Machine Learning-Based Shipping Warehouse Demand Prediction Method and System

Abstract

The present invention provides a method and system for predicting shipping warehousing requirements based on machine learning. By collecting the dynamic dataset of shipping logistics at the target port, through multi-modal feature analysis, joint feature extraction is performed on the port throughput time series, container transportation map, and warehousing equipment status log to generate a composite feature vector containing ship scheduling time series, cargo turnover cycle, and warehousing capacity fluctuation. Then, the composite feature vector is subjected to spatio-temporal coding processing. The time dependence feature of ship arrival time is extracted using a recurrent convolutional network, and the correlation of cargo flow transfer across warehousing areas is captured by combining the graph attention mechanism. After generating a multi-dimensional spatio-temporal coding matrix, it is input into a pre-trained basic shipping warehousing requirement prediction model, and preliminary warehousing requirement distribution data including the expected cargo accumulation volume and storage cycle probability curve of each warehousing area is output, and the target warehousing requirement prediction result is corrected to achieve more accurate and detailed prediction of shipping warehousing requirements, which helps to reasonably plan warehousing resources.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning. Specifically, it relates to a method and system for predicting shipping warehousing demand based on machine learning. Background Art

[0002] Under the background of the increasingly prosperous global trade today, shipping, as one of the most important transportation methods in international trade, its related warehousing management is crucial for the efficiency and economy of port operations. Accurately predicting shipping warehousing demand can help ports reasonably plan warehousing space, arrange human and material resources, thereby reducing operating costs and improving service quality.

[0003] Most traditional methods for predicting shipping warehousing demand are based on simple statistical models and empirical rules. These methods often only consider data from a single dimension. For example, they only predict future warehousing demand based on historical port throughput, or only focus on the approximate cycle of goods turnover, while ignoring other key factors such as ship scheduling and the status of warehousing equipment. This way of considering data from a single dimension cannot comprehensively reflect the complexity and dynamics of the shipping warehousing system, resulting in poor accuracy of prediction results and being difficult to meet the actual operating needs of ports.

[0004] With the development of data technology, some existing technologies have begun to attempt to use multi-source data for prediction, but these methods only simply splice or aggregate data from different sources, and fail to deeply explore the internal connections and synergistic effects between different types of data. For example, although both port throughput and container transportation data are obtained, the association between them in terms of time series and spatial distribution is not essentially analyzed, so that the multi-source data fails to exert its due advantages and the improvement of prediction effect is limited. Summary of the Invention

[0005] In view of the above-mentioned problems, in combination with the first aspect of the present invention, embodiments of the present invention provide a method for predicting shipping warehousing demand based on machine learning. The method includes:

[0006] Collect a dynamic dataset of shipping logistics in the target port, perform multi-modal feature analysis on the dynamic dataset of shipping logistics, and generate a composite feature vector including ship scheduling time series, goods turnover cycle, and warehousing capacity fluctuation; the multi-modal feature analysis includes joint feature extraction of port throughput time series, container transportation atlas, and warehousing equipment status log;

[0007] Perform spatio-temporal encoding processing on the composite feature vector to generate a multi-dimensional spatio-temporal encoding matrix; the spatio-temporal encoding processing includes using a recurrent convolutional network to extract time-dependent features of ship arrival time, and combining graph attention mechanism to capture the correlation of goods transfer across warehousing areas;

[0008] Input the multi-dimensional spatio-temporal encoding matrix into a pre-trained basic shipping warehouse demand prediction model to output preliminary warehouse demand distribution data; the preliminary warehouse demand distribution data includes the expected cargo accumulation volume and storage cycle probability curve of each warehouse area;

[0009] Correct the preliminary warehouse demand distribution data to generate a target warehouse demand prediction result.

[0010] On the other hand, an embodiment of the present invention also provides a shipping warehouse demand prediction system based on machine learning, including a processor and a machine-readable storage medium. The machine-readable storage medium is connected to the processor. The machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0011] Based on the above aspects, the embodiments of the present application collect the shipping logistics dynamic data set of the target port and perform multi-modal feature analysis to generate a composite feature vector, breaking through the limitations of traditional single data sources. By jointly extracting features from the port throughput time series, container transportation map and warehouse equipment status log, it not only comprehensively covers the key information in the shipping warehouse system, but also can capture the complex internal relationships between different modal data, generating a composite feature vector including ship scheduling time series, cargo turnover cycle and warehouse capacity fluctuation, providing rich and accurate basic data for subsequent prediction and greatly improving the information value of the data.

[0012] In the feature encoding link, perform spatio-temporal encoding processing on the composite feature vector. Using a recurrent convolutional network to extract the time dependence features of ship arrival time can accurately grasp the dynamic change rules in the ship arrival time series, and combining with the graph attention mechanism to capture the cargo flow correlation across warehouse areas, enabling the model to automatically focus on important warehouse area associations and adaptively allocate attention weights, thus accurately capturing the complex patterns of cargo flow between different warehouse areas. The multi-dimensional spatio-temporal encoding matrix generated by this spatio-temporal encoding processing method can effectively integrate the information in the time and space dimensions, providing a more comprehensive and in-depth feature representation for warehouse demand prediction and enhancing the model's modeling ability for complex real-world scenarios, while traditional methods often have difficulty effectively dealing with the complex cargo flow relationships across regions.

[0013] Input the multi-dimensional space-time coding matrix into the pre-trained basic shipping warehouse demand prediction model. The output preliminary warehouse demand distribution data includes the expected cargo accumulation volume and storage cycle probability curve of each warehouse area. Compared with the general data provided by traditional prediction methods, it provides more detailed and accurate information for port operators. Operators can plan the allocation of warehouse resources in advance based on this data. For example, reasonably arrange the cargo storage plans of different warehouse areas, optimize the warehouse layout, improve the utilization rate of warehouse space, and reduce cargo backlog and handling costs.

[0014] Finally, correct the preliminary warehouse demand distribution data to generate the target warehouse demand prediction result, further improving the accuracy and reliability of the prediction. By considering more complex factors and uncertainties in actual scenarios, optimize and adjust the preliminary results, so that the final prediction result can better reflect the actual situation, which helps the port to make more flexible and efficient decisions, reasonably allocate resources, improve the overall operation efficiency, and reduce operation risks when facing the changing shipping logistics demands. Brief Description of the Drawings

[0015] Figure 1 It is a schematic execution flowchart of a shipping warehouse demand prediction method based on machine learning provided by an embodiment of the present invention.

[0016] Figure 2 It is a schematic diagram of the hardware architecture of a shipping warehouse demand prediction system based on machine learning provided by an embodiment of the present invention. Detailed Embodiments

[0017] The present invention will be specifically described below in conjunction with the accompanying drawings of the specification. Figure 1 It is a schematic flowchart of a shipping warehouse demand prediction method based on machine learning provided by an embodiment of the present invention. The shipping warehouse demand prediction method based on machine learning will be introduced in detail below.

[0018] Step S110, collect the dynamic dataset of shipping logistics of the target port, perform multi-modal feature analysis on the dynamic dataset of shipping logistics, and generate a composite feature vector including ship scheduling time series, cargo turnover cycle, and warehouse capacity fluctuation. The multi-modal feature analysis includes joint feature extraction of port throughput time series, container transportation map, and warehouse equipment status log.

[0019] Specifically, the target port is the specific port targeted by the entire shipping warehouse demand prediction process. This port is the object of data collection, analysis, and prediction, and has its own unique operating characteristics, infrastructure, cargo types, and traffic conditions.

[0020] The described shipping logistics dynamic dataset covers a data set of dynamic information in all aspects of the shipping logistics process at the target port. This information includes, but is not limited to, the situation of ships entering and leaving the port, the situation of cargo loading, unloading and transportation, the operating status of warehousing equipment, etc., and this data is continuously updated and changed over time, reflecting the real-time and historical dynamic situation of port shipping logistics.

[0021] Therefore, the multi-modal feature solution aims to extract meaningful features from multiple different types (modalities) of data sources. In this scenario, it is to comprehensively understand the situation of shipping logistics from different perspectives, and by exploring the internal connections between different modal data, richer and more accurate information can be obtained.

[0022] The described composite feature vector is a data structure that integrates feature information related to ship scheduling time series, cargo turnover cycle, and warehousing capacity fluctuation obtained from multi-modal feature analysis. This composite feature vector is the basis for subsequent analysis and prediction, combining multiple related features together for unified processing and analysis.

[0023] The described port throughput time series is a data sequence that records the quantity of goods handled at the port in chronological order. Port throughput is an important indicator to measure the operation scale and efficiency of the port. This time series can reflect the flow change rules of goods entering and leaving the port in different time periods, such as the peaks and valleys of cargo flow in different seasons, weekdays and non-weekdays, etc.

[0024] The described container transportation map is a data structure that details the transportation information of each container. It contains various information in the entire transportation process of the container from the departure port to the destination port, such as the residence time at each terminal, whether it undergoes special inspections, the transfer route, etc., which helps to comprehensively understand the flow of containerized goods and is of great significance for analyzing the cargo turnover cycle, etc.

[0025] The described warehousing equipment status log is a log that records the change of the operating status of warehousing equipment in the port over time. These warehousing equipment include shelves, stackers, conveyors, etc. in the automated stereoscopic warehouse. The log content covers information such as the operating speed, load-bearing weight, and conveying efficiency of the equipment, which can reflect the working status of the warehousing equipment and its impact on cargo storage and turnover, and thus reflect the situation of warehousing capacity fluctuation, etc.

[0026] Joint feature extraction is a process of using specific algorithms to simultaneously extract features related to ship scheduling time series, cargo turnover cycle, and warehousing capacity fluctuation from three different types of data: port throughput time series, container transportation map, and warehousing equipment status log. By jointly extracting features, the correlation relationships between different data can be explored, so as to more comprehensively describe the overall situation of shipping logistics.

[0027] In this embodiment, consider a large port, such as Port AA. During daily operations, there are numerous data sources for constructing a dynamic dataset of shipping logistics. For the port throughput time series, precise metering devices are installed at each import and export channel of the port. These devices can record in real time the amount of goods carried by each ship when entering and leaving the port. For example, different types of ships enter and leave the port every day, such as large oil tankers, container ships, etc. The cargo volume of an oil tanker may be measured in tons, while that of a container ship is measured in twenty-foot equivalent units (TEUs). Over a one-month time span, the port management department can collect a vast amount of throughput data, and when these data are arranged in chronological order, they form the port throughput time series.

[0028] The container transportation map details the transportation information of each container. At the container terminal, each container has a unique identification code. When a container is unloaded from a ship, then transferred to the terminal yard, and then loaded onto another ship or transported to an inland warehouse, every link in this process is recorded. This includes information such as its port of departure, port of destination, stay time at the terminal, and whether it has undergone special inspections. The aggregation of this information constitutes the container transportation map.

[0029] The status log of warehousing equipment covers the equipment conditions in each warehousing area within the port. Taking an automated storage and retrieval system as an example, devices such as shelves, stacker cranes, and conveyors are equipped with sensors. The sensors can monitor the operating status of the equipment in real time, such as the operating speed of the stacker crane, the load-bearing capacity of the shelves, and the conveying efficiency of the conveyor. When a cargo needs to be stored, first, the conveyor transports the cargo to the designated shelf area, and then the stacker crane stores the cargo in a suitable position. During this process, if the conveying speed of the conveyor suddenly drops, it may affect the cargo turnover cycle; if the load-bearing capacity of the shelves approaches the upper limit, it will reflect the fluctuations in warehousing capacity.

