnormal transportation routes, it is determined that there is a deviation in the planning result, and the warehouse network simulation verification result is generated.

[0042] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0043] In this invention, data such as warehouse quantity, coordinates, transportation routes, and cargo demand are input into a graph neural network to perform warehouse network spatial structure deduction and logistics flow direction learning. Based on multi-dimensional information interaction, refined modeling and prediction of the warehouse network structure are achieved, significantly improving the ability to organize complex warehouse network spaces. Based on transportation route feedback data, the net cargo flow difference is calculated and error analysis is performed by combining inflow, outflow, and transit time, enabling the dynamic state of the logistics network to be quantitatively perceived and abnormal route combinations deviating from the planned expectations to be identified in a timely manner. Based on the acquisition of warehouse inbound and outbound volumes and inventory utilization rates, net inventory changes and capacity changes are further calculated, and residual trajectories are analyzed under multiple timestamps, effectively transforming time-series data into diagnostic indicators reflecting the health of warehouse operations. Through trend analysis, incremental residuals of abnormal inventory are identified, capturing weak but persistent operational deviations. By combining the connection relationship between multi-warehouse residual trajectories and the transportation network, closed-loop simulation verification of overall planning deviations is finally achieved. This logic organically combines multi-source feedback data, time series analysis, and network structure modeling, enabling the logistics system simulation verification process to have continuous perception and trend recognition capabilities. It effectively compensates for the problems of traditional methods such as delayed response to dynamic anomalies and insufficient fine-grained modeling, thereby improving the reliability and robustness of warehouse network planning decisions. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a schematic diagram of the steps of the present invention;

[0046] Figure 2 This is a detailed schematic diagram of S1 of the present invention;

[0047] Figure 3 This is a detailed schematic diagram of S2 of the present invention;

[0048] Figure 4 This is a detailed schematic diagram of S3 of the present invention;

[0049] Figure 5 This is a detailed schematic diagram of S4 of the present invention;

[0050] Figure 6 This is a detailed schematic diagram of S5 of the present invention;

[0051] Figure 7 This is a system module diagram of the present invention. Detailed Implementation

[0052] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0053] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0054] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0055] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0056] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0057] Please see Figure 1 This invention provides a warehouse network planning method based on a large model, comprising the following steps:

[0058] S1: Obtain the number of warehouses, coordinates, transportation routes, and cargo requirements, input them into a graph neural network to perform warehouse network spatial structure deduction and logistics flow direction learning, and construct warehouse network planning results;

[0059] S2: Obtain feedback cargo inflow, outflow and transit time for transportation routes, calculate the net cargo flow difference and feedback on-transit inventory reference value, perform error analysis with warehouse network planning results and determine whether the difference threshold is exceeded, and generate abnormal route groups.

[0060] S3: Obtain the feedback of goods in and out of warehouses associated with abnormal route groups and inventory utilization rate, calculate the net inventory change and warehouse capacity change, analyze the warehouse operation feedback residuals under multiple timestamps, and generate residual time series.

[0061] S4: Based on the residual time series, summarize the residuals by time period for each warehouse and analyze the residual change trend. If the residuals are in the same direction and the residual change trend is increasing, mark the warehouse as an anomaly and generate a set of residual trend anomaly trajectories.

[0062] S5: Obtain and accumulate the residual trajectories of all abnormal warehouses in the residual trend abnormal trajectory set. When the number of abnormal warehouses exceeds half the threshold and there is a connection with abnormal transportation routes, it is determined that there is a deviation in the planning result, and the warehouse network simulation verification result is generated.

[0063] The warehouse network planning results include the number of warehouses, warehouse coordinates, and transportation routes. The abnormal route group specifically includes the net cargo flow difference, the feedback reference value of in-transit inventory, and the difference threshold. The residual time series includes the net inventory change, the warehouse capacity change, and the warehouse operation feedback residual. The residual trend abnormal trajectory set specifically refers to the residual direction, the residual change trend, and the abnormal warehouse marker. The warehouse network simulation verification results specifically include the number of abnormal warehouses, the cumulative value of the abnormal warehouse residual trajectory, and the connection relationship of abnormal transportation routes.

[0064] Please see Figure 2 The specific steps of S1 are as follows:

[0065] S101: Obtain the number of warehouses, warehouse coordinates, and transportation route parameters. Then, perform a weighted summation of the number of warehouses and the lengths of multiple routes in the transportation route. Finally, compare the summation with the demand for goods, filter out the set of routes that meet the demand for goods, and generate a route matching set.

