Low-altitude UAV delivery demand prediction method and system based on AI prediction model

Abstract

The present invention provides a low-altitude unmanned aerial vehicle delivery demand prediction method and system based on an AI prediction model, which relates to the field of artificial intelligence. First, a basic low-altitude delivery demand prediction network is preliminarily trained using a sample delivery interaction data sequence to generate a temporary low-altitude delivery demand prediction network. Subsequently, new iterative sample delivery interaction data are continuously introduced to perform parameter learning and optimization on the temporary low-altitude delivery demand prediction network, and a more accurate iterative low-altitude delivery demand prediction network is gradually generated. The method not only makes full use of historical delivery data, but also continuously corrects and optimizes the low-altitude delivery demand prediction network through an iterative learning mechanism, so that the network can better adapt to changes in actual delivery demand. The target low-altitude delivery demand prediction network finally generated can accurately and efficiently output prediction results when facing target delivery interaction data in the actual low-altitude unmanned aerial vehicle delivery process, thereby effectively improving the efficiency of unmanned aerial vehicle delivery services and user satisfaction.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method and system for predicting low-altitude unmanned aerial vehicle delivery demand based on an AI prediction model. Background Art

[0002] With the rapid development of science and technology, low-altitude drone delivery has gradually become an emerging trend in the logistics industry, showing great potential in improving delivery efficiency, reducing operating costs, and meeting instant delivery needs. However, the prediction of low-altitude drone delivery demand faces many challenges. Since delivery demand is affected by many factors, such as weather conditions, traffic conditions, changes in customer demand, etc., the prediction of delivery demand becomes complex and changeable.

[0003] Traditional forecasting methods, such as time series analysis and regression analysis, often have difficulty accurately capturing the correlation between these complex factors, resulting in limited accuracy of forecast results. Especially when faced with large-scale, high-dimensional delivery interaction data, the processing capabilities of traditional methods are insufficient and it is difficult to meet the high-precision requirements of actual delivery demand forecasting. Summary of the invention

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides a method for predicting low-altitude UAV delivery demand based on an AI prediction model, the method comprising:

[0005] Acquire a sample delivery interaction data sequence in a low-altitude UAV delivery process; the sample delivery interaction data sequence includes sample delivery interaction data corresponding to a plurality of low-altitude delivery demand prediction modes;

[0006] Obtain candidate sample delivery interaction data corresponding to the candidate low-altitude delivery demand prediction mode, perform parameter learning on the graph convolution conversion unit of the basic low-altitude delivery demand prediction network according to the candidate sample delivery interaction data, and generate a temporary low-altitude delivery demand prediction network;

[0007] Acquire iterative sample delivery interaction data corresponding to the next low-altitude delivery demand prediction mode, perform parameter learning on the graph convolution conversion unit of the temporary low-altitude delivery demand prediction network according to the iterative sample delivery interaction data, and generate an iterative low-altitude delivery demand prediction network;

[0008] Loading the previously learned sample delivery interaction data of the iterative sample delivery interaction data into the graph convolution conversion units of the temporary low-altitude delivery demand prediction network and the iterative low-altitude delivery demand prediction network, respectively, to generate corresponding first graph convolution conversion results and second graph convolution conversion results;

[0009] According to the first graph convolution conversion result and the second graph convolution conversion result corresponding to the same sample delivery interaction data, parameter learning is performed on the classification conversion unit of the iterative low-altitude delivery demand prediction network. After generating an optimized temporary low-altitude delivery demand prediction network, return to the step of obtaining the iterative sample delivery interaction data corresponding to the next low-altitude delivery demand prediction mode until the training termination condition is met, and generate a target low-altitude delivery demand prediction network according to the temporary low-altitude delivery demand prediction network corresponding to the training termination condition;

[0010] The target delivery interaction data in the low-altitude UAV delivery process is obtained, the target delivery interaction data is loaded into the target low-altitude delivery demand prediction network, and the target low-altitude delivery demand prediction result corresponding to the target delivery interaction data is generated.