[0030] Joint feature extraction is performed on the port throughput time series, container transportation map, and warehousing equipment status log through a specific algorithm. For the port throughput time series, the peaks and troughs of cargo flow in different time periods can be analyzed, which helps to determine the ship scheduling sequence. For example, if the throughput suddenly increases in a certain period, it may mean that multiple ships arrive at the port simultaneously, and it is necessary to reasonably arrange the docking and loading / unloading sequence of the ships. From the container transportation map, the residence time of goods in different transportation links can be calculated to determine the cargo turnover cycle. The equipment load situation in the warehousing equipment status log can reflect the fluctuation of warehousing capacity. By integrating this information, a composite feature vector containing the ship scheduling sequence, cargo turnover cycle, and warehousing capacity fluctuation can be generated. For example, the composite feature vector may be represented as a multi-dimensional array, where the ship scheduling sequence is represented in the form of time intervals, the cargo turnover cycle is represented in days, and the warehousing capacity fluctuation is represented as a percentage of capacity utilization rate.

[0031] Step S120: Perform spatio-temporal encoding processing on the composite feature vector to generate a multi-dimensional spatio-temporal encoding matrix. The spatio-temporal encoding processing includes using a recurrent convolutional network to extract the time dependence features of ship arrival times and combining graph attention mechanisms to capture the cargo flow correlation across warehousing areas.

[0032] Specifically, the spatio-temporal encoding processing is an operation for processing the composite feature vector, aiming to encode the information in the feature vector according to the time and space dimensions, so as to better capture the spatio-temporal relationships in shipping logistics, which helps to better utilize the time dependence features and spatial correlation features in the subsequent prediction model.

[0033] The multi-dimensional spatio-temporal encoding matrix is a matrix data structure obtained after spatio-temporal encoding processing, which contains the shipping logistics feature information encoded in multiple dimensions. These dimensions may include time, warehousing area, cargo features, etc. It is a more comprehensive and structured representation of the composite feature vector in the spatio-temporal dimension, preparing for input into the pre-trained basic shipping and warehousing demand prediction model.

[0034] The recurrent convolutional network is a neural network structure used to extract the time dependence features of ship arrival times. It can process sequence data (such as ship arrival time series), and capture the time dependence relationships in the data through convolutional layers and recurrent structures (such as LSTM or GRU units). In this scenario, it can learn the long-term and short-term dependence relationships between ship arrival times, such as ships arriving at the port according to a fixed cycle or fluctuations in arrival times due to external factors.

[0035] The time-dependence characteristics of vessel arrivals refer to the characteristics reflecting the interrelationships among vessel arrival times, including long-term periodic patterns (such as the fixed arrival cycles of certain cargo ships) and short-term fluctuation factors (such as changes in arrival times caused by weather, shipping markets, etc.). These characteristics are of great significance for the rational arrangement of vessel scheduling and warehousing resources at ports.

[0036] The graph attention mechanism is a mechanism for processing graph-structured data. In this scenario, it is used to construct a graph structure with warehousing areas as nodes and cargo transfer paths as edges, and capture the cargo transfer correlations across warehousing areas. It can dynamically focus on the relationships between different nodes (warehousing areas) in the graph, adjust the attention degree to the relationships between regions according to the actual situation of cargo transfer, so as to better analyze the complex correlations of cargo transfer.

[0037] The cargo transfer correlation across warehousing areas describes the characteristics of the cargo transfer relationship between different warehousing areas. When cargo is transferred from one warehousing area to another within a port, there are complex correlation relationships in this transfer process. For example, the flow direction of the cargo, the transfer sequence, and the mutual influence of the cargo flow between different areas are affected by various factors such as cargo types, warehousing equipment layout, and market demand.

[0038] Continuing with Port AA as an example, after obtaining the composite feature vector, spatio-temporal encoding processing is performed. For the process of the recurrent convolutional network extracting the time-dependence characteristics of vessel arrivals, assume that the port has vessel arrival records for the past year. The arrival time of each vessel is a key time node. The recurrent convolutional network will use these arrival time series as input, and it can capture the dependence relationships between the arrival times of different vessels. For example, some large cargo ships may arrive at the port according to a fixed cycle, such as once every two months, which is a long-term time dependence. In the short term, due to factors such as weather and shipping markets, the arrival times of vessels may fluctuate. The recurrent convolutional network can learn these time-dependence characteristics. For example, in a specific season, due to the impact of the typhoon season, the arrival times of vessels will be relatively dispersed, avoiding congestion at the port caused by concentrated arrivals.

[0039] Meanwhile, the graph attention mechanism is combined to capture the correlation of cargo flow across storage areas. Port AA has multiple storage areas, each storing different types of goods or serving different shipping routes. When a cargo is unloaded from a ship, it will be transferred to a specific storage area. During this process, there are complex correlations in cargo flow. For example, some imported electronic products may need to be inspected first in a temporary storage area near the dock and then transferred to a dedicated electronic product storage area. The graph attention mechanism will construct a graph structure with storage areas as nodes and cargo flow paths as edges. Each storage area has its own attributes, such as area, storage capacity, and the degree of equipment advancement. Through the graph attention mechanism, the cargo flow relationship between different storage areas can be dynamically focused on. If there is a backlog of goods in a certain storage area, the graph attention mechanism can analyze which upstream storage areas have poor cargo transfer or which downstream storage areas have insufficient demand, resulting in the inability to transfer goods in a timely manner. In this way, the time-dependence feature of ship arrival time and the correlation of cargo flow across storage areas are comprehensively processed to generate a multi-dimensional spatio-temporal coding matrix. This multi-dimensional spatio-temporal coding matrix can be a three-dimensional matrix, where one dimension represents time, one dimension represents storage areas, and the other dimension represents the characteristic information of the goods.

[0040] Step S130, input the multi-dimensional spatio-temporal coding matrix into a pre-trained basic shipping warehouse demand prediction model, and output preliminary warehouse demand distribution data. The preliminary warehouse demand distribution data includes the expected cargo accumulation volume and storage cycle probability curve of each storage sub-area.

[0041] Specifically, the pre-trained basic shipping warehouse demand prediction model is a model pre-trained with a large amount of data for predicting shipping warehouse demand. It can be constructed based on a sample shipping warehouse operation data set and includes components such as a time feature extractor, a space feature extractor, and a fully connected prediction layer. It can predict the preliminary warehouse demand distribution data of each storage sub-area according to the input multi-dimensional spatio-temporal coding matrix.

[0042] The preliminary warehouse demand distribution data is the data output by the pre-trained basic shipping warehouse demand prediction model, including the expected cargo accumulation volume and storage cycle probability curve of each storage sub-area. The expected cargo accumulation volume represents the possible amount of goods that may accumulate in each storage sub-area in the future, and the storage cycle probability curve describes the probability of goods staying in the storage sub-area for different durations. These data are the preliminary warehouse demand prediction results.

[0043] The expected cargo accumulation volume of each warehousing zone is the quantity of goods that may accumulate within a specific future time period predicted by the basic shipping warehousing demand prediction model for each warehousing zone within the port. This quantity is expressed in a certain measurement unit (such as tons or standard TEU), reflecting the cargo storage demand situation of each warehousing zone, and is an important basis for warehousing planning and resource allocation.

[0044] The storage cycle probability curve is a curve representing the probability distribution of the residence time of goods in the warehousing zone. Due to factors such as market demand and cargo characteristics, the residence time of different types of goods in the warehousing zone is uncertain. The storage cycle probability curve can quantify this uncertainty. For example, for some seasonal goods, the storage cycle is shorter before the peak season and may be longer in the off-season. This curve helps warehousing operators reasonably arrange the storage and turnover of goods.

[0045] Taking a specific warehousing zone in Port AA as an example, the multi-dimensional spatio-temporal coding matrix is input into the pre-trained basic shipping warehousing demand prediction model. This model is constructed based on a large amount of historical data, including data on ship arrivals, cargo transfers, and warehousing equipment operations over the past years. Inside the model, for each warehousing zone, the model will perform complex calculations based on the input multi-dimensional spatio-temporal coding matrix.

[0046] Suppose a certain warehousing zone mainly stores textiles. According to the model's calculations, it will predict the expected cargo accumulation volume of this warehousing zone in the future. If there are a large number of textile import orders within a certain time period, and the ship arrival plan shows that multiple ships carrying textiles are about to arrive, the model will predict an increase in the cargo accumulation volume of this warehousing zone. This expected cargo accumulation volume may be in tons or standard TEU.

[0047] At the same time, the model will also output the storage cycle probability curve. For the textile warehousing zone, the storage cycle may be affected by various factors, such as seasonal fluctuations in market demand and the type of textiles (whether it is high-end fashion fabric or ordinary home textile fabric). If it is high-end fashion fabric, there may be a higher demand before the season alternation, and the storage cycle is relatively short; while the storage cycle of ordinary home textile fabric may be relatively long and less affected by market fluctuations. Through learning and analyzing historical data, the model can predict the probabilities under different storage cycles. For example, the probability of a storage cycle of 1 month is 20%, 2 months is 30%, 3 months is 40%, etc. These preliminary warehousing demand distribution data provide a basis for subsequent corrections.

[0048] Step S140: Correct the preliminary warehousing demand distribution data to generate the target warehousing demand prediction result.

[0049] In this embodiment, the target warehousing demand prediction result is the final warehousing demand prediction result obtained after correcting the preliminary warehousing demand distribution data. This result is more accurate and reliable than the preliminary result, taking into account more factors such as the demand impact factor of warehousing nodes and spatio-temporal explanatory feature traceability data, and can better reflect the actual warehousing demand situation, providing decision-making support for the warehousing management of the port.

[0050] For the correction process of the preliminary warehousing demand distribution data, the textile warehousing area of Port AA is taken as an example for illustration.

[0051] Based on the local sensitive feature domain and storage cycle probability curve of the multi-dimensional spatio-temporal coding matrix, calculate the demand impact factors of different warehousing nodes (here, the warehousing nodes can be understood as different storage areas or storage locations within the warehousing area). Extract the feature slices corresponding to each warehousing node from the local sensitive feature domain of the multi-dimensional spatio-temporal coding matrix. Suppose there are three different storage areas, A, B, and C, in the textile warehousing area. For storage area A, its corresponding feature slice contains information such as the recent maintenance situation of the equipment in this area and the goods transfer efficiency with adjacent areas. Calculate the local sensitive index of this feature slice. For example, if the equipment in this area has just undergone a comprehensive maintenance and the equipment operation efficiency is high, then the contribution degree of this feature slice to the prediction of the goods stacking volume may be relatively high, and the local sensitive index will be relatively large.

[0052] At the same time, calculate the goods retention risk scores of each warehousing node based on the storage cycle probability curve. For storage area A, if the storage cycle probability curve shows that the goods in this area have a high probability of being stored for more than 3 months, due to the fact that textiles may be affected by the warehousing environment (such as getting damp, changing color, etc.), the goods retention risk score will be relatively high. Then, normalize and weight the local sensitive index and the retention risk score to generate a comprehensive demand impact factor matrix. For example, the local sensitive index of storage area A is 0.6, and the retention risk score after normalization is 0.8. After weighting, the comprehensive demand impact factor is 0.7.