[0066] First, the total number of valid warehouses is obtained; in this example, there are three. Then, the geographical coordinates of the multiple warehouses are retrieved, and all transportation route parameters connecting the warehouses and distribution points are collected. Warehouse coordinates are obtained through the built-in GPS module of the warehouse management system. Specifically, warehouse W1 is located at longitude 121.25°, latitude 31.38°, warehouse W2 at longitude 121.22°, latitude 31.03°, and warehouse W3 at longitude 121.11°, latitude 31.15°. Transportation route parameters are obtained by calling the electronic map service application interface, resulting in five potential transportation routes. Route P1 is a single-segment path from W1 to the destination, with a length of 25 kilometers. Route P2 is a single-segment path from W2 to the destination, with a length of 20 kilometers. Route P3 is a single-segment path from W3 to the destination, with a length of 35 kilometers. Route P4 is a multi-segment path, consisting of a 15-kilometer segment from W1 to W3 and a 35-kilometer segment from W3 to the destination. Route P5 is also a multi-segment path, consisting of a 10-kilometer segment from W2 to W3 and a 35-kilometer segment from W3 to the destination. Next, the number of warehouses and the lengths of the multiple segments in the transportation route are weighted and accumulated. The weight for the number of warehouses is set to 5, referencing the relationship between fixed and variable costs in warehouse network operation, based on historical data analysis. The weight for the multiple segments is set to 1.0, referencing the additional costs incurred in transit links, based on experimental data analysis, and using a weight of 1.14 for a single segment and 1.14 for multiple segments. Based on this, the weighted values ​​for the multiple routes are calculated. The weighted value for route P1 is 1*5 + 25*1.0, resulting in 30.0. The weighted value for route P2 is 1*5 + 20*1.0, resulting in 25.0. The weighted value for route P3 is 1*5 + 35*1.0, resulting in 40.0. The weighted value for route P4 is the sum of 2*5 + 15*1.14 and 35*1.0, resulting in 62.1. The weighted value for route P5 is the sum of 2*5 + 10*1.14 and 35*1.0, resulting in 56.4. Finally, the accumulated value is compared with the demand for goods. Assume the total demand for goods is 5000 units, and the available inventory for W1, W2, and W3 is 3000, 4000, and 2500 units respectively. Routes P1, P2, and P3 are eliminated because the inventory in a single warehouse is less than 5000 units. The total inventory of W1 and W3 involved in route P4 is 5500 units, meeting the demand. The total inventory of W2 and W3 involved in route P5 is 6500 units, meeting the demand. After filtering, a path matching set is generated, which includes route P4 and its data, and route P5 and its data.

[0067] S102: Based on the path matching set, construct the corresponding graph structure data frame, vectorize the warehouse coordinates and the demand for goods, and concatenate them with the graph structure data frame to obtain the warehouse network input feature set;

[0068] Based on the generated path matching set, the corresponding graph structure data is first constructed. This path matching set indicates that there are valid connections between warehouses W1 and W3, and between W2 and W3. Accordingly, an adjacency data structure describing the topological connections between warehouse nodes is constructed. In this structure, there are connections between W1 and W3, and between W2 and W3, but no direct connection exists between W1 and W2. Subsequently, warehouse coordinates and cargo demand are vectorized. The geographical coordinates of the warehouses are used directly as spatial features. A non-numerical data type, "warehouse type," is introduced and quantified based on its functional positioning. The specific quantification standard is as follows: ordinary regional warehouses are assigned a value of 1, specialized warehouses with special processing capabilities are assigned a value of 2, and large central hub warehouses with both storage and distribution functions are assigned a value of 3. This standard was developed after analyzing the roles and cost structures of 10 different functional warehouses in the network. In this embodiment, W1 and W2 are ordinary regional warehouses with a quantification value of 1; W3 is a central hub warehouse with a quantification value of 3. Next, the vectorized features are concatenated. The concatenation operation here refers to organizing the features of multiple warehouse nodes into a feature set. Each node's feature vector consists of longitude, latitude, quantified warehouse type, and current inventory reflecting its supply capacity. Specifically, the feature vector of W1 is (121.25, 31.38, 1, 3000); the feature vector of W2 is (121.22, 31.03, 1, 4000); and the feature vector of W3 is (121.11, 31.15, 3, 2500). Finally, the graph structure data describing node connections is integrated with the set containing multi-node features to obtain the complete warehouse network input feature set. This feature set encapsulates all node attribute information and the connections between nodes in the warehouse network, serving as input data for the next step.

[0069] S103: Based on the input feature set of the warehouse network, perform layer-by-layer propagation calculation in the graph neural network, and normalize and compare the state values ​​output by multiple propagation layers with the node demand, select stable and convergent propagation output results, and obtain the warehouse network planning results.

[0070] Based on the acquired warehouse network input feature set, layer-by-layer propagation computation is performed in the graph neural network. In each layer, the new state of each warehouse node is generated by aggregating and transforming the state of its previous layer and the states of all its neighboring nodes. This process uses a transformation matrix to map the input features to the hidden state; the values ​​of this matrix are obtained through regression analysis of 1000 historical optimal scheduling schemes. The specific actions of the propagation computation include: First, normalizing the graph structure describing the node connection relationships, adding self-loops to each node and adjusting them according to the number of connections. Then, performing an aggregation operation, the new feature of each node is a weighted sum of its own and its neighboring nodes' previous layer features. Finally, applying the transformation matrix to the aggregated features for linear transformation and executing a modified linear unit activation function, setting all calculation results less than zero to zero. This process is repeated multiple times; in this example, it is set to three layers. The advantage of this approach is that, through layer-by-layer information propagation and nonlinear transformation, the final state vector of each node can incorporate the structural and feature information of its multi-level neighbors. After executing three propagation layers, three sets of output state values ​​are obtained. The state values ​​output from multiple propagation layers are normalized and compared to filter out stable and convergent propagation outputs. The specific action for convergence judgment is to calculate the sum of the absolute differences between all corresponding elements of the third-layer output state and the second-layer output state, and compare this sum with a preset convergence threshold of 0.5. This threshold was determined through simulation experiments on 50 warehouse network planning problems of different complexities. The experiments show that the state changes corresponding to effective solutions are all less than 0.5. If the calculated sum of differences, for example, 0.6, is greater than this threshold, it indicates that the state has not yet converged, and the number of propagation layers needs to be increased and the comparison repeated until the difference is less than 0.5. Once a stable and convergent propagation output is obtained, it is converted into specific cargo allocation quantities through an output mapping layer. The mapping layer converts the state vectors of multiple nodes into scalar values, i.e., the cargo allocation quantities that multiple warehouses should undertake; this set of values ​​is the warehouse network planning result.

[0071] Please see Figure 3 The specific steps of S2 are as follows:

[0072] S201: Obtain the feedback cargo inflow, feedback cargo outflow, and transit time parameters of the transportation route, perform differential calculation on the inflow and outflow, and then weight the differential value with the transit time parameter to generate the net cargo flow difference.