[0011] On the other hand, an embodiment of the present invention also provides a low-altitude UAV delivery demand prediction system based on an AI prediction model, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0012] Based on the above aspects, the embodiment of the present application significantly improves the prediction accuracy of low-altitude drone delivery demand by iteratively learning and optimizing the low-altitude delivery demand prediction network. The method first uses the sample delivery interaction data sequence to perform preliminary training on the basic low-altitude delivery demand prediction network to generate a temporary low-altitude delivery demand prediction network; then, by continuously introducing new iterative sample delivery interaction data, the temporary low-altitude delivery demand prediction network is parameter-learned and optimized, and a more accurate iterative low-altitude delivery demand prediction network is gradually generated. It not only makes full use of historical delivery data, but also continuously corrects and optimizes the low-altitude delivery demand prediction network through an iterative learning mechanism, so that it can better adapt to changes in actual delivery demand. The target low-altitude delivery demand prediction network finally generated can accurately and efficiently output prediction results when facing the target delivery interaction data in the actual low-altitude drone delivery process, thereby facilitating the scheduling and planning of drone delivery, and effectively improving the efficiency and user satisfaction of drone delivery services. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is a schematic diagram of the execution flow of the low-altitude UAV delivery demand prediction method based on the AI ​​prediction model provided in an embodiment of the present invention.

[0014] Figure 2 It is a schematic diagram of the hardware architecture of a low-altitude UAV delivery demand prediction system based on an AI prediction model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0015] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 It is a flow chart of a method for predicting low-altitude UAV delivery demand based on an AI prediction model provided by an embodiment of the present invention. The method for predicting low-altitude UAV delivery demand based on an AI prediction model is introduced in detail below.

[0016] Step S110, obtaining a sample delivery interaction data sequence in the low-altitude UAV delivery process. The sample delivery interaction data sequence includes sample delivery interaction data corresponding to multiple low-altitude delivery demand prediction modes.

[0017] In this embodiment, in the low-altitude drone delivery scenario, the low-altitude delivery demand involves many factors. For example, consider a low-altitude drone delivery service in a large urban area, where the types of items delivered are diverse, including refrigerated food, medicines, small electronic products, etc. The delivery area covers different types of geographical areas such as commercial areas, residential areas, and industrial areas, and each geographical area has different environmental characteristics and delivery restrictions.

[0018] For the acquisition of sample delivery interaction data, multiple prior delivery interaction data are first acquired. Assuming that there is an order for drug delivery at a certain moment in a commercial area, the prior delivery interaction data may include the time when the order is placed (for example, 9 a.m., which may affect the priority of delivery, because some drugs need to be delivered within a specific time to ensure patient needs), the destination of the order (residents in a certain residential building, which involves the navigation path planning of the drone, and needs to avoid dense areas of high-rise buildings and possible temporary airspace control in the area), the weight of the order (5 kg, the weight will affect the endurance and flight stability of the drone, heavier items may require a more powerful drone or batch delivery), and the customer's special requirements for delivery time (such as delivery before 11 a.m., which is a very strict time limit, which may be related to the business preparation time of the pharmacy or drug storage), and this prior delivery interaction data is marked with prior low-altitude delivery requirements, such as the need to deliver as soon as possible and ensure patient needs under all restrictions.

[0019] For a drug delivery order in a residential area, the a priori delivery interaction data includes the order time (3 p.m., the urgency of drug delivery may vary. The time requirements for common drugs for some patients with chronic diseases may be relatively loose, but some emergency drugs need to be delivered as soon as possible), the destination (a specific residential address in a community, which requires consideration of the impact of the building layout in the community on drone flight. There may be some signal interference areas or no-fly zones), the type of drug (such as emergency heart disease drugs, the nature of which determines the urgency of delivery), and the patient's contact number (used for communication during the delivery process, such as notifying the patient to receive the drug when the drone arrives nearby). The a priori low-altitude delivery demand may be to deliver the drugs to the patient safely and quickly. At the same time, it is necessary to consider the possible children's activity areas in the residential area to avoid flight safety risks.

[0020] Cluster the prior delivery interaction data corresponding to the same prior low-altitude delivery demand. For example, group all the prior delivery interaction data related to refrigerated food delivery into one category to form a first cluster sequence corresponding to a prior low-altitude delivery demand, because they all have similar delivery requirements, such as ensuring freshness and delivering to the destination in the commercial area on time.