[0053] Then, screen out the target extractors with significant interpretability from multiple feature extractors of the basic shipping warehousing demand prediction model. Identify the interpretability indicators of each feature extractor in the basic shipping warehousing demand prediction model, and this interpretability indicator is obtained by calculating the backpropagation saliency map. Suppose there are multiple feature extractors in the model, such as a time feature extractor and a space feature extractor. If the interpretability indicator of the time feature extractor exceeds the preset threshold, it will be used as the target extractor. Extract the intermediate feature representations of the target extractor in the multi-layer network of the basic shipping warehousing demand prediction model to form a multi-granularity feature set. For example, the time feature extractor may extract the goods turnover features at different time scales at different layers, and summarize these features to form a multi-granularity feature set.

[0054] Hierarchically weight the multi-granularity feature set according to the demand impact factor to generate feature traceability data with spatio-temporal interpretability. For different features in the multi-granularity feature set, different weights are given according to the demand impact factor. If a certain feature is highly correlated with the risk of cargo detention, then its weight will be relatively large.

[0055] Construct a spatio-temporal attention distribution map reflecting the weights of the cargo flow pattern based on the feature traceability data. Input the feature traceability vector of the feature traceability data into the multi-head self-attention module to generate an initial attention weight matrix. For example, this initial attention weight matrix will represent the degree of attention to the cargo flow in different time and space dimensions. Decompose the initial attention weight matrix in the spatio-temporal dimension to obtain a time attention sub-matrix and a space attention sub-matrix. Use a gating mechanism to fuse these two sub-matrices to generate a spatio-temporal joint attention distribution. Then, sparsify the spatio-temporal joint attention distribution, retain the top K attention connections with weight values higher than the dynamic threshold, and construct a spatio-temporal attention distribution map reflecting the weights of the cargo flow pattern. This map may show that in certain specific time periods, to which storage areas the goods tend to flow from which areas.

[0056] Perform feature enhancement fusion on the spatio-temporal attention distribution map and the shipping cargo dynamic data set to generate an optimized input matrix. Convert the spatio-temporal attention distribution map into a feature enhancement mask matrix, standardize the original features in the shipping cargo dynamic data set to obtain a normalized feature tensor. For example, normalize the original cargo throughput data so that it is within a certain numerical range. Then perform a Hadamard product operation on the normalized feature tensor and the feature enhancement mask matrix to obtain a primary enhanced feature. Add the primary enhanced feature and the original feature through a residual connection to generate the final optimized input matrix.

[0057] Finally, correct the preliminary storage demand distribution data based on the optimized input matrix to generate the target storage demand prediction result. For example, for the expected cargo accumulation volume in the textile storage area, according to the information in the optimized input matrix, if it is found that the textile orders on a ship about to arrive at the port are cancelled, then the expected cargo accumulation volume will be reduced accordingly. At the same time, for the storage period probability curve, it will also be adjusted according to the new information. For example, due to the emergence of new competitors in the market, the sales speed of textiles may accelerate, and the storage period probability curve will be adjusted towards the short-period direction, so as to generate a more accurate target storage demand prediction result.

[0058] Based on the above steps, the embodiment of the present application collects the shipping logistics dynamic data set of the target port, performs multi-modal feature parsing, and generates a composite feature vector, breaking through the limitations of traditional single data sources. By jointly extracting features from the port throughput time series, container transportation map, and warehousing equipment status log, it not only comprehensively covers the key information in the shipping and warehousing system, but also can capture the complex internal correlations between different modal data, generating a composite feature vector containing ship scheduling time series, cargo turnover cycle, and warehousing capacity fluctuations, providing rich and accurate basic data for subsequent predictions, and greatly improving the information value of the data.

[0059] In the feature encoding stage, spatio-temporal encoding processing is performed on the composite feature vector. Using a recurrent convolutional network to extract the time dependence features of ship arrival times can accurately grasp the dynamic change laws in the ship arrival time series, and combining with the graph attention mechanism to capture the cargo flow correlation across warehousing areas enables the model to automatically focus on important warehousing area associations, adaptively allocate attention weights, and thus accurately capture the complex patterns of cargo flow between different warehousing areas. The multi-dimensional spatio-temporal encoding matrix generated by this spatio-temporal encoding processing method can effectively integrate the information in the time and space dimensions, providing a more comprehensive and in-depth feature representation for warehousing demand prediction, enhancing the model's modeling ability for complex real-world scenarios, while traditional methods often struggle to effectively handle the complex cargo flow relationships across regions.

[0060] Input the multi-dimensional spatio-temporal encoding matrix into the pre-trained basic shipping and warehousing demand prediction model. The output preliminary warehousing demand distribution data includes the expected cargo accumulation volume and storage cycle probability curve for each warehousing sub-region. Compared with the general data provided by traditional prediction methods, it provides more detailed and accurate information for port operators. Operators can plan the allocation of warehousing resources in advance based on this data, such as reasonably arranging the cargo storage plans for different warehousing sub-regions, optimizing the warehouse layout, improving the utilization rate of warehousing space, and reducing cargo backlogs and handling costs.

[0061] Finally, the preliminary warehousing demand distribution data is corrected to generate the target warehousing demand prediction result, further improving the accuracy and reliability of the prediction. By considering more complex factors and uncertainties in actual scenarios, the preliminary result is optimized and adjusted, so that the final prediction result can better reflect the real situation, helping the port to make more flexible and efficient decisions, reasonably allocate resources, improve the overall operation efficiency, and reduce operation risks when facing the ever-changing shipping logistics demands.

[0062] In a possible implementation manner, the pre-trained basic shipping and warehousing demand prediction model is constructed through the following steps:

[0063] Step S101: Obtain a sample shipping warehouse operation dataset, where the sample shipping warehouse operation dataset includes ship arrival and departure records, a cargo type distribution matrix, and the time series of the operating status of warehouse equipment.

[0064] In this embodiment, taking Port AA as an example, this sample shipping warehouse operation dataset contains rich and comprehensive information. The ship arrival and departure records detail the precise time of each ship entering and leaving Port AA, the ship name, the company to which the ship belongs, and other information. For example, a large container ship named "Oriental" arrived at a certain dock in Port AA at 9:00 am on May 1, 2023, and left at 5:00 pm on May 3. Such records cover different types of ships, including tankers, bulk carriers, container ships, and many other ships operating in Port AA, with a time span that may be several years. The cargo type distribution matrix presents the types of goods entering and leaving the port and their quantity distribution in each time period. For example, in a certain quarter, electronic products accounted for 30% of the total cargo volume, textiles accounted for 20%, chemical products accounted for 15%, etc., and it will be detailed to the breakdown ratios of different types of electronic products, textiles, etc. The time series of the operating status of warehouse equipment records the changes in the operating conditions of the equipment in each warehouse area in Port AA over time. For example, the stacker in the automated stereoscopic warehouse, its operating speed, fault conditions, and the changes in the load-bearing weight of the shelves at different times will be recorded. These data provide a basic data source for constructing a basic shipping warehouse demand prediction model.

[0065] Step S102: Construct a bidirectional gated recurrent unit network as a time feature extractor to process the sample shipping warehouse operation dataset, perform time-dependency modeling on the ship arrival interval and the cargo turnover cycle, and generate a time-dependency modeling result.

[0066] Still taking Port AA as an example, for the ship arrival intervals, when processing the ship arrival and departure records, the bidirectional gated recurrent unit network will analyze the arrival time of each ship as a time node. In the long run, there are certain regularities in the operation cycles of different ships. For example, large container ships on some international routes may arrive at the port at fixed intervals. The network will process the ship arrival time series separately from the forward propagation layer and the backward propagation layer. The forward propagation layer can capture the positive change trend of the ship arrival intervals over time. For example, due to the development of the shipping market, the arrival intervals of ships on some routes have gradually shortened. The backward propagation layer can analyze from back to front and discover some time-dependent relationships that may have been overlooked. For example, during a certain special period in the past, due to the upgrade and transformation of port facilities, the ship arrival intervals had short-term fluctuations, and these fluctuations had a long-term impact on the subsequent ship arrival intervals. By separately extracting the forward time-dependent features and the backward time-dependent features, outputting the forward hidden state sequence and the backward hidden state sequence, and then performing stride concatenation, a fused hidden state matrix containing bidirectional temporal correlations is generated.

[0067] For the time-dependent modeling of the cargo turnover cycle, extract the time slice sequence of the cargo type distribution matrix from the sample shipping and warehousing operation dataset, and perform step-by-step concatenation with the periodic pattern feature vector of the ship scheduling time series to form the joint time feature input. Assume that for electronic product cargo, its turnover cycle varies in different seasons. Before the new product release season, the import volume of electronic products will increase and the turnover cycle will be shorter because the goods need to be transported to the sales channels as soon as possible. The deep stacking layer of the bidirectional gated recurrent unit network will capture the long-term and short-term dependencies of the cargo turnover cycle step by step through multiple gating mechanisms, and output a multi-level time feature tensor. For example, in the short term, the turnover of a certain batch of electronic products may be affected by the recent logistics efficiency at the port; in the long term, it may be affected by the change in the supply and demand relationship in the entire electronic product market. Then, perform time dimension compression on the multi-level time feature tensor, use the adaptive pooling layer to extract the maximum time response feature and the average time trend feature and concatenate them as the initial result of the time-dependent modeling. Then construct a time attention mechanism module, calculate the contribution weights of the features at different time steps in the initial result to the warehousing demand prediction, and perform dynamic weighted adjustment on the initial result according to the contribution weights to generate an optimized time-dependent feature representation. Finally, input the optimized time-dependent feature representation into the regularization layer for feature scaling and noise suppression, output the standardized time feature vector, and input it into the prediction unit to generate the predicted values of the ship arrival intervals and the cargo turnover cycle, calculate the prediction error of the time-dependent modeling with the true labels, and use the backpropagation algorithm to jointly optimize the parameters of the bidirectional gated recurrent unit network, the time attention mechanism module, and the prediction unit until the prediction error reaches the convergence condition to generate the final time-dependent modeling result.

[0068] Step S103, construct a graph convolutional network as a spatial feature extractor to process the sample shipping warehouse operation dataset, and extract the spatial association features of the cargo transfer path based on the processed warehouse area topology graph.

[0069] Taking the warehouse area of Port AA as an example, based on the warehouse area coordinates and transfer path connection relationships in the sample shipping warehouse operation dataset, a warehouse area topology graph with warehouse nodes as vertices and cargo transfer paths as edges can be constructed. This topology graph includes a node attribute matrix and an adjacency relationship tensor. Each node in the node attribute matrix represents a warehouse area, and its attributes include the area of the warehouse area, storage capacity, equipment type, etc. For example, a large open-air warehouse area may have a large area, but the types of stored goods are relatively single, mainly bulk goods, and its equipment is mainly large cranes. Dynamically expand the node attribute matrix according to the infrastructure capacity and real-time cargo throughput of the warehouse nodes to generate a fused node feature vector containing static attributes and dynamic states. If the cargo throughput of a certain warehouse area has increased recently, this dynamic information will be incorporated into the fused node feature vector.