[0073] First, feedback data on cargo inflow, outflow, and transit time for each transportation route is obtained. These parameters are collected in real time using RFID readers deployed at warehouse entrances and exits, and GPS devices on transport vehicles. RFID devices record the time and quantity of goods passing through specific nodes, generating inflow and outflow data; GPS is used to calculate the actual transit time between two nodes. For routes P4 (segment W1W3) and P5 (segment W2W3) in the aforementioned path matching set, the feedback data is as follows: Route P4 has an outflow of 900 boxes / hour at the W1 exit and an inflow of 850 boxes / hour at the W3 entrance, with an actual transit time of 0.6 hours. Route P5 has an outflow of 1950 boxes / hour at the W2 exit and an inflow of 1900 boxes / hour at the W3 entrance, with an actual transit time of 0.4 hours. Next, differential calculations are performed on the inflow and outflow volumes for multiple routes. This calculation involves subtracting the corresponding inflow from the outflow within a specific time period. The flow difference value for line P4 = Boxes / hour. The flow differential value for line P5 = The calculated difference is then weighted with the transit time parameter to generate the net cargo flow difference. This weighting operation introduces a "time cost conversion factor," which quantifies the transit time dimension into a cost metric with the same unit of cargo volume. This factor is set based on a statistical analysis of 100 historical transport data batches. This analysis shows that for every additional hour of transit time, the indirect management costs incurred due to delays, cargo status monitoring, and capital tied up are equivalent to the value of 20 boxes of cargo after financial accounting. To verify the rationality of this setting, a 30-day comparative experiment was conducted, setting up two groups of transport tasks: one using standard transit time, and the other artificially increasing transit time. By comparing the differences in total operating costs between the two groups, the linear relationship between transit time and indirect costs was verified, and the effectiveness of the conversion factor of 20 was confirmed. Therefore, the time cost conversion factor was set to 20 boxes / hour. The specific execution of the weighted calculation is: Net Cargo Flow Difference = Flow Difference Score + Transit Time Time cost conversion factor. Net freight volume difference for line P4 = The net freight flow difference of line P5 = A higher net cargo flow difference indicates that the route not only faces physical cargo loss or delays during the current period, but also a higher risk of time costs.

[0074] S202: Based on the net cargo flow difference, the in-transit time parameter is called and multiplied with the feedback cargo inflow, and the result is compared with the outflow to calculate the feedback in-transit inventory reference value.

[0075] Based on the acquired real-time feedback parameters of the transportation routes, namely the transit time and the feedback cargo outflow volume, the two are multiplied. The purpose of this calculation is to estimate the total amount of cargo actually in transit at a specific point in time. The calculation does not rely on planned data but directly uses real-time data collected from the logistics execution process. The specific values ​​of the referenced parameters are as follows: the transit time for route P4 is 0.6 hours, and the feedback cargo outflow volume is 900 boxes / hour; the transit time for route P5 is 0.4 hours, and the feedback cargo outflow volume is 1950 boxes / hour. The transit time of each route is multiplied by the corresponding feedback cargo outflow volume. The transit inventory of route P4 is calculated as follows: Calculation of the in-transit inventory of route P5 = Subsequently, the result of the above multiplication operation is used as a reference value for the in-transit inventory. This reference value is a quantitative estimate of the dynamic in-transit inventory based on real-time data. The advantage of this method is that by using real-time outflow and transit time data, it replaces the traditional static estimation method based on planned values ​​or historical averages, thus making the monitoring of in-transit inventory more accurate and able to reflect fluctuations in the number of in-transit items caused by unforeseen circumstances such as traffic congestion and loading / unloading delays. The feedback in-transit inventory reference value for route P4 is 540 items, indicating that 540 items of goods are en route between W1 and W3. The feedback in-transit inventory reference value for route P5 is 780 items, indicating that 780 boxes of goods are en route between W2 and W3.

[0076] S203: Compare the feedback on the reference value of the inventory in transit with the warehouse network planning results item by item, judge the difference between the comparison value and the preset difference threshold, filter the routes that exceed the difference threshold, and obtain the abnormal route group.

[0077] The calculated feedback on in-transit inventory reference value is compared item by item with the generated warehouse network planning results. The warehouse network planning results include the planned transportation volume for multiple routes. In this embodiment, the warehouse network planning results specify that route P4 (W1W3 segment) should carry 1000 boxes of goods, and route P5 (W2W3 segment) should carry 2000 boxes of goods. Based on this planning, the planned in-transit inventory is calculated: the planned in-transit inventory for route P4 is... The planned inventory of Line P5 is as follows: The feedback on in-transit inventory reference values ​​are 540 boxes for line P4 and 780 boxes for line P5. Comparison value = |Feedback on in-transit inventory reference value | / Planned in-transit inventory 100%. Comparison value for line P4 = The comparison value of line P5 = Next, the calculated comparison value is compared with a preset difference threshold. The difference threshold is set based on the analysis of 200 historical transportation task execution data. When the difference between planned and actual in-transit inventory is less than 5%, the probability of negative events such as delays, cargo damage, and customer complaints remains at a low level (below 2.5%). Once the difference exceeds 5%, the incidence of negative events jumps to 16.0% and increases sharply as the difference increases. Through receiver operating characteristic curve analysis, setting the difference threshold at 5% optimally balances the sensitivity and specificity of the warning. Therefore, the preset difference threshold is 5%. The judgment process is to compare the comparison value of each route with 5%. The comparison value of route P4 is 8%, which is greater than 5%. The comparison value of route P5 is 2.5%, which is less than 5%. Based on this judgment result, routes with comparison values ​​exceeding the difference threshold are filtered out. In this case, the comparison value of route P4 exceeds the 5% threshold and is therefore filtered out. The comparison value of route P5 does not exceed the threshold and is not filtered out. All filtered routes are aggregated to obtain the abnormal route group. In this embodiment, the abnormal line group obtained is {P4}.