[0021] The training value of each prior delivery interaction data is determined based on the prior low-altitude delivery demand. Taking refrigerated food delivery as an example, prior delivery interaction data with larger order weights, stricter delivery time limits, and destinations near traffic congestion areas (which may affect the time of ground transportation to the delivery starting point, and thus affect the overall delivery efficiency) may have higher training value because these data can reflect more complex delivery situations.

[0022] The first cluster sequence is used as the low-altitude distribution demand prediction sub-pattern, and each low-altitude distribution demand prediction sub-pattern corresponding to the same training value is clustered to generate a second cluster sequence corresponding to each training value. For example, the low-altitude distribution demand prediction sub-patterns of refrigerated food distribution and emergency medicine distribution with higher training value may be divided into a second cluster sequence, because although they distribute different items, they are similar in terms of urgency and special requirements for the distribution environment. This second cluster sequence includes each sample distribution interaction data corresponding to the same low-altitude distribution demand prediction mode. Finally, a sample distribution interaction data sequence is generated based on each second cluster sequence.

[0023] Step S120, obtaining candidate sample delivery interaction data corresponding to the candidate low-altitude delivery demand prediction mode, performing parameter learning on the graph convolution conversion unit of the basic low-altitude delivery demand prediction network according to the candidate sample delivery interaction data, and generating a temporary low-altitude delivery demand prediction network.

[0024] Assume that in the previously constructed sample delivery interaction data sequence, a sample delivery interaction data related to the delivery of emergency items (including emergency medicines and other emergency items) with a candidate low-altitude delivery demand prediction mode is selected as the candidate sample delivery interaction data. These candidate sample delivery interaction data contain various complex information, such as order information for emergency delivery of emergency medicines in the city center, including the surrounding traffic flow when the order is placed (obtained through data interaction with the city traffic management system. High traffic flow may affect the time from ground transportation to the drone take-off point, and thus affect the overall delivery time), weather conditions (for example, cloudy days may affect the drone's visual navigation system, but the impact is smaller than that of severe weather such as heavy rain), airspace activities near the destination (whether there are other drones or small aircraft nearby, which involves flight safety and route planning), etc.

[0025] The graph convolutional conversion unit of the basic low-altitude delivery demand prediction network begins to use these candidate sample delivery interaction data for parameter learning. The graph convolutional conversion unit includes multiple graph convolutional conversion subunits. For the graph convolutional conversion subunit corresponding to the order-related interaction data, it processes order-related information, such as the content of the order (whether it is room temperature medicine or refrigerated medicine), the weight of the order, the urgency of the order, etc. Taking the emergency medicine in the order as an example, the graph convolutional conversion subunit will perform graph convolution conversion processing on the name, weight, urgency and other information of the medicine, and convert this information into feature representations that the network can understand.

[0026] For the graph convolution conversion subunit corresponding to the delivery-related interaction data, it processes relevant information in the delivery process, such as the starting point of the delivery (hospital pharmacy or cold storage warehouse), the delivery route (including the geographical path from the starting point to the destination and the obstacles and restricted areas that may be encountered on the way), and the environmental conditions during delivery (weather, airspace activities, etc.). For example, for the delivery route from the hospital pharmacy to the patient's residence, the graph convolution conversion subunit will perform graph convolution conversion processing on the geographical coordinates on the route, possible high-rise buildings (as obstacles), nearby airports (as airspace restricted areas), and other information.

[0027] During the processing, the basic low-altitude delivery demand prediction network is used to predict any input candidate sample delivery interaction data, and generate an estimated low-altitude delivery demand corresponding to any input candidate sample delivery interaction data. For example, based on the input candidate sample delivery interaction data of a certain emergency medicine delivery, the network may predict an estimated low-altitude delivery demand, including the recommended drone type (selecting a drone with appropriate power and range based on the weight and distance of the medicine), the estimated delivery time (considering factors such as traffic, weather, and routes), and the recommended flight altitude (avoiding high-rise buildings and other airspace activity areas).