[0070] Calculate the edge weight coefficients of each transportation path to generate a weighted adjacency matrix. This edge weight coefficient is determined based on the path usage frequency and transportation time consumption in historical cargo transportation records. For example, for a transportation path connecting a dock and a nearby warehousing area, since it is frequently used and has a short transportation time consumption, its edge weight coefficient will be relatively high. Construct a multi-layer graph convolutional network, input the fused node feature vector and the weighted adjacency matrix into the first graph convolutional layer, and generate a primary spatial embedding representation through neighborhood node feature aggregation. Then input the primary spatial embedding representation into the second graph convolutional layer, and perform weighted feature propagation in combination with the edge weight coefficient to generate a high-order spatial feature matrix. Divide the space into clusters based on the physical layout of the warehousing area, perform a graph pooling operation on the high-order spatial feature matrix, and extract the feature summary vectors of different space clusters. For example, according to the functional division of the warehousing area, the warehousing areas storing the same type of goods or serving the same type of shipping routes are divided into one space cluster. According to the temporal variation law of the cargo transportation path, dynamically update the adjacency matrix to generate a time-aware dynamic graph structure. For example, as the season changes, the transportation paths of some goods may change, and this information will be reflected in the dynamic graph structure. Input the dynamic graph structure and the feature summary vector into the temporal graph attention layer, and capture the spatial dependencies across time slices through the self-attention mechanism. For example, under different seasons, the cargo transfer dependencies between each warehousing area may be different, and the temporal graph attention layer can capture this change. Perform multi-scale fusion on the attention-weighted spatial features to generate a spatially associated feature map with a hierarchical structure. Finally, perform cross-modal alignment on the spatially associated feature map and the time-dependency modeling result, and output a joint spatio-temporal feature tensor for processing by the fully connected prediction layer.

[0071] In step S104, after tensor concatenation of the outputs of the time feature extractor and the space feature extractor, input them into the fully connected prediction layer to generate a preliminary demand prediction value.

[0072] Taking the warehousing demand prediction of Port AA as an example, the features after tensor concatenation include time features such as ship arrival intervals and cargo turnover cycles, as well as information such as spatially associated features of cargo transportation paths. The fully connected prediction layer will perform complex calculations based on these comprehensive features to predict the preliminary demands of each warehousing sub-area. For example, for the textile warehousing sub-area, based on the input features, the fully connected prediction layer will predict preliminary demand prediction values such as the possible cargo accumulation volume and the required warehousing space size in the future period.

[0073] Step S105, the parameters of the time feature extractor, the spatial feature extractor, and the fully connected prediction layer are jointly optimized using an adaptive moment estimation algorithm. After the prediction error converges, the basic shipping warehouse demand prediction model is generated, where the basic shipping warehouse demand prediction model includes the time feature extractor, the spatial feature extractor, and the fully connected prediction layer.

[0074] In this embodiment, during the optimization process, the error between the predicted value and the actual value is continuously calculated, and this actual value is obtained based on the accurate past warehouse operation data of Port AA. As the number of iterations increases, the prediction error gradually converges. For example, at the beginning, the error in predicting the cargo accumulation in the textile warehouse area may be relatively large, but as the algorithm continuously adjusts the parameters of the time feature extractor, the spatial feature extractor, and the fully connected prediction layer, the prediction error gradually decreases. When the prediction error converges to an acceptable range, the basic shipping warehouse demand prediction model is generated. This model includes the optimized time feature extractor, spatial feature extractor, and fully connected prediction layer, and can be used for subsequent warehouse demand prediction work.

[0075] In a possible implementation manner, step S102 includes:

[0076] Step S1021, perform time alignment processing on the ship arrival and departure records and the time series of the operation status of warehouse equipment in the sample shipping warehouse operation dataset to generate multi-dimensional time series data with unified timestamps.

[0077] In this embodiment, taking the shipping warehouse operation of Port AA as an example, for example, the "Ocean Star" ship arrived at the port at 10:00 am on March 1, 2023, and left the port at 3:00 pm on March 3; the "Voyage" arrived at the port at 8:00 am on March 5, 2023, and left the port at 11:00 am on March 7, etc. The time series of the operation status of warehouse equipment records information such as the operation duration of cranes and the load changes of shelves, and the recording time intervals of these data may be different. In order to enable unified analysis, time alignment processing is required. The data from different sources are divided according to the same time interval, for example, taking one hour as a time unit, so as to generate multi-dimensional time series data with unified timestamps. In this way, at the same timestamp, the ship arrival and departure situation and the operation status of warehouse equipment can be obtained simultaneously, forming a time series data matrix containing multi-faceted information.

[0078] Step S1022, input the multi-dimensional time series data into the forward propagation layer and the backward propagation layer of a bidirectional gated recurrent unit network, respectively extract the forward time-dependent features and the backward time-dependent features, and output the forward hidden state sequence and the backward hidden state sequence.

[0079] Taking the forward propagation layer as an example, it starts from the starting point of the time series and processes the data in chronological order. When processing the ship arrival intervals, for each ship arrival event, it analyzes the time interval with the previous ship arrival event and combines it with the changes in the operating status of the warehousing equipment. For example, if the warehousing equipment was operating at a high load when the previous ship arrived and the next ship has a short arrival interval, the forward propagation layer will capture this time-dependent relationship that may affect the warehousing demand, extract the forward time-dependent features, and output the forward hidden state sequence. The backward propagation layer starts from the end of the time series and processes the data in reverse. It can discover some relationships that are not easily noticeable in the forward propagation. For example, during a certain period, the maintenance plan of the warehousing equipment may affect the subsequent ship arrival arrangements, and the backward propagation layer can capture this reverse time-dependent relationship and output the backward hidden state sequence.

[0080] Step S1023: Perform a stride splicing on the forward hidden state sequence and the backward hidden state sequence to generate a fused hidden state matrix containing bidirectional temporal correlations.

[0081] Suppose the forward hidden state sequence is [state1, state2, state3] and the backward hidden state sequence is [state4, state5, state6]. By splicing according to a specific stride rule, a fused hidden state matrix containing bidirectional temporal correlations such as [state1, state4, state2, state5, state3, state6] may be obtained. Each element in this fused hidden state matrix integrates the time-dependent information from the forward and backward propagation layers.

[0082] Step S1024: Based on the feature vectors at each time step in the fused hidden state matrix, construct a time-dependent feature encoding for the ship arrival intervals, and perform multi-scale aggregation on the time-dependent feature encoding through a sliding time window to generate a periodic pattern feature vector for the ship scheduling time series.

[0083] For each time step, a specific encoding is generated according to the historical data of the ship arrival intervals and information such as the status of the associated warehousing equipment. For example, when the ship arrival interval is short and the warehousing equipment is operating normally, the encoding may be represented as a specific vector [value1, value2, value3]. Then, multi-scale aggregation is performed on this time-dependent feature encoding through a sliding time window. The sliding time window can be set to different sizes, such as 3 hours, 6 hours, etc. Taking a 3-hour sliding time window as an example, the feature encodings of the ship arrival intervals within 3 hours will be aggregated, comprehensively considering the ship arrival situations and the changes in the status of the warehousing equipment within these 3 hours, thereby generating a periodic pattern feature vector for the ship scheduling time series. This vector can reflect the periodic patterns of the ship arrival intervals at different time scales.

[0084] Step S1025: Extract the time slice sequence of the cargo type distribution matrix from the multi-dimensional time series data, and splice it with the periodic pattern feature vector step by step in time to form a joint time feature input.

[0085] In this embodiment, the cargo type distribution matrix records the quantity change of different cargo types. For example, on a certain day, there are 100 twenty-foot equivalent units (TEUs) of electronic products, 50 TEUs of textiles, etc. Extract the time slices of this cargo type distribution matrix in chronological order to form a time slice sequence. Splice this time slice sequence with the periodic pattern feature vector of the ship scheduling time series step by step in time to form a joint time feature input. For example, if at a certain time step, the periodic pattern feature vector of the ship scheduling time series indicates that the arrival of ships is relatively intensive, and the time slice of the cargo type distribution matrix shows an increase in the import volume of electronic products, then this joint time feature input contains information on both aspects and can more comprehensively reflect the status of shipping and warehousing operations.

[0086] Step S1026: Input the joint time feature into the deep stacked layer of the gated recurrent unit network, capture the long-term and short-term dependencies of the cargo turnover cycle step by step through multiple gating mechanisms, output a multi-level time feature tensor, compress the time dimension of the multi-level time feature tensor, use the adaptive pooling layer to extract the maximum time response feature and the average time trend feature, and splice the two to form the initial result of time-dependent modeling.

[0087] In this embodiment, in this deep stacked layer, the long-term and short-term dependencies of the cargo turnover cycle are captured step by step through multiple gating mechanisms. Taking the cargo turnover cycle of electronic products as an example, in the short term, it may be affected by the immediate logistics operation efficiency within the port, such as the loading and unloading speed, congestion of the transfer path, etc.; in the long term, it may be affected by the supply and demand changes in the entire electronic product market, such as the release cycle of new products, seasonal consumption demand changes, etc. Each layer of the gating mechanism will analyze and extract these long-term and short-term dependencies, and finally output a multi-level time feature tensor. This tensor contains the time feature information of the cargo turnover cycle extracted from different levels of the gating mechanism.

[0088] Since the multi-level time feature tensor contains rich time feature information, but a more compact representation may be required in subsequent processing. The adaptive pooling layer is used to extract the maximum time response feature and the average time trend feature. The maximum time response feature may reflect the key changes in the goods turnover cycle at a specific time point. For example, during a certain promotion season, the turnover speed of electronic products reaches its maximum value. The average time trend feature reflects the average change trend of the goods turnover cycle over a long period of time. These two features are concatenated to obtain the initial result for time-dependent modeling.

[0089] Step S1027: Construct a time attention mechanism module, calculate the contribution weights of the features at different time steps in the initial result to the warehousing demand prediction, and perform dynamic weighted adjustment on the initial result according to the contribution weights to generate an optimized time-dependent feature representation.

[0090] For example, within a certain time period, the change in the goods turnover cycle may have a greater impact on the warehousing demand. Then the contribution weight of the time step feature corresponding to this time period in the warehousing demand prediction will be higher. Perform dynamic weighted adjustment on the initial result according to this contribution weight to generate an optimized time-dependent feature representation. This can highlight the time features that are more critical for the warehousing demand prediction and improve the prediction accuracy.

[0091] Step S1028: Input the optimized time-dependent feature representation into a regularization layer for feature scaling and noise suppression, output a standardized time feature vector, input the standardized time feature vector into a prediction unit to generate predicted values of the ship arrival interval and the goods turnover cycle, and calculate the prediction error of the time-dependent modeling with the true label.

[0092] Step S1029: Use the backpropagation algorithm to jointly optimize the parameters of the bidirectional gated recurrent unit network, the time attention mechanism module, and the prediction unit until the prediction error reaches the convergence condition to generate the final time-dependent modeling result.

[0093] In the actual shipping warehousing operation data, there may be some data fluctuations or noises. For example, outliers in the goods turnover cycle due to measurement errors or temporary special situations. The regularization layer will reasonably scale each feature value in the feature representation to make it within a suitable numerical range, and at the same time suppress the influence of these noises, and output a standardized time feature vector.

[0094] Finally, the prediction unit calculates the predicted values of the ship arrival interval and the cargo turnover cycle based on the input standardized time feature vector and the pre-trained model parameters. Then, these predicted values are used to calculate the prediction error of the time dependence modeling with the true labels (actual ship arrival interval and cargo turnover cycle data). The backpropagation algorithm is used to jointly optimize the parameters of the bidirectional gated recurrent unit network, the time attention mechanism module, and the prediction unit. During the optimization process, the parameters of each component are adjusted according to the magnitude of the prediction error. For example, if the predicted value of the ship arrival interval differs significantly from the true value, the backpropagation algorithm will adjust the parameters related to the ship arrival interval in the bidirectional gated recurrent unit network and the parameters related to the calculation of the time step feature weights of the ship arrival interval in the time attention mechanism module, etc. This process is iterated until the prediction error reaches the convergence condition, generating the final time dependence modeling result. This final result can accurately reflect the time dependence relationship of the ship arrival interval and the cargo turnover cycle, providing important time feature information for the prediction of shipping and warehousing demand.