[0078] Please see Figure 4 The specific steps of S3 are as follows:

[0079] S301: Obtain feedback goods inbound quantity, feedback goods outbound quantity, and inventory utilization rate data of the warehouse associated with the abnormal route group, perform differential calculation on the inbound quantity and outbound quantity, weight the differential result with the inventory utilization rate parameter, and generate the net inventory change.

[0080] First, the system acquires feedback data on inbound and outbound goods, as well as inventory utilization, from warehouses (W1 and W3) associated with the identified abnormal route group {P4}. This data is acquired in real-time through a warehouse management system (WMS). Inbound and outbound volumes are collected by RFID readers deployed at warehouse entrances as goods pass through. Inventory utilization is calculated by the system based on real-time inventory levels and the total warehouse capacity. At the current time point in this embodiment, the acquired data is as follows: warehouse W1 has an inbound volume of 0 boxes / hour, an outbound volume of 900 boxes / hour, a total capacity of 5000 boxes, and a current inventory of 3000 boxes. The calculated inventory utilization rate is... Warehouse W3 reports an inbound volume of 850 boxes / hour (from line P4) and an outbound volume of 0 boxes / hour. Its total capacity is 4000 units, and the current inventory is 2500 boxes. The calculated inventory utilization rate is... Next, a difference calculation is performed on the inbound and outbound quantities for each associated warehouse. Specifically, the outbound quantity is subtracted from the inbound quantity. The inventory difference value for warehouse W1 is... Boxes / hour. The inventory difference for warehouse W3 is [value missing]. Boxes / hour. Subsequently, the difference result is weighted with the inventory utilization rate parameter to generate the net inventory change. This weighting operation introduces an "inventory utilization rate risk coefficient" to quantify the operational risks such as congestion and mismatches caused by high inventory utilization, and converts them into equivalent goods. The coefficient is set with reference to a regression analysis experiment on the operational efficiency of 50 warehouses at different utilization rate levels. Experimental data shows that when the inventory utilization rate exceeds the safety line of 70%, each additional 1% increase in utilization rate results in an additional operational cost (such as increased picking time and temporary storage costs) that is on average equivalent to the cost of handling 15 goods. Therefore, the inventory utilization rate risk coefficient is set to 15. The specific execution of the weighted calculation is: Net Inventory Change = Inventory Difference Value + Inventory Utilization Rate Inventory utilization risk coefficient. Net inventory change of warehouse W1 = The results show that W1's inventory is experiencing a net outflow at a rate equivalent to 891 boxes per hour. The net inventory change in warehouse W3 = This result indicates that W3's inventory is experiencing a net inflow at a rate equivalent to 859.375 units per hour.

[0081] S302: Based on the net inventory change, obtain the warehouse's capacity parameters and inventory utilization data, calculate the remaining capacity of multiple warehouses, and perform a warehouse-by-warehouse comparison with the calculation results and the net inventory change to generate the warehouse capacity change.

[0082] Based on the calculated net inventory change, the capacity parameters and inventory utilization data of the associated warehouses (W1 and W3) in the abnormal route group are first obtained. The capacity parameter is the total design capacity of the warehouse, and the inventory utilization data is directly retrieved from S301. Specifically, warehouse W1 has a total capacity of 5000 boxes and an inventory utilization rate of 60%; warehouse W3 has a total capacity of 4000 boxes and an inventory utilization rate of 62.5%. Next, the remaining capacity of multiple warehouses is calculated. The specific execution action is: Remaining capacity = Total warehouse capacity (1. Inventory utilization rate). Based on this, the remaining capacity of warehouse W1 = Boxes. Remaining capacity of warehouse W2 = Boxes. Then, the calculated remaining capacity of multiple warehouses is compared warehouse-by-warehouse with the corresponding warehouse's net inventory change obtained in S301 to generate the warehouse capacity change. This comparison is a subtraction operation, substituting the results from the previous steps: the net inventory change for warehouse W1 is 891, and the net inventory change for warehouse W3 is 859.375. Warehouse capacity change for warehouse W1 = The result indicates that, under the current net outflow trend, the remaining capacity of W1 is expanding. The change in the warehouse capacity of warehouse W3 = Boxes. This result indicates that under the current net inflow trend, the remaining capacity of W3 is shrinking, but there is still a buffer space of 640.625 boxes. The warehouse's capacity to accommodate changes quantifies the expected remaining storage capacity of the warehouse after the current inventory change trend continues for one unit of time.

[0083] S303: Based on the changes in warehouse capacity and net inventory, perform hourly differential calculations on warehouse feedback data under multiple time stamps, compare the differential residual of each time stamp with a preset residual threshold, extract the residual information of all time stamps, and obtain the residual time series.

[0084] Based on the generated warehouse capacity variation and the generated net inventory variation, hourly differential calculations are performed on the warehouse feedback data under multiple timestamps. Taking the destination warehouse W3 of the abnormal route as an example, this process collects feedback data at the current time t and the two hours before (t1, t2), and the specific data is presented in the table below.