[0028] Then, the network parameter information of the basic low-altitude distribution demand prediction network is optimized based on the network learning supervision data corresponding to any input candidate sample delivery interaction data and the loss function value of the estimated low-altitude distribution demand. The network learning supervision data may be the actual distribution result data, such as the type of drone actually used, the actual delivery time, the actual flight altitude, etc. By comparing the estimated low-altitude distribution demand and the actual network learning supervision data, the loss function value is calculated. If the loss function value is large, it means that the predicted result deviates greatly from the actual result. It is necessary to adjust the network parameters, such as adjusting the convolution kernel parameters and weights in the graph convolution conversion subunit, until the training termination conditions are met. For example, when the loss function value is less than a preset threshold or reaches a certain number of training rounds, a temporary low-altitude distribution demand prediction network is generated.

[0029] Step S130, obtaining the iterative sample delivery interaction data corresponding to the next low-altitude delivery demand prediction mode, performing parameter learning on the graph convolution conversion unit of the temporary low-altitude delivery demand prediction network based on the iterative sample delivery interaction data, and generating an iterative low-altitude delivery demand prediction network.

[0030] Assume that the latter low-altitude delivery demand prediction model is about the sample delivery interaction data related to the delivery of electronic parts to a specific area (such as an enterprise in an industrial park) during a specific time period (such as nighttime) as iterative sample delivery interaction data. These data contain some special information, such as the lighting conditions in the industrial park at night (some areas may have insufficient lighting, affecting the visual navigation of the drone), the company's receiving process (some companies may require specific personnel to receive at a specific location, and may have strict safety inspection processes, which will affect the final stage of delivery), the nature of the parts (such as whether it is a precision electronic component, which may have special requirements for environmental conditions such as vibration, temperature, humidity, etc. during the delivery process), etc.

[0031] The graph convolutional transformation unit of the temporary low-altitude delivery demand prediction network begins to use these iterative sample delivery interaction data for parameter learning. For the graph convolutional transformation subunit corresponding to the order-related interaction data, it processes the relevant information of the electronic component order, such as the model, quantity, production batch, etc. of the component, and converts this information into a feature representation that the network can understand. For the graph convolutional transformation subunit corresponding to the delivery-related interaction data, it processes the starting point of the delivery (electronic component warehouse), the delivery route (the route from the warehouse to the enterprise in the industrial park, which needs to consider the road layout, building distribution and special circumstances at night in the industrial park), and the environmental conditions during delivery (lighting conditions, temperature, humidity, etc.).

[0032] During the processing, the temporary low-altitude delivery demand prediction network is also used to predict any input iterative sample delivery interaction data to generate estimated low-altitude delivery demand. For example, it predicts the appropriate drone model (considering the weight of the parts and the requirements for the environment), the estimated delivery time (considering the traffic conditions at night, the traffic rules in the industrial park, and the time of the receiving process), and the recommended flight route (avoiding areas with insufficient lighting and safety risks).

[0033] Then, the network parameter information of the temporary low-altitude distribution demand prediction network is optimized based on the network learning supervision data corresponding to the iterative sample delivery interaction data and the loss function value of the estimated low-altitude distribution demand. The network learning supervision data may be the actual delivery results, such as the actual drone used, the actual delivery time, the actual flight route, etc. By comparing the estimated low-altitude distribution demand and the actual network learning supervision data, the loss function value is calculated. If the loss function value is large, the parameters of the temporary low-altitude distribution demand prediction network are adjusted until the training termination conditions are met, thereby generating an iterative low-altitude distribution demand prediction network.

[0034] Step S140, loading the previously learned sample delivery interaction data of the iterative sample delivery interaction data into the graph convolution conversion units of the temporary low-altitude delivery demand prediction network and the iterative low-altitude delivery demand prediction network respectively, and generating corresponding first graph convolution conversion results and second graph convolution conversion results.

[0035] Taking the previously mentioned iterative sample delivery interaction data of delivering electronic parts to enterprises in an industrial park at night as an example, the prior learning sample delivery interaction data may be data collected earlier in similar delivery scenarios.