[0095] For example, in a possible implementation manner, step S103 includes:

[0096] Step S1031: Based on the storage area coordinates and transfer path connection relationships in the sample shipping and warehousing operation dataset, a storage area topology graph is generated with storage nodes as vertices and cargo transfer paths as edges. The topology graph includes a node attribute matrix and an adjacency relationship tensor.

[0097] Taking Port AA as an example, the storage area coordinates determine the geographical location of each storage area within the port. For example, Storage Area A is located in the northeast of the port with coordinates (x1, y1); Storage Area B is located in the southwest with coordinates (x2, y2), etc. The transfer path connection relationships describe the channels for cargo transfer between storage areas. For example, there is a direct transfer path from Storage Area A to Storage Area B, and to go from Storage Area A to Storage Area C, it needs to pass through Storage Area B, etc. This storage area topology graph includes a node attribute matrix and an adjacency relationship tensor. Each row in the node attribute matrix corresponds to a storage node and contains attribute information such as the area and storage capacity of the storage area. For example, the area of Storage Area A is 1000 square meters and the storage capacity is 500 twenty-foot equivalent units (TEU); the area of Storage Area B is 800 square meters and the storage capacity is 400 TEU, etc. The adjacency relationship tensor represents the connection relationships between storage nodes. For example, if there is a direct transfer path connection between Storage Area A and Storage Area B, there will be corresponding markings at the corresponding positions in the adjacency relationship tensor.

[0098] Step S1032: Dynamically expand the node attribute matrix according to the infrastructure capacity and real-time cargo throughput of the warehousing node to generate a fused node feature vector containing static attributes and dynamic states.

[0099] Suppose the infrastructure capacity of Warehouse Area A is a shelf that can accommodate 500 TEUs, and the real-time cargo throughput at a certain moment is 300 TEUs. Incorporate this dynamic information into the node attribute matrix. The original attribute vector may be expanded from [area, storage capacity] to [area, storage capacity, infrastructure capacity, real-time cargo throughput], generating a fused node feature vector containing static attributes and dynamic states. Such a fused node feature vector can more comprehensively reflect the actual situation of the warehousing node.

[0100] Step S1033: Calculate the edge weight coefficients of each transfer path based on the path usage frequency and transportation time in the historical cargo transfer records, and generate a weighted adjacency matrix.

[0101] For example, for the transfer path from Warehouse Area A to Warehouse Area B, it has been used 10 times in the past month, and the average transportation time is 2 hours. According to a specific calculation rule (such as the higher the usage frequency and the shorter the transportation time, the higher the edge weight coefficient), calculate the edge weight coefficient of this transfer path. Perform such calculations for all transfer paths to generate a weighted adjacency matrix. The weight values in this adjacency matrix can reflect the importance of the transfer path in the cargo transfer process.

[0102] Step S1034: Construct a multi-layer graph convolutional network, input the fused node feature vector and the weighted adjacency matrix into the first graph convolutional layer, and generate a primary spatial embedding representation through neighborhood node feature aggregation.

[0103] In the first graph convolutional layer, a primary spatial embedding representation is generated through neighborhood node feature aggregation. For each warehousing node, it aggregates the feature information of its neighborhood nodes (nodes connected to it by transfer paths). For example, the primary spatial embedding representation of Warehouse Area A will contain its own fused node feature vector and partial feature information of Warehouse Areas B and C connected to it, and fuse this information together through a specific aggregation algorithm (such as graph convolutional operation).

[0104] Step S1035: Input the primary spatial embedding representation into the second graph convolutional layer, perform weighted feature propagation in combination with the edge weight coefficients, and generate a high-order spatial feature matrix.

[0105] In the second graph convolutional layer, the edge weight coefficients affect the intensity of feature propagation. If the edge weight coefficient of a transfer path is high, then the feature information propagated along this path will be more emphasized. Through this weighted feature propagation, a high-order spatial feature matrix is generated. This high-order spatial feature matrix contains higher-level spatial feature information after being processed by two graph convolutional layers, and can more deeply reflect the spatial relationship between storage areas.

[0106] Step S1036: Based on the physical layout of the storage areas, divide spatial clusters, and perform a graph pooling operation on the high-order spatial feature matrix to extract the feature summary vectors of different spatial clusters.

[0107] For example, according to factors such as the function of the storage area, the type of goods, or the geographical location, the storage area storing electronic products is divided into one spatial cluster, and the storage area storing textiles is divided into another spatial cluster, etc. Perform a graph pooling operation on the high-order spatial feature matrix to extract the feature summary vectors of different spatial clusters. The graph pooling operation can compress and summarize the high-order spatial features within each spatial cluster to obtain a feature summary vector that can represent the overall features of this spatial cluster.

[0108] Step S1037: According to the temporal variation law of the goods transfer path, dynamically update the adjacency matrix to generate a time-aware dynamic graph structure.

[0109] For example, during different seasons or when market demand changes, the goods transfer path may change. If in a certain season, due to an increase in the import volume of electronic products, a new transfer path is opened directly from the dock to the electronic product storage area, then this new connection relationship needs to be added to the adjacency matrix, so as to generate a dynamic graph structure that can reflect this temporal change.

[0110] Step S1038: Input the dynamic graph structure and the feature summary vectors into the temporal graph attention layer, and capture the spatial dependence relationships across time slices through the self-attention mechanism.

[0111] For example, during different time periods, the goods transfer dependence relationships between various spatial clusters may change. During the promotion season, the goods transfer between the electronic product storage area and the sales channel storage area will be more frequent, and the temporal graph attention layer can capture the changes in the spatial dependence relationships across time slices.

[0112] Step S1039: Perform multi-scale fusion on the attention-weighted spatial features to generate a spatial association feature map with a hierarchical structure, perform cross-modal alignment on the spatial association feature map and the time dependence modeling results, and output a joint spatio-temporal feature tensor for processing by the fully connected prediction layer.

[0113] In this embodiment, the spatial features at different levels in this spatial association feature map reflect the spatial association relationship between the storage areas from local to global. For example, at a lower level, it may reflect the cargo transfer relationship between adjacent storage areas, and at a higher level, it may reflect the cargo transfer relationship between different functional areas.

[0114] Cross-modal alignment is to enable the effective fusion and analysis of temporal features and spatial features in the same framework. The joint spatio-temporal feature tensor contains temporal features such as the interval between ship arrivals and the cargo turnover cycle, as well as information such as the spatial association features of the cargo transfer path, and can provide comprehensive input for the fully connected prediction layer for predicting the shipping storage demand.

[0115] In a possible implementation manner, step S140 may include:

[0116] Step S141, calculating the demand impact factors of different storage nodes based on the locally sensitive feature domain of the multi-dimensional spatio-temporal coding matrix and the storage cycle probability curve. The locally sensitive feature domain is the feature subspace in the spatio-temporal coding matrix that reflects the cargo retention risk.

[0117] In this embodiment, taking the storage area of Port AA as an example, the multi-dimensional spatio-temporal coding matrix covers a lot of information related to shipping storage. Among them, the locally sensitive feature domain is the feature subspace that reflects the cargo retention risk. For each storage node, its corresponding feature slice is extracted from this locally sensitive feature domain. For example, in Storage Area A of Port AA, its feature slice may include information such as the recent change in cargo flow volume in this area, the distribution of cargo types, and the operating efficiency of relevant storage equipment. Calculate the local sensitivity index of this feature slice, and this index reflects the contribution degree of this feature slice to the prediction of the cargo accumulation volume. Suppose the recent cargo flow volume in Storage Area A fluctuates greatly and the equipment operating efficiency has decreased, then the contribution degree of this feature slice to the prediction of the cargo accumulation volume may be relatively high, and the corresponding local sensitivity index is also relatively large.

[0118] Meanwhile, calculate the cargo detention risk scores for each warehousing node based on the storage cycle probability curve. The storage cycle probability curve describes the probability of the cargo staying at the warehousing node for different durations. For Warehouse Area A, if the storage cycle probability curve shows that there is a high probability of the cargo being detained for a long time, for example, the probability of being stored for more than 3 months reaches 30%, since long-term detention may bring risks such as cargo deterioration and excessive occupation of warehousing space, the cargo detention risk score will be relatively high. Then, normalize and weight the local sensitivity index and the detention risk score to generate a comprehensive demand impact factor matrix. For example, the normalized local sensitivity index of Warehouse Area A is 0.6, and the normalized detention risk score is 0.8. According to a specific weighting rule (such as the weighting coefficient is determined based on historical data and experience), the calculated comprehensive demand impact factor is 0.7. This comprehensive demand impact factor matrix comprehensively considers the impacts of different factors of each warehousing node on demand and provides an important basis for subsequent corrections.

[0119] Step S142: Screen out the target extractors with significant interpretability from multiple feature extractors of the basic shipping warehousing demand prediction model, and dynamically weight and fuse the output features of the target extractors in combination with the demand impact factor to generate feature traceability data representing the driving factors of warehousing demand.

[0120] In the basic shipping warehousing demand prediction model, there are multiple feature extractors, such as the time feature extractor and the space feature extractor. Identify the interpretability indicators of each feature extractor, which are obtained by calculating the backpropagation saliency map. Taking the time feature extractor as an example, during the process of modeling the time dependence of the ship arrival interval and the cargo turnover cycle, different parts have different contributions to the final warehousing demand prediction. Through the backpropagation saliency map, the interpretability indicator of each part can be calculated. If the interpretability indicator of a certain part of the time feature extractor exceeds the preset threshold, it indicates that this part has significant interpretability in the warehousing demand prediction.

[0121] Select the feature extractor whose interpretability index exceeds the preset threshold as the target extractor. Suppose that the parts of the time feature extractor related to the long-term and short-term dependence analysis of the goods turnover cycle and the parts of the spatial feature extractor related to the extraction of spatial association features of the goods transfer path between storage areas have relatively high interpretability indexes. Then, the corresponding feature extractors of these two parts are selected as the target extractors. Extract the intermediate feature representations of the target extractors in the multi-layer network of the basic shipping warehouse demand prediction model to form a multi-granularity feature set. For example, the time feature extractor may extract goods turnover features at different time scales in different layers, from short-term daily turnover features to long-term seasonal turnover features, etc. The spatial feature extractor will also extract spatial association features of storage areas with different granularities in different layers. Summarizing these features from different layers forms a multi-granularity feature set.

[0122] Hierarchically weight the multi-granularity feature set according to the demand impact factor to generate a feature traceability vector with spatio-temporal interpretability. Different weights are given to different features in the multi-granularity feature set according to the demand impact factor. For example, the goods turnover cycle feature highly related to the goods retention risk will be given a larger weight according to the demand impact factor due to its importance in warehouse demand prediction. In this way, the demand impact factor is incorporated into the multi-granularity feature set to generate a feature traceability vector that can represent the driving factors of warehouse demand. This vector can explain the reasons for the changes in warehouse demand from both time and space dimensions.

[0123] Step S143: Construct a spatio-temporal attention distribution map reflecting the weights of the goods flow pattern based on the feature traceability data, and perform feature enhancement fusion on the spatio-temporal attention distribution map and the shipping dynamic dataset to generate an optimized input matrix.