[0085]

[0086] Table 1 shows the operational feedback data for warehouse W3 over three consecutive hours. Next, the calculation process of S301 is repeated for each timestamp to obtain the net inventory change for each timestamp. Timestamp t2: Net inventory change = Timestamp t1: Net inventory change = Timestamp t: Net inventory change (introduced from S301) = 859.375. Then, perform hourly differential operations on the net inventory changes of adjacent timestamps to calculate the differential residual. The differential residual is calculated by subtracting the net inventory change of the previous timestamp from the current timestamp's net inventory change. The differential residual for timestamp t1 = The difference residual of timestamp t = Subsequently, the differential residuals calculated for each timestamp are compared with a preset residual threshold. The residual threshold is set based on data analysis of 1000 stable operating hours. This data shows that, in the absence of abnormal events, 99% of the hourly month-on-month fluctuations in net inventory changes have an absolute value below 30. When the fluctuation exceeds 30, it is related to significant changes in operational rhythm, such as concentrated vehicle arrivals and backlogs of outbound orders. Therefore, the residual threshold is set to 30. The comparison process involves comparing the absolute value of the differential residuals with 30. The differential residual for timestamp t1 is 20.33, which is less than 30. The differential residual for timestamp t is 30.345, which is greater than 30. Finally, the differential residual information for all timestamps is extracted to form a residual time series. This series records the fluctuation at each time point. In this embodiment, the obtained residual time series is {20.33, 30.345}.

[0087] Please see Figure 5 The specific steps of S4 are as follows:

[0088] S401: Based on the residual time series, the residual values ​​of multiple time periods are accumulated and summed, and the residuals of multiple time periods under the same warehouse are summarized and arranged to generate a warehouse residual series;

[0089] Based on the residual time series {20.33, 30.345} obtained from the dataset, the residual values ​​across these multiple time periods are summed. Specifically, this calculation involves adding all residual values ​​from the beginning of the time series. In this embodiment, the series contains residuals at two time points, and the summation result is the cumulative residual at time point t. The calculation process is: Cumulative Residual = The cumulative residual value of 50.675 quantifies the cumulative deviation of the net inventory change of warehouse W3 from time point t-2 to t from the steady state. Next, the residuals of multiple time periods under the same warehouse are summarized and arranged. The summary arrangement here refers to organizing the residual time series data belonging to the same warehouse identifier and constructing a data structure indexed by warehouse. Since only line P4 was identified as an anomaly in the process from S203 to S303, and the correlation analysis is focused on warehouse W3, there is currently only the residual time series of warehouse W3. This arrangement process generates a mapping where the key is the warehouse identifier "W3" and the value is its corresponding residual time series array {20.33, 30.345}. If in the aforementioned steps, another warehouse, such as W5, is also identified and its residual time series is calculated to be {15.2, 18.5}, then the summarized and arranged data structure will be W3: {20.33, 30.345}, W5: {15.2, 18.5}. After this step, the residual sequence generated for warehouse W3 is W3: {20.33, 30.345}.

[0090] S402: Call the residual sequence of the warehouse, extract the sign of the direction of change of residual value for multiple time periods, sort and judge the residual difference values ​​of adjacent time periods, if the signs are the same, it is determined that the direction is the same and recorded, thus obtaining the residual direction sequence;

[0091] The generated warehouse residual sequence W3, {20.33, 30.345}, is used to extract signs for the changes in residual values ​​across multiple time periods. This extraction process involves comparing each residual value in the sequence with zero and assigning a sign based on the comparison result. The quantification standard is: +1 for a residual value greater than zero, -1 for a residual value less than zero, and 0 for a residual value equal to zero. This standard is established through analysis of inventory inflow and outflow pressure in historical operational data, where positive values ​​represent net inflow pressure and negative values ​​represent net outflow pressure. In this embodiment, the extraction process for the sequence {20.33, 30.345} is as follows: the first residual value, 20.33, is greater than zero, so the sign is extracted as +1. The second residual value, 30.345, is also greater than zero, so the sign is extracted as +1. This results in a sign sequence {+1, +1}. Next, the residual difference values ​​of adjacent time periods are sorted and judged. The judgment here refers to checking whether the signs of all elements in the symbol sequence generated in the previous step are consistent. Specifically, it iterates through the symbol sequence {+1, +1}, comparing whether all elements are equal. In this example, all elements in the sequence are +1, so the signs are considered consistent. If the signs are consistent, the directions are considered the same and recorded. The recorded results constitute the residual direction sequence. For warehouse W3, its residual direction sequence is {+1, +1}. The results show that, within the observed continuous time period, the net inventory change fluctuation of warehouse W3 always moves in the same direction (the direction of increasing net inflow pressure).

[0092] S403: Based on the residual direction sequence, perform monotonicity determination on the difference results of the residual values. If the residual directions are the same and the difference results are increasing, mark the corresponding warehouse index position as an anomaly, and aggregate the time series trajectories of all anomaly warehouses to generate a set of residual trend anomaly trajectories.

[0093] Based on the obtained residual direction sequence W3: {+1, +1} and the warehouse residual sequence W3: {20.33, 30.345}, a monotonicity determination is performed on the difference results of the residual values. The determination involves two consecutive actions. The first action checks whether the residual directions are the same. According to the result of S402, all elements in the residual direction sequence {+1, +1} are +1, therefore the directions are determined to be the same. The second action checks whether the residual values ​​themselves are increasing. This check specifically compares the magnitudes of two adjacent elements in the warehouse residual sequence {20.33, 30.345}. The residual value 30.345 at time point t is compared with the residual value 20.33 at time point t-1. Since... The difference is The value is greater than zero, therefore the sequence is determined to be increasing. When both conditions of "same residual direction" and "increasing difference result" are met simultaneously, the index position of the corresponding warehouse is marked as an anomaly. In this embodiment, warehouse W3 meets both conditions, so warehouse W3 is marked as an anomaly. Subsequently, the time series trajectories of all warehouses marked as anomalies are aggregated. Aggregation here refers to collecting and combining the complete residual time series of all warehouses marked as anomalies into a set. Since only warehouse W3 is marked as an anomaly in this example, the result of aggregation is to extract the residual time series {20.33, 30.345} of warehouse W3. The generated set of residual trend anomaly trajectories is W3: {t-1: 20.33, t: 30.345}.