[0036] For the temporary low-altitude distribution demand prediction network, the previously learned sample distribution interaction data is loaded into the graph convolution conversion unit. For example, an early electronic parts distribution order data is loaded into the graph convolution conversion subunit corresponding to the order-related interaction data. This order data contains information such as the model, quantity, and order time of the order. The graph convolution conversion subunit performs graph convolution conversion processing on this information to generate order interaction graph features. At the same time, the distribution-related information (such as the distribution starting point, route, and environmental conditions at the time) is loaded into the graph convolution conversion subunit corresponding to the distribution-related interaction data to generate distribution interaction graph features. The basic graph convolution conversion result is generated based on the distribution interaction graph features and the order interaction graph features. This result is part of the first graph convolution conversion result. If the graph convolution conversion unit also includes a graph convolution conversion subunit corresponding to the interactive session node extraction unit, etc., it is also processed according to the corresponding rules, such as first using the interactive session node extraction unit to extract the interactive session nodes from the data, and then performing graph convolution conversion processing to finally generate the first graph convolution conversion result.

[0037] For the iterative low-altitude delivery demand prediction network, the previously learned sample delivery interaction data is also loaded into the graph convolution conversion unit. In the same way as the temporary low-altitude delivery demand prediction network, the order-related interaction data and the delivery-related interaction data are processed to generate the corresponding order interaction graph features and delivery interaction graph features, and finally generate the second graph convolution conversion result.

[0038] Step S150, based on the first graph convolution conversion result and the second graph convolution conversion result corresponding to the same sample delivery interaction data, the classification conversion unit of the iterative low-altitude delivery demand prediction network is subjected to parameter learning, and after generating an optimized temporary low-altitude delivery demand prediction network, the step of obtaining the iterative sample delivery interaction data corresponding to the next low-altitude delivery demand prediction mode is returned to until the training termination condition is met, and the target low-altitude delivery demand prediction network is generated based on the temporary low-altitude delivery demand prediction network corresponding to the training termination condition.

[0039] Assume that the same sample delivery interaction data is the complete data of a nighttime delivery of specific electronic parts to enterprises in an industrial park. The first graph convolution transformation result corresponding to this data contains the first graph convolution transformation sub-result generated by each graph convolution transformation sub-unit of the temporary low-altitude delivery demand prediction network, and the second graph convolution transformation result contains the second graph convolution transformation sub-result generated by each graph convolution transformation sub-unit of the iterative low-altitude delivery demand prediction network.

[0040] The second graph convolution transformation sub-results corresponding to the previously learned sample delivery interaction data are loaded into the corresponding classification transformation sub-units in the iterative low-altitude delivery demand prediction network to generate the corresponding prediction classification transformation sub-results. For example, loading the graph convolution transformation sub-results about order model, quantity, etc. into the corresponding classification transformation sub-unit may obtain a prediction result about the order type classification (such as the classification result of a high-value electronic parts order).

[0041] The graph convolutional transformation subunit and the classification transformation subunit that establish the mapping relationship are obtained as the target graph convolutional transformation subunit and the target classification transformation subunit. For example, the graph convolutional transformation subunit related to the order model and the classification transformation subunit for the order type classification are a pair that establish the mapping relationship.

[0042] The network learning error is calculated based on the first graph convolution conversion sub-result generated by the target graph convolution conversion sub-unit and the predicted classification conversion sub-result generated by the target classification conversion sub-unit corresponding to the same sample delivery interaction data, and the network learning error corresponding to each classification conversion sub-unit is generated. For example, if the first graph convolution conversion sub-result indicates that the order is a high-value electronic component, but the predicted classification conversion sub-result is incorrectly classified as an ordinary electronic component, a large network learning error will be generated.

[0043] According to the learning errors of each network, the corresponding classification conversion subunits in the iterative low-altitude distribution demand prediction network are optimized, such as adjusting the weight parameters in the classification conversion subunits, etc., until the training termination conditions are met, such as the network learning error is less than a preset value or a certain number of training rounds is reached, and the optimized temporary low-altitude distribution demand prediction network is generated. Then return to the step of obtaining the iterative sample distribution interaction data corresponding to the next low-altitude distribution demand prediction mode, and repeat this process until the training termination conditions are met. When the training termination conditions are met, the corresponding temporary low-altitude distribution demand prediction network becomes the target low-altitude distribution demand prediction network.