[0124] This initial attention weight matrix reflects the degree of attention to the goods flow in different time and space dimensions. For example, within a certain time period or in a certain storage area, the goods flow may be more frequent or have a greater impact on the warehouse demand. Then, there will be a higher weight value at the corresponding position in the initial attention weight matrix. Decompose the initial attention weight matrix in the spatio-temporal dimension to obtain a time attention sub-matrix and a space attention sub-matrix. The time attention sub-matrix mainly focuses on the changing rules of the goods flow over time, such as the differences in the goods flow volume in different seasons or on different working days; the space attention sub-matrix focuses on the goods flow relationship between storage areas, for example, which storage areas have more frequent goods transfers between them.

[0125] The gating mechanism is adopted to fuse the temporal attention sub-matrix and the spatial attention sub-matrix to generate a spatio-temporal joint attention distribution. The gating mechanism can reasonably adjust the attention weights in the temporal and spatial dimensions according to the actual situation of the goods flow. For example, during a certain promotion season, if the flow of goods between specific storage areas is not only more concentrated in time but also has a specific flow direction in space, the gating mechanism will adjust the weights accordingly, so that the spatio-temporal joint attention distribution can accurately reflect this goods flow pattern. The spatio-temporal joint attention distribution is sparsified, and the top K attention connections with weight values higher than the dynamic threshold are retained to construct a spatio-temporal attention distribution map reflecting the weights of the goods flow pattern. This map can highlight the key temporal and spatial relationships of the goods flow. For example, within a certain time period, the goods flow from storage area A to storage area B is a goods flow pattern that needs to be focused on.

[0126] The spatio-temporal attention distribution map and the shipping goods flow dynamic dataset are fused with feature enhancement to generate an optimized input matrix. First, the spatio-temporal attention distribution map is converted into a feature enhancement mask matrix. The element values of this feature enhancement mask matrix correspond to the weight values in the spatio-temporal attention distribution map and are used to perform feature enhancement on the shipping goods flow dynamic dataset. The original features in the shipping goods flow dynamic dataset are standardized to obtain a normalized feature tensor. For example, for the goods throughput data, it is processed according to a specific standardization method (such as mean-standard deviation standardization) to make its values within a suitable range. Then, the normalized feature tensor and the feature enhancement mask matrix are subjected to a Hadamard product operation to obtain a primary enhanced feature. The Hadamard product operation will enhance or suppress the features in the normalized feature tensor according to the weight values in the feature enhancement mask matrix. For example, if a certain feature has a high weight in the spatio-temporal attention distribution map, then after the Hadamard product operation, the value of this feature in the primary enhanced feature will be relatively large.

[0127] The primary enhanced feature and the original feature are added together through a residual connection to generate a final optimized input matrix. The residual connection can retain the useful information in the original feature while integrating the enhanced information in the primary enhanced feature. For example, the original goods type distribution feature, after passing through the residual connection, not only retains its basic information but also integrates the enhanced information related to the goods flow pattern brought by the spatio-temporal attention distribution map, so that the final optimized input matrix can more comprehensively and accurately reflect the actual situation of the shipping and storage.

[0128] Step S144, based on the optimized input matrix, correct the preliminary storage demand distribution data to generate a target storage demand prediction result.

[0129] For the expected cargo accumulation volume and storage cycle probability curve of each warehousing zone in the preliminary warehousing demand distribution data, taking the textile warehousing zone of Port AA as an example, it is corrected according to the information in the optimization input matrix. If the cargo flow pattern in the optimization input matrix shows that a batch of new textile orders have been cancelled recently, then the expected cargo accumulation volume will be correspondingly reduced. For the storage cycle probability curve, if it is found that due to intensified market competition, the sales speed of textiles has accelerated, then the storage cycle probability curve will be adjusted towards the short-cycle direction. In this way, the comprehensive information in the optimization input matrix is used to comprehensively and meticulously correct the preliminary warehousing demand distribution data, so as to generate a more realistic and accurate target warehousing demand prediction result.

[0130] In a possible implementation manner, step S141 includes:

[0131] Step S1411, extracting the feature slices corresponding to each warehousing node from the locally sensitive feature domain of the multi-dimensional spatio-temporal coding matrix, and calculating the local sensitivity index of each feature slice. The local sensitivity index reflects the contribution degree of this feature slice to the prediction of the cargo accumulation volume.

[0132] Step S1412, calculating the cargo retention risk score of each warehousing node based on the storage cycle probability curve, and the retention risk score is non-linearly positively correlated with the storage cycle.

[0133] Step S1413, performing normalization weighting on the local sensitivity index and the retention risk score to generate a comprehensive demand impact factor matrix.

[0134] Step S1414, performing dynamic smoothing processing on the demand impact factor matrix through a sliding time window to eliminate short-term fluctuation noise.

[0135] In this embodiment, taking the shipping warehousing of Port AA as an example, the multi-dimensional spatio-temporal coding matrix contains rich information related to shipping warehousing, and the locally sensitive feature domain therein is crucial for analyzing the demand impact factors of each warehousing node.

[0136] First, extract the feature slices corresponding to each warehousing node from the locally sensitive feature domain of the multi-dimensional spatio-temporal coding matrix. Taking the three warehousing nodes A, B, and C in Port AA as an example, for warehousing node A, its feature slices cover a lot of information. For example, from the time dimension, it includes the correlation information between the ship arrival frequency within a specific time period and the cargo handling volume of this warehousing node; from the space dimension, it contains information such as the cargo transfer efficiency between this warehousing node and surrounding warehousing nodes, and the impact of the spatial layout of warehousing equipment on cargo flow. When calculating the local sensitivity index of each feature slice, it is necessary to consider the contribution degree of these factors to the prediction of cargo accumulation volume. Suppose the ship arrival frequency of warehousing node A has increased recently, but the cargo handling efficiency has not increased correspondingly, resulting in the accumulation of goods near the dock. Then, the feature slices related to the ship arrival frequency and cargo handling efficiency have a greater contribution degree to the prediction of cargo accumulation volume, and their local sensitivity index will be relatively high accordingly.

[0137] Calculate the cargo retention risk score for each warehousing node based on the storage cycle probability curve. The storage cycle probability curve reflects the probability distribution of the duration of goods staying at the warehousing node. For warehousing node A, if the storage cycle probability curve shows a relatively high probability that the goods will be retained for more than 3 months, since the long-term retention of goods will occupy warehousing space, increase warehousing costs, and may cause risks such as damage to goods, the cargo retention risk score of this warehousing node will be relatively high. This retention risk score has a non-linear positive correlation with the storage cycle, that is, as the storage cycle increases, the growth rate of the retention risk score may gradually increase to reflect the greater risks brought by long-term retention.

[0138] Normalize and weight the local sensitivity index and the retention risk score to generate a comprehensive demand impact factor matrix. Suppose the normalized local sensitivity index of warehousing node A is 0.6, and the normalized cargo retention risk score is 0.8. When calculating the weight, according to the pre-determined weighting rule (this rule may be based on the analysis of historical data and expert experience), for example, give a weight of 0.4 to the local sensitivity index and a weight of 0.6 to the retention risk score. Calculate the comprehensive demand impact factor as 0.4 * 0.6 + 0.6 * 0.8 = 0.72. The same calculation is also carried out for warehousing nodes B, C, etc., so as to generate a comprehensive demand impact factor matrix. This comprehensive demand impact factor matrix comprehensively considers the local sensitive features and cargo retention risks of each warehousing node, providing a more comprehensive basis for subsequent analysis.

[0139] The demand impact factor matrix is dynamically smoothed by sliding a time window to eliminate short-term fluctuation noise. In actual shipping warehouse operations, short-term fluctuations in demand impact factors may occur due to some temporary factors (such as short-term weather changes affecting ship arrival times, small-scale equipment failures, etc.). Taking Warehouse Node A as an example, assume that due to a short foggy weather, the number of ships arriving at the port on a certain day decreases significantly, causing abnormal fluctuations in the demand impact factor calculated on that day. By performing dynamic smoothing through a sliding time window, for example, setting the sliding time window to 7 days, the demand impact factors within these 7 days are weighted and averaged (higher weights for times closer to the current time and lower weights for more distant times), thereby eliminating this short-term fluctuation noise and enabling the demand impact factor to more stably and accurately reflect the demand situation of the warehouse node.

[0140] In a possible implementation manner, step S142 includes:

[0141] Step S1421, identifying the interpretability metrics of each feature extractor in the basic shipping warehouse demand prediction model, where the interpretability metrics are obtained by calculating the backpropagation saliency map.

[0142] Step S1422, selecting the feature extractors whose interpretability metrics exceed a preset threshold as target extractors.

[0143] Step S1423, extracting the intermediate feature representations of the target extractors in the multi-layer network of the basic shipping warehouse demand prediction model to form a multi-granularity feature set.

[0144] Step S1424, hierarchically attention-weighting the multi-granularity feature set according to the demand impact factor to generate a spatio-temporal interpretable feature traceability vector.

[0145] In the shipping warehouse demand prediction model of Port AA, there are multiple feature extractors, such as time feature extractors and space feature extractors, etc. Each feature extractor plays a different role in the process of processing shipping warehouse data.

[0146] First, identify the interpretability metrics of each feature extractor in the basic shipping warehouse demand prediction model, which are obtained by calculating the backpropagation saliency map. Taking the time feature extractor as an example, in the process of modeling the time dependence of ship arrival intervals and cargo turnover cycles, different network layers and parameters have different contributions to the final warehouse demand prediction result. Through the backpropagation saliency map, the role of each part in the prediction result can be analyzed in detail, thereby calculating its interpretability metric. For example, in the time feature extractor, the part related to the analysis of peak ship arrival times may have a relatively high interpretability metric for the prediction result because peak ship arrival times often have a greater impact on warehouse demand.

[0147] Select the feature extractor whose interpretability index exceeds the preset threshold as the target extractor. Assume that the preset threshold is 0.5. If the interpretability index of the part related to the long-term and short-term dependence analysis of the goods turnover cycle in the time feature extractor is 0.6, and the interpretability index of the part related to the extraction of spatial association features of the goods transfer path between storage areas in the spatial feature extractor is 0.7, then the corresponding feature extractors of these two parts are selected as the target extractors. This means that these two feature extractors have more significant interpretability in the prediction of storage demand, and their output features are more crucial for accurately predicting storage demand.

[0148] Extract the intermediate feature representation of the target extractor in the multi-layer network of the basic shipping storage demand prediction model to form a multi-granularity feature set. For the selected time feature extractor, in its multi-layer network, different layers may extract goods turnover features at different time scales. For example, at a relatively shallow layer, the daily-scale goods turnover changes may be extracted, and at a relatively deep layer, the seasonal-scale goods turnover patterns may be extracted. These features from different layers together constitute the multi-granularity features of the time feature extractor. Similarly, for the spatial feature extractor, different layers will also extract spatial association features of storage areas at different granularities, such as the goods transfer relationship between local storage areas and the impact of the overall storage area layout on goods flow. Summarize these features from different target extractors and different network layers to form a multi-granularity feature set.

[0149] Perform hierarchical attention weighting on the multi-granularity feature set according to the demand impact factor to generate a feature traceability vector with spatio-temporal interpretability. For each feature in the multi-granularity feature set, different weights are assigned according to the demand impact factor. Taking Warehouse Node A as an example, if its comprehensive demand impact factor is relatively high, it means that the situation of this warehouse node has a greater impact on the overall storage demand. Then the weights of the goods turnover cycle features, spatial association features of storage areas, etc. related to Warehouse Node A in the multi-granularity feature set will be increased accordingly. For example, the feature related to the long-term and short-term dependence of the goods turnover cycle of Warehouse Node A may be assigned a weight of 0.8 according to the size of the demand impact factor, while the weight of the feature with a weaker relationship with other warehouse nodes may be only 0.2. By this way of hierarchical attention weighting, the demand impact factor is incorporated into the multi-granularity feature set, and the generated feature traceability vector can explain the driving factors of storage demand from both the time and space dimensions, providing valuable information for further analysis.