[0094] Please see Figure 6 The specific steps of S5 are as follows:

[0095] S501: Based on the residual trend anomaly trajectory set, retrieve the residual trajectory corresponding to all abnormal warehouses, and perform an accumulation operation on the residual trajectory of the same warehouse in the time dimension to generate an abnormal warehouse residual accumulation sequence.

[0096] Based on the generated set of residual trend anomaly trajectories, subsequent processing is performed. To fully demonstrate the implementation process, the example scenario is expanded here. It is assumed that in the previous steps, in addition to warehouse W3, warehouse W2 also exhibits an anomaly trend because it is connected to another transportation route P6, which is not detailed in S203 but also shows anomalies. A set of residual trend anomaly trajectories containing the two warehouses is generated, specifically W3: {t-1: 20.33, t: 30.345}, W2: {t-1: 18.0, t: 25.0}. First, the residual trajectories corresponding to all warehouses marked as anomalies are retrieved from this trajectory set. Specifically, the set is traversed, and each key-value pair is extracted, where the key is the warehouse identifier and the value is the corresponding residual time series. In this embodiment, the retrieved content is the residual trajectory {20.33, 30.345} of warehouse W3 and the residual trajectory {18.0, 25.0} of warehouse W2. Next, the residual trajectories of the same warehouse are summed over time. Specifically, this operation involves summing all values ​​in the residual time series for a single warehouse. This calculation is performed independently for each retrieved abnormal warehouse trajectory. For warehouse W3, the summation process is as follows: the residual value of 20.33 at time point t-1 is added to the residual value of 30.345 at time point t. For warehouse W2, the cumulative calculation process is as follows: add the residual value of 18.0 at time point t-1 to the residual value of 25.0 at time point t. Finally, the cumulative results of each abnormal warehouse are summarized to generate an abnormal warehouse residual cumulative sequence. The sequence is a data structure indexed by warehouse identifiers and storing the corresponding residual cumulative values. In this embodiment, the generated abnormal warehouse residual cumulative sequence is {W3: 50.675, W2: 43.0}. The sequence clearly quantifies the cumulative deviation of each abnormal warehouse, providing direct data input for subsequent global risk assessment.

[0097] S502: Call the abnormal warehouse residual accumulation sequence to count the number of abnormal warehouses and calculate the ratio with the total number of warehouses. If the ratio is greater than half the threshold, record the abnormal warehouse set that meets the condition to obtain the abnormal warehouse set.

[0098] The generated residual accumulation sequence of abnormal warehouses {W3: 50.675, W2: 43.0} is invoked. First, the number of abnormal warehouses contained in this sequence is counted. Specifically, the data structure is traversed, and the number of keys (i.e., warehouse identifiers) contained therein is accumulated. In this embodiment, the sequence contains two keys, "W3" and "W2", so the count result is 2. Next, the total number of warehouses in the system is obtained, and the ratio is calculated with the aforementioned number of abnormal warehouses. According to the setting in S101, the total number of valid warehouses in the system is 3 (W1, W2, W3). The ratio calculation is performed by dividing the number of abnormal warehouses by the total number of warehouses. The calculated ratio is: Subsequently, this ratio is compared with a preset half threshold. The half threshold is set based on historical data analysis of the stability of the warehouse network system, combined with a fixed scaling factor of 0.5. The selection of this scaling factor of 0.5 is based on a retrospective analysis of 500 outage events in 20 warehouse networks of different sizes over a one-year operating cycle.

[0099]

[0100] As shown in Table 2, when the proportion of abnormal warehouse nodes to the total number of nodes is less than 50%, the network can recover on its own through local adjustments, and the occurrence rate of systemic outages is low (below 10%). When this proportion exceeds 50%, the occurrence rate of systemic outages rises sharply to 76.5%, indicating that the network has entered a globally unstable state. Therefore, the half threshold is set to 0.5. The judgment action here is to compare the calculated ratio of 0.667 with the half threshold of 0.5. The judgment process is as follows: Since the calculated ratio is greater than half the threshold, the judgment condition is met. If the judgment is met, the set of all warehouses currently identified as abnormal is recorded. This recording process involves extracting the identifiers of all abnormal warehouses involved in the calculation in S501 and forming a set. In this embodiment, the extracted warehouse identifiers are "W2" and "W3". The obtained abnormal warehouse set is {W2, W3}. This result indicates that the abnormal state of the current warehouse network has spread to most nodes, constituting a systemic risk.

[0101] S503: Based on the abnormal warehouse set, retrieve the connection relationship between multiple abnormal warehouses and abnormal transportation routes, and make a logical judgment on the index position of the warehouses with connections. If the judgment is valid, mark the mapping between the associated warehouses and routes, and establish the warehouse network simulation verification results.

[0102] The half threshold is set by statistically counting the total number of warehouses and multiplying it by a fixed proportion factor of 0.5.