[0044] Step S160, obtaining target delivery interaction data during the low-altitude UAV delivery process, loading the target delivery interaction data into the target low-altitude delivery demand prediction network, and generating a target low-altitude delivery demand prediction result corresponding to the target delivery interaction data.

[0045] In this embodiment, it is assumed that the target delivery interaction data is a new delivery order, and the order content is to deliver a batch of medical equipment to a scientific research institution located in the suburbs. This target delivery interaction data contains detailed information of the order, such as the name, weight, volume, special transportation requirements (such as the need for constant temperature transportation) of the instrument, and the address of the scientific research institution (including its geographical coordinates, surrounding topography, such as nearby mountains that may affect the flight path of the drone), the contact information of the recipient, etc.

[0046] The target delivery interaction data is loaded into the target low-altitude delivery demand prediction network. The graph convolution conversion unit of the target low-altitude delivery demand prediction network first processes the order-related interaction data, such as the name, weight, special transportation requirements and other information of the instrument, and generates order interaction graph features. At the same time, the distribution-related interaction data (such as the distribution starting point, the address of the scientific research institution, the surrounding topography, etc.) is convolutionally converted to generate distribution interaction graph features, and the basic graph convolution conversion results are generated based on these two features. Then the classification conversion unit classifies the basic graph convolution conversion results and finally generates the target low-altitude delivery demand prediction results. This result may include the recommended drone type (considering the weight, volume and special transportation requirements of the instrument, a specially designed drone cargo cabin with a constant temperature function may be required), the estimated delivery time (considering the distance to the suburbs, the topography and possible meteorological conditions), the recommended flight route (avoiding obstacles such as mountains), etc., thereby providing a comprehensive demand forecast for the low-altitude drone delivery order.

[0047] Based on the above steps, the embodiment of the present application significantly improves the prediction accuracy of low-altitude drone delivery demand by iteratively learning and optimizing the low-altitude delivery demand prediction network. The method first uses the sample delivery interaction data sequence to perform preliminary training on the basic low-altitude delivery demand prediction network to generate a temporary low-altitude delivery demand prediction network; then, by continuously introducing new iterative sample delivery interaction data, the temporary low-altitude delivery demand prediction network is parameter-learned and optimized, and a more accurate iterative low-altitude delivery demand prediction network is gradually generated. It not only makes full use of historical delivery data, but also continuously corrects and optimizes the low-altitude delivery demand prediction network through an iterative learning mechanism, so that it can better adapt to changes in actual delivery demand. The target low-altitude delivery demand prediction network finally generated can accurately and efficiently output prediction results when facing the target delivery interaction data in the actual low-altitude drone delivery process, thereby facilitating the scheduling and planning of drone delivery, and effectively improving the efficiency and user satisfaction of drone delivery services.

[0048] In a possible implementation, step S110 includes:

[0049] Step S111, obtaining a plurality of prior delivery interaction data in the low-altitude UAV delivery process. The prior delivery interaction data is marked with a priori low-altitude delivery demand.

[0050] In this embodiment, in a low-altitude drone delivery service system covering multiple areas of the city, the prior delivery interaction data comes from a wide range of sources. Taking delivery to residential areas as an example, when a pharmacy or drug warehouse places an order for drugs through a delivery platform at a specific time, the prior delivery interaction data covers the specific time of the order. The time information accurate to the minute level reflects the pharmacy's requirements for drugs and the urgency of delivery, because different drugs have different requirements at different time periods, which is directly related to the prior low-altitude delivery demand. At the same time, the destination information of the order contains detailed geographic coordinates such as the specific floor and entrance location of the residential building where the pharmacy is located, which is crucial for the drone to accurately navigate to the destination, and this information is related to the timely and accurate delivery in the prior low-altitude delivery demand. The type and weight of the drugs in the order are also part of the prior delivery interaction data. Different types of drugs have different requirements for storage environment and transportation