[0150] In a possible implementation manner, step S143 includes:

[0151] Step S1431: Input the feature traceability vector of the feature traceability data into the multi-head self-attention module to generate an initial attention weight matrix.

[0152] Step S1432: Perform spatio-temporal dimensional decomposition on the initial attention weight matrix to obtain a temporal attention sub-matrix and a spatial attention sub-matrix.

[0153] Step S1433: Use a gating mechanism to fuse the temporal attention sub-matrix and the spatial attention sub-matrix to generate a spatio-temporal joint attention distribution.

[0154] Step S1434: Perform sparsification processing on the spatio-temporal joint attention distribution, and retain the top K attention connections with weight values higher than the dynamic threshold to construct a spatio-temporal attention distribution map reflecting the weight of the cargo flow pattern.

[0155] Step S1435: Convert the spatio-temporal attention distribution map into a feature enhancement mask matrix.

[0156] Step S1436: Perform standardization processing on the original features in the shipping dynamic dataset to obtain a normalized feature tensor.

[0157] Step S1437: Perform a Hadamard product operation on the normalized feature tensor and the feature enhancement mask matrix to obtain a primary enhanced feature.

[0158] Step S1438: Add the primary enhanced feature and the original feature through a residual connection to generate a final optimized input matrix.

[0159] Based on the shipping and warehousing scenario of Port AA, input the feature traceability vector of the feature traceability data into the multi-head self-attention module to generate an initial attention weight matrix. The multi-head self-attention module can learn the information relationships in the feature traceability vector from multiple representation subspaces. For example, for the cargo turnover cycle feature and the warehousing area space association feature in the feature traceability vector, the multi-head self-attention module will analyze them separately from different "heads" (representation subspaces). Suppose there are 3 heads. The first head may pay more attention to the short-term changes in the cargo turnover cycle feature and the local transfer relationship in the warehousing area space association feature; the second head may focus on the long-term trend in the cargo turnover cycle feature and the overall layout relationship in the warehousing area space association feature; the third head may comprehensively consider other complex interaction relationships. In this way, an initial attention weight matrix that can reflect the complex relationships between different features is generated. Each element in this initial attention weight matrix represents the importance weight of the corresponding feature in the analysis of the cargo flow pattern.

[0160] Decompose the initial attention weight matrix in the spatio-temporal dimension to obtain a temporal attention sub-matrix and a spatial attention sub-matrix. The temporal attention sub-matrix focuses on the changing patterns of cargo flow over time. For example, in different seasons at Port AA, the volume and direction of cargo flow will vary. During peak seasons, there may be more cargo inflows and outflows, and the time intervals between inflows and outflows are shorter; the opposite is true during off-peak seasons. The temporal attention sub-matrix can capture such temporal changing patterns and highlight the importance weights of different time points for the cargo flow pattern. The spatial attention sub-matrix, on the other hand, focuses on the cargo flow relationships between storage areas. For instance, between Storage Area A and Storage Area B, due to their adjacent geographical locations and convenient transportation facilities, cargo transfer is frequent. Then, in the spatial attention sub-matrix, the connection weight corresponding to the connection between these two storage areas will be higher to reflect this spatial cargo flow relationship.

[0161] Adopt a gating mechanism to fuse the temporal attention sub-matrix and the spatial attention sub-matrix to generate a spatio-temporal joint attention distribution. The gating mechanism can reasonably adjust the attention weights in the temporal and spatial dimensions according to the actual situation of cargo flow. For example, during a special period such as holidays, due to changes in consumer demand, not only does the temporal pattern of cargo flow change (such as concentrated inflows before holidays and outflows after holidays), but the spatial pattern of cargo flow also adjusts (such as certain storage areas being dedicated to storing holiday-related goods, and more cargo flowing to these areas). The gating mechanism will adjust the weights in the temporal attention sub-matrix and the spatial attention sub-matrix according to this situation, so that the fused spatio-temporal joint attention distribution can accurately reflect the cargo flow pattern during this special period.

[0162] Perform sparsification processing on the spatio-temporal joint attention distribution, retaining the top K attention connections with weight values higher than the dynamic threshold to construct a spatio-temporal attention distribution map reflecting the weight of the cargo flow pattern. In actual shipping and warehousing operations, the cargo flow pattern is affected by numerous factors, but some key cargo flow connections play a dominant role in the overall cargo flow pattern. By setting a dynamic threshold (this threshold may be dynamically adjusted according to historical data and current operating conditions), only the top K attention connections with higher weight values are retained. For example, at Port AA, if it is found that the cargo transfer relationship between certain storage areas has a very large impact on the overall cargo flow pattern during a specific period, then the connection weights between these storage areas will be higher than the dynamic threshold and will be retained in the spatio-temporal attention distribution map. This spatio-temporal attention distribution map can clearly display the key weight relationships in the cargo flow pattern, providing an important basis for subsequent feature enhancement and fusion.

[0163] Furthermore, continuing with the example of the shipping and warehousing operation data of Port AA, first convert the spatio-temporal attention distribution map into a feature enhancement mask matrix. The weight values in the spatio-temporal attention distribution map reflect the key relationships in the cargo flow pattern, and these weight values are converted into the element values of the feature enhancement mask matrix. For example, if the weight of the cargo flow connection from Warehouse Area A to Warehouse Area B in the spatio-temporal attention distribution map is relatively high, then in the feature enhancement mask matrix, there will be a relatively high value at the position corresponding to the features related to the cargo flow from Warehouse Area A to Warehouse Area B, and this value will be used to enhance the original features in the shipping cargo flow dynamic dataset.

[0164] Standardize the original features in the shipping cargo flow dynamic dataset to obtain a normalized feature tensor. The shipping cargo flow dynamic dataset contains various original features, such as cargo throughput, ship arrival time, warehousing equipment status, etc. Taking cargo throughput as an example, for different cargo types and different time periods, the numerical ranges of cargo throughput may vary greatly. In order to make these features comparable in subsequent calculations, standardization is required. The mean - standard deviation standardization method can be used, that is, first calculate the mean and standard deviation of the cargo throughput, and then subtract the mean from each cargo throughput value and divide by the standard deviation to obtain the standardized cargo throughput value. The same processing is also performed on other original features in the shipping cargo flow dynamic dataset, thereby obtaining a normalized feature tensor.

[0165] Perform the Hadamard product operation on the normalized feature tensor and the feature enhancement mask matrix to obtain the primary enhanced features. The Hadamard product operation is an operation of multiplying corresponding elements. In this process, the element values in the feature enhancement mask matrix will enhance or suppress the features in the normalized feature tensor. For example, if the value at the position related to the cargo throughput of Warehouse Area A in the feature enhancement mask matrix is relatively high, then after the Hadamard product operation, the feature value corresponding to the cargo throughput of Warehouse Area A in the normalized feature tensor will be enhanced, which means that in subsequent analysis, the feature of the cargo throughput of Warehouse Area A will be given higher importance because it plays a key role in the cargo flow pattern.

[0166] Add the primary enhanced features to the original features through a residual connection to generate the final optimized input matrix. The role of the residual connection is to integrate the enhanced information in the primary enhanced features while retaining the useful information in the original features. For example, for the original feature of cargo throughput, after obtaining the primary enhanced features through normalization and Hadamard product operations, add the primary enhanced features to the original cargo throughput feature through a residual connection. In this way, the cargo throughput feature in the final optimized input matrix contains both the basic information in the original data and the enhanced information related to the cargo flow pattern reflected by the spatio-temporal attention distribution map. The same operation is performed on other original features in the shipping dynamic dataset to generate the final optimized input matrix. This optimized input matrix can more comprehensively and accurately reflect the actual situation of shipping warehousing, providing higher-quality input data for further correction of warehousing demand prediction.

[0167] In a possible implementation manner, the method further includes:

[0168] Step S150, quantify the uncertainty of the target warehousing demand prediction result to generate a demand confidence interval for each warehousing area.

[0169] In this embodiment, taking the shipping warehousing of Port AA as an example, the target warehousing demand prediction result includes important information such as the expected cargo accumulation volume and storage period of each warehousing area. Uncertainty quantification aims to measure the reliability of these prediction results. It is achieved by using statistical analysis methods, such as techniques based on Bayesian inference or Monte Carlo simulation. Taking the storage of electronic products in a certain warehousing area as an example, when calculating the demand confidence interval, considering many factors affecting the warehousing demand of electronic products, such as the market demand fluctuation of electronic products, the stability of the supply chain, and the uncertainty of ship arrivals. Based on historical data and current market dynamics and other information, use Bayesian inference to construct a probability model. This model will comprehensively consider the probability distributions of various factors, such as the different probability distributions of the market demand for electronic products in the peak season and off-season, and the probability of ships arriving on time. Calculate the demand confidence interval at a certain confidence level (for example, 95% confidence level) through this model. Suppose the average expected cargo accumulation volume of the predicted electronic products in a certain time period is 1000 twenty-foot equivalent units (TEU), and the confidence interval obtained through uncertainty quantification may be [800 TEU, 1200 TEU], which means that there is a 95% probability that the actual cargo accumulation volume will fall within this interval.

[0170] Step S160, detect abnormal prediction nodes where the demand confidence interval exceeds a preset fluctuation threshold, and trigger an artificial review mechanism.

[0171] Still taking Port AA as an example, the preset fluctuation threshold is a reasonable range determined based on historical data and operation experience. If the demand confidence interval of a certain warehousing area exceeds this preset fluctuation threshold, it is regarded as an abnormal prediction node. For example, for the textile warehousing area, the preset fluctuation threshold is set to a 30% fluctuation up and down of the expected cargo accumulation. If the confidence interval obtained through uncertainty quantification is [500 TEU, 1500 TEU], and the expected cargo accumulation in this area under normal circumstances is 1000 TEU, and the range of 30% fluctuation up and down is [700 TEU, 1300 TEU], then this confidence interval exceeds the preset fluctuation threshold. At this time, the manual review mechanism will be triggered. During the manual review process, professional warehousing management personnel will carefully review various data related to this warehousing area, including the recent ship arrival plan, the market demand trend of textiles, the operating status of warehousing equipment, etc. They will check whether there are special situations not considered by the prediction model, such as whether there is a sudden increase or cancellation of new large textile orders, or whether there is a warehousing equipment about to be repaired that affects the storage capacity, etc.

[0172] Step S170: Compare the prediction result after review and confirmation with historical data to generate a warehousing layout optimization suggestion plan.

[0173] In Port AA, after the manual review is completed, the prediction result after review and confirmation is obtained. Compare this result with historical data, which covers the shipping warehousing operation situation in the past many years, including warehousing demands, cargo turnover situations, etc. under different seasons and different market environments. Taking a specific warehousing area storing chemical products as an example, if the prediction result after review and confirmation shows that the cargo accumulation of chemical products will increase significantly in a future period, and compared with historical data, it is found that in the past in similar situations, the storage capacity of this warehousing area was insufficient to handle such a large amount of goods. Then a warehousing layout optimization suggestion plan will be generated. This plan may include re-planning the warehousing space, such as whether it is necessary to increase the number of shelves or adjust the shelf layout to improve storage efficiency; whether it is necessary to adjust the cargo storage strategy, for example, whether it is necessary to adjust the storage location for some perishable chemical products to better control the environmental conditions; whether it is necessary to optimize the cargo transfer process, such as adding transfer equipment or adjusting the transfer path to improve the cargo turnover speed, etc. Through such comparison and analysis, the generated warehousing layout optimization suggestion plan can better adapt to future warehousing demands and improve the efficiency and effectiveness of the entire shipping warehousing operation.