[0103] Based on the abnormal warehouse set {W2, W3} obtained from the previous step, the connection relationships between multiple abnormal warehouses and abnormal transportation routes are first retrieved. The abnormal transportation routes here originate from the output of S203. To maintain consistency with the aforementioned extended example scenario, it is assumed that the output of S203 is the abnormal route group {P4, P5}. The retrieval of connection relationships is completed by querying the warehouse network topology data defined in S101. The query results are as follows: Abnormal warehouse W2: connected to transportation route P5 (W2W3 segment). Abnormal warehouse W3: connected to both transportation routes P4 (W1W3 segment) and P5 (W2W3 segment). Abnormal route P4: connects warehouses W1 and W3. Abnormal route P5: connects warehouses W2 and W3. Next, logical judgments are performed on the index positions of the connected warehouses. The specific execution action of the judgment is to check whether the two endpoint warehouses connected to each abnormal transportation route simultaneously exist in the abnormal warehouse set. For abnormal route P4 (connecting W1 and W3): endpoints are W1 and W3. Compare these two endpoints with the abnormal warehouse set {W2, W3}. W3 exists in the set, but W1 does not. For abnormal route P5 (connecting W2 and W3): endpoints are W2 and W3. Compare these two endpoints with the abnormal warehouse set {W2, W3}. Both W2 and W3 exist in the set. If the determination is valid, the associated warehouse-route mapping is marked. The criterion for "valid determination" here is: when at least one endpoint warehouse of an abnormal transportation route also belongs to the abnormal warehouse set, the mapping relationship between the route and the corresponding abnormal warehouse is determined to be valid. For route P4: endpoint W3 is an abnormal warehouse, therefore the mapping relationship "connection between P4 and W3" is determined to be valid. For route P5: endpoints W2 and W3 are both abnormal warehouses, therefore both the mapping relationships "connection between P5 and W2" and "connection between P5 and W3" are determined to be valid. Finally, integrate all marked mapping relationships to establish the warehouse network simulation verification results. The result is a structured dataset that clearly identifies the core fault subgraph constituting the current systemic risk of the warehouse network. In this embodiment, the simulation verification result of the established warehouse network is as follows: the set of marked associations: {(P4, W3), (P5, W2), (P5, W3)}; the set of core abnormal nodes: {W2, W3}; and the set of core abnormal paths: {P4, P5}. This result indicates that warehouses W2 and W3, as well as the line P5 connecting them, plus the line P4 extending from W3, together constitute the core area of ​​the problem.

[0104] Please see Figure 7 A warehouse network simulation and verification system based on a large model includes:

[0105] The structural deduction module obtains the number of warehouses, coordinates, transportation routes, and cargo requirements. It then inputs these data into a graph neural network to perform warehouse network spatial structure deduction and logistics flow direction learning. Finally, it constructs the warehouse network planning results and transmits them to the error analysis module.

[0106] The error analysis module obtains the feedback cargo inflow, outflow and transit time of the transportation route, calculates the net cargo flow difference and the feedback in-transit inventory reference value, performs error analysis with the warehouse network planning results and determines whether the difference threshold is exceeded, generates abnormal route groups and transmits them to the residual generation module.

[0107] The residual generation module obtains the feedback cargo inbound and outbound volume and inventory utilization rate of the warehouses associated with the abnormal route group, calculates the net inventory change and warehouse capacity change, analyzes the warehouse operation feedback residuals under multiple timestamps, generates residual time series and transmits them to the trend recognition module.

[0108] The trend recognition module, based on the residual time series, summarizes the residuals by time period for each warehouse and analyzes the residual change trend. If the residuals are in the same direction and the residual change trend is increasing, the warehouse is marked as abnormal, and a set of residual trend abnormal trajectory is generated and transmitted to the simulation verification module.

[0109] The simulation verification module acquires and accumulates the residual trajectories of all abnormal warehouses in the residual trend abnormal trajectory set. When the number of abnormal warehouses exceeds half the threshold and there is a connection with abnormal transportation routes, it is determined that there is a deviation in the planning result, and the warehouse network simulation verification result is generated.

[0110] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A large model-based warehouse network planning method, characterized in that, Comprise the following steps: S1: Obtain the warehouse quantity, coordinate and transportation line and goods demand, input the graph neural network to deduce the warehouse network space structure and learn the logistics direction, and construct the warehouse network planning result; The specific steps of S1 are: S101: Obtain the warehouse quantity, warehouse coordinate and transportation line parameter, and weight and accumulate the warehouse quantity and the length of the multiple path in the transportation line, then compare the accumulated value with the goods demand quantity, screen the path set that meets the goods demand quantity, and generate the path matching set; S102: Based on the path matching set, construct the corresponding graph structure data frame, vectorize the warehouse coordinate and goods demand quantity, and perform splicing operation with the graph structure data frame to obtain the warehouse network input feature set; S103: According to the warehouse network input feature set, execute layer-by-layer propagation calculation in the graph neural network, and normalize and compare the state value output by multiple propagation layers and the node demand quantity, screen the stable convergence propagation output result, and obtain the warehouse network planning result; S2: Obtain the feedback goods inflow and outflow and in-transit time of the transportation line, calculate the net goods flow difference and feedback in-transit stock reference value, perform error analysis on the warehouse network planning result and judge whether it exceeds the difference threshold, and generate the abnormal line group; The specific steps of S2 are: S201: Obtain the feedback goods inflow, feedback goods outflow and in-transit time parameter of the transportation line, perform difference calculation on the inflow and outflow, and then weight the difference value with the in-transit time parameter to generate the net goods flow difference; S202: Based on the net goods flow difference, call the in-transit time parameter and feedback goods inflow to perform product operation, and compare the result with the outflow to calculate the feedback in-transit stock reference value; S203: According to the feedback in-transit stock reference value and the warehouse network planning result, compare them item by item, judge the comparison value with the preset difference threshold, and screen the line that exceeds the difference threshold to obtain the abnormal line group; S3: Obtain the feedback goods in and out of the warehouse and the inventory utilization rate of the warehouse associated with the abnormal line group, calculate the net inventory change and the warehouse capacity change, and analyze the warehouse operation feedback residual error under multiple time stamps to generate the residual time sequence; S4: Based on the residual time sequence, aggregate the residual error by single warehouse and time period, and analyze the residual error change trend, if the residual error direction is the same and the residual error change trend is increasing, mark the warehouse as abnormal, generate the residual trend abnormal trajectory set; S5: Obtain the residual trajectory of all abnormal warehouses in the residual trend abnormal trajectory set and accumulate, if the number of abnormal warehouses exceeds the half threshold and there is a connection relationship with the abnormal transportation line, it is judged that the planning result has deviation, and the warehouse network simulation verification result is generated.

2. The warehouse planning method based on a large model according to claim 1, characterized in that, The warehouse network planning result includes the number of warehouses, warehouse coordinates, and transportation routes, the abnormal route group specifically refers to net cargo flow difference, feedback in-transit inventory reference value, and difference threshold, the residual time sequence includes net inventory change, warehouse accommodation change, and warehouse operation feedback residual, the residual trend abnormal trajectory set specifically refers to residual direction, residual change trend, and abnormal warehouse marker, and the warehouse network simulation verification result specifically refers to abnormal warehouse number, abnormal warehouse residual trajectory cumulative value, and abnormal transportation route connection relationship.

3. The warehouse planning method based on a large model according to claim 1, characterized in that, The difference threshold is obtained by statistically sampling the feedback cargo inflow, feedback cargo outflow and in-transit time parameters of the transportation route, calculating the mean and variance of the net cargo flow difference of multiple routes, and then linearly weighting the mean and variance results.

4. The warehouse planning method based on a large model according to claim 1, characterized in that, The specific steps of S3 are: S301: Obtain the feedback cargo storage quantity, feedback cargo out-of-storage quantity and inventory utilization rate data of the warehouse associated with the abnormal route group, perform difference calculation on the storage quantity and the out-of-storage quantity, weight the difference result and the inventory utilization rate parameter, and generate the net inventory change; S302: Based on the net inventory change, obtain the accommodation parameter and inventory utilization rate data of the warehouse, calculate the remaining accommodation of multiple warehouses, and perform warehouse-by-warehouse comparison between the calculation result and the net inventory change to generate the warehouse accommodation change; S303: According to the warehouse accommodation change and the net inventory change, perform hour-by-hour difference operation on the warehouse feedback data under multiple time stamps, compare the difference residual of each time stamp with a preset residual threshold, extract the residual information of all time stamps, and obtain the residual time sequence.

5. The large model-based warehouse planning method of claim 4, wherein, The residual threshold is obtained by statistically sampling the difference results of the net inventory change and the warehouse accommodation change in the warehouse feedback data under multiple time stamps, calculating the mean and variance of the residual value, and then nonlinearly weighting the mean and variance.

6. The warehouse planning method based on a large model according to claim 1, characterized in that, The specific steps of S4 are: S401: Based on the residual time sequence, accumulate and sum the residual values of multiple time periods, and arrange the residuals of multiple time periods for the same warehouse, to generate a warehouse-specific residual sequence; S402: Call the warehouse-specific residual sequence, extract the sign of the residual value change direction for multiple time periods, and sort and judge the residual difference values of adjacent time periods, if the signs are consistent, it is determined that the directions are the same and recorded, to obtain a residual direction sequence; S403: According to the residual direction sequence, perform monotonicity judgment on the difference result of the residual value, if the residual directions are the same and the difference result is in an increasing relationship, mark the corresponding warehouse index position as abnormal, and aggregate the time sequence trajectories of all abnormal warehouses to generate a residual trend abnormal trajectory set.

7. The warehouse planning method based on a large model according to claim 1, characterized in that, The specific steps of S5 are: S501: Based on the residual trend abnormal trajectory set, retrieve the residual trajectories of all abnormal warehouses, and perform accumulation operation on the residual trajectories of the same warehouse in the time dimension to generate an abnormal warehouse residual accumulation sequence; S502: Call the exception warehouse residual accumulation sequence, count the number of abnormal warehouses, and calculate the ratio with the total number of warehouses. If the ratio is greater than half the threshold value, record the abnormal warehouse set that meets the conditions to obtain the abnormal warehouse set; S503: According to the abnormal warehouse set, search for the connection relationship between multiple abnormal warehouses and abnormal transportation lines, and logically judge the index position of the warehouse that exists. If the judgment is correct, mark the associated warehouse and line mapping, and establish the warehouse network simulation verification result; The half threshold is set by multiplying the total number of warehouses by the fixed proportion factor 0.

5.

8. A large model-based warehouse network simulation verification system, characterized in that, The system is used to implement the warehouse network planning method based on a large model according to any one of claims 1-7, and the system comprises: A structure deduction module obtains the number of warehouses, coordinates, transportation lines, and cargo demand, inputs a graph neural network to deduce the warehouse network space structure and learn logistics flow, constructs a warehouse network planning result, and transmits it to the error analysis module; An error analysis module obtains the feedback cargo inflow, outflow, and in-transit time of the transportation line, calculates the net cargo flow difference and feedback in-transit inventory reference value, analyzes the error of the warehouse network planning result, and judges whether it exceeds the difference threshold value, generates an abnormal line group, and transmits it to the residual generation module; A residual generation module obtains the feedback cargo in-out warehouse quantity and inventory utilization rate of the associated warehouse of the abnormal line group, calculates the net inventory change and warehouse capacity change, and analyzes the warehouse operation feedback residual under multiple time stamps, generates a residual time sequence, and transmits it to the trend identification module; A trend identification module based on the residual time sequence, aggregates residuals by single warehouse and time period, and analyzes the residual change trend. If the residual direction is the same and the residual change trend is increasing, mark the warehouse as abnormal, generate a residual trend abnormal trajectory set, and transmit it to the simulation verification module; A simulation verification module obtains the residual trajectory of all abnormal warehouses in the residual trend abnormal trajectory set and accumulates it. If the number of abnormal warehouses exceeds half the threshold value and there is a connection relationship with the abnormal transportation line, it is judged that the planning result has deviation, and a warehouse network simulation verification result is generated.

Images