[0174] Figure 2 Fig. shows the hardware structure diagram of a shipping warehousing demand prediction system 100 based on machine learning for implementing the above-mentioned machine learning-based shipping warehousing demand prediction method provided by an embodiment of the present invention, as Figure 2As shown, the machine learning-based shipping warehouse demand prediction system 100 may include a processor 110, a machine-readable storage medium 120, a bus 130, and a communication unit 140.

[0175] The machine-readable storage medium 120 may store data and / or instructions. In some embodiments, the machine-readable storage medium 120 may store data obtained from an external terminal. In some embodiments, the machine-readable storage medium 120 may store the data and / or instructions used by the machine learning-based shipping warehouse demand prediction system 100 to execute or use to complete the exemplary methods described in the present invention.

[0176] In a specific implementation process, one or more processors 110 execute the computer-executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the machine learning-based shipping warehouse demand prediction method of the above method embodiments. The processor 110, the machine-readable storage medium 120, and the communication unit 140 are connected through the bus 130, and the processor 110 may be used to control the transceiver actions of the communication unit 140.

[0177] For the specific implementation process of the processor 110, reference may be made to the respective method embodiments executed by the machine learning-based shipping warehouse demand prediction system 100 above. The implementation principles and technical effects are similar, and will not be elaborated here in this embodiment.

[0178] In addition, an embodiment of the present invention also provides a readable storage medium, in which computer-executable instructions are preset. When the processor executes the computer-executable instructions, the above machine learning-based shipping warehouse demand prediction method is implemented.

[0179] It should be noted that, in order to simplify the description of the present invention disclosure and thus help the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, sometimes multiple features are merged into one embodiment, drawing, or description thereof.

Claims

1. A method for predicting shipping warehousing demand based on machine learning, characterized in that: The method comprises: Collect the dynamic data set of shipping logistics of the target port, perform multimodal feature analysis on the dynamic data set of shipping logistics, and generate a composite feature vector including ship scheduling time series, cargo turnover cycle and storage capacity fluctuation; the multimodal feature analysis includes joint feature extraction of port throughput time series, container transportation map and storage equipment status log; Performing spatiotemporal coding processing on the composite feature vector to generate a multi-dimensional spatiotemporal coding matrix; the spatiotemporal coding processing includes extracting the time dependency features of the ship arrival time by using a recurrent convolutional network, and capturing the correlation of cargo flow across storage areas in combination with a graph attention mechanism; Input the multi-dimensional spatiotemporal coding matrix into a pre-trained basic shipping storage demand prediction model to output preliminary storage demand distribution data; the preliminary storage demand distribution data includes the expected cargo accumulation volume and storage cycle probability curve of each storage partition; The preliminary storage demand distribution data is modified to generate a target storage demand forecast result.

2. The method for predicting shipping warehousing demand based on machine learning according to claim 1, characterized in that: The step of correcting the preliminary storage demand distribution data to generate a target storage demand forecast result includes: Based on the local sensitive feature domain of the multidimensional spatiotemporal coding matrix and the storage cycle probability curve, the demand influencing factors of different storage nodes are calculated; the local sensitive feature domain is a feature subspace in the spatiotemporal coding matrix that reflects the risk of cargo detention; Selecting a target extractor with significant explanatory power from multiple feature extractors of the basic shipping storage demand forecasting model, dynamically weighting and fusing the output features of the target extractor in combination with the demand influencing factors, and generating feature traceability data representing the driving factors of storage demand; A spatiotemporal attention distribution map reflecting the weight of the cargo flow pattern is constructed based on the feature traceability data, and the spatiotemporal attention distribution map is fused with the shipping logistics dynamic data set for feature enhancement to generate an optimized input matrix; The preliminary storage demand distribution data is modified based on the optimized input matrix to generate a target storage demand forecast result.

3. The method for predicting shipping warehousing demand based on machine learning according to claim 1, characterized in that: The pre-trained basic shipping storage demand prediction model is constructed by the following steps: Acquire a sample shipping and warehousing operation data set, wherein the sample shipping and warehousing operation data set includes ship arrival and departure records, a cargo type distribution matrix, and a warehousing equipment operation status time series; Constructing a bidirectional gated recurrent unit network as a time feature extractor to process the sample shipping warehousing operation data set to perform time-dependent modeling on the ship arrival interval and the cargo turnover cycle, and generating a time-dependent modeling result; Constructing a graph convolutional network as a spatial feature extractor to process the sample shipping warehousing operation data set, so as to extract spatial correlation features of the cargo transshipment path based on the processed warehousing area topology map; After tensor splicing, the outputs of the temporal feature extractor and the spatial feature extractor are input into a fully connected prediction layer to generate a preliminary demand forecast value; An adaptive moment estimation algorithm is used to jointly optimize the parameters of the temporal feature extractor, the spatial feature extractor and the fully connected prediction layer until the prediction error converges, thereby generating the basic shipping warehousing demand prediction model, wherein the basic shipping warehousing demand prediction model includes the temporal feature extractor, the spatial feature extractor and the fully connected prediction layer.

4. The method for predicting shipping warehousing demand based on machine learning according to claim 3 is characterized in that: The step of constructing a bidirectional gated recurrent unit network as a time feature extractor to process the sample shipping warehousing operation data set to perform time-dependent modeling on the ship arrival interval and the cargo turnover cycle, and generating a time-dependent modeling result includes: Performing time alignment processing on the ship arrival and departure records and storage equipment operation status time series in the sample shipping and storage operation data set to generate multidimensional time series data with a unified timestamp; Inputting the multidimensional time series data into the forward propagation layer and the backward propagation layer of the bidirectional gated recurrent unit network, extracting the forward time-dependent features and the backward time-dependent features respectively, and outputting the forward hidden state sequence and the backward hidden state sequence; Perform stride-length concatenation on the forward hidden state sequence and the backward hidden state sequence to generate a fused hidden state matrix containing bidirectional temporal correlation; Based on the feature vector of each time step in the fused hidden state matrix, a time-dependent feature code of the ship arrival interval is constructed, and the time-dependent feature code is multi-scale aggregated through a sliding time window to generate a periodic pattern feature vector of the ship scheduling time series; Extracting a time slice sequence of a cargo type distribution matrix from the multidimensional time series data, and splicing it with the periodic pattern feature vector time step by time to form a joint time feature input; The joint time feature is input into the deep stacking layer of the gated recurrent unit network, and the long-term and short-term dependencies of the goods turnover cycle are captured step by step through the multi-layer gating mechanism, and a multi-level time feature tensor is output; Compressing the multi-level time feature tensor in time dimension, extracting the maximum time response feature and the average time trend feature by using an adaptive pooling layer, and concatenating the two as the initial result of time-dependent modeling; Constructing a time attention mechanism module, calculating the contribution weights of different time step features in the initial results to the storage demand prediction, dynamically weighting and adjusting the initial results according to the contribution weights, and generating an optimized time-dependent feature representation; Inputting the optimized time-dependent feature representation into a regularization layer for feature scaling and noise suppression, and outputting a standardized time feature vector; The standardized time feature vector is input into the prediction unit to generate the predicted values ​​of the ship arrival interval and the cargo turnover cycle, and the prediction error of the time-dependent modeling is calculated with the true label; The back-propagation algorithm is used to jointly optimize the parameters of the bidirectional gated recurrent unit network, the temporal attention mechanism module and the prediction unit until the prediction error reaches the convergence condition to generate the final time-dependent modeling result.

5. The method for predicting shipping warehousing demand based on machine learning according to claim 2, characterized in that: The calculation of demand influencing factors of different storage nodes based on the local sensitive feature domain of the multidimensional spatiotemporal coding matrix and the storage cycle probability curve includes: Extracting feature slices corresponding to each storage node from the local sensitive feature domain of the multidimensional spatiotemporal coding matrix, and calculating the local sensitivity index of each feature slice; the local sensitivity index reflects the contribution of the feature slice to the prediction of the cargo accumulation amount; Calculating a cargo detention risk score for each storage node based on the storage cycle probability curve, wherein the detention risk score is nonlinearly positively correlated with the storage cycle; Normalizing and weighting the local sensitivity index and the retention risk score to generate a comprehensive demand impact factor matrix; The demand influencing factor matrix is ​​dynamically smoothed through a sliding time window to eliminate short-term fluctuation noise.

6. The method for predicting shipping warehousing demand based on machine learning according to claim 2, characterized in that: The step of selecting a target extractor with significant explanatory power from multiple feature extractors of the basic shipping storage demand forecasting model, dynamically weighting and fusing the output features of the target extractor in combination with the demand influencing factors, and generating feature traceability data representing the driving factors of storage demand includes: Identifying an explanatory index of each feature extractor in the basic shipping storage demand forecasting model, wherein the explanatory index is obtained by calculating a back-propagation saliency map; Selecting the feature extractor whose explanatory index exceeds a preset threshold as the target extractor; Extracting the intermediate feature representation of the target extractor in the multi-layer network of the basic shipping storage demand prediction model to form a multi-granularity feature set; According to the demand influencing factors, hierarchical attention weighting is performed on the multi-granularity feature set to generate a feature traceability vector with spatiotemporal interpretability.

7. The method for predicting shipping warehousing demand based on machine learning according to claim 6, characterized in that: The step of constructing a spatiotemporal attention distribution map reflecting the weight of the cargo flow pattern according to the characteristic traceability data comprises: Inputting the feature tracing vector of the feature tracing data into a multi-head self-attention module to generate an initial attention weight matrix; Decomposing the initial attention weight matrix in time and space dimensions to obtain a time attention sub-matrix and a space attention sub-matrix; The temporal attention sub-matrix and the spatial attention sub-matrix are fused using a gating mechanism to generate a spatiotemporal joint attention distribution; The spatiotemporal joint attention distribution is thinned, and the top K attention connections with weight values ​​higher than a dynamic threshold are retained to construct a spatiotemporal attention distribution map reflecting the weight of the cargo flow pattern.

8. The method for predicting shipping warehousing demand based on machine learning according to claim 2, characterized in that: The step of performing feature enhancement fusion on the spatiotemporal attention distribution map and the shipping logistics dynamic data set to generate an optimized input matrix includes: Converting the spatiotemporal attention distribution map into a feature enhancement mask matrix; Standardizing the original features in the shipping logistics dynamic data set to obtain a normalized feature tensor; Performing a Hadamard product operation on the normalized feature tensor and the feature enhancement mask matrix to obtain primary enhanced features; The primary enhanced features are added to the original features through residual connections to generate a final optimized input matrix.

9. The method for predicting shipping warehousing demand based on machine learning according to any one of claims 1 to 8, characterized in that: The method further comprises: Quantifying the uncertainty of the target storage demand forecast result to generate a demand confidence interval for each storage partition; Detecting abnormal prediction nodes where the demand confidence interval exceeds a preset fluctuation threshold, triggering a manual review mechanism; Compare the verified and confirmed forecast results with historical data to generate storage layout optimization proposals.

10. A shipping storage demand forecasting system based on machine learning, characterized in that: The shipping storage demand forecasting system based on machine learning includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the shipping storage demand forecasting method based on machine learning as described in any one of claims 1 to 9.

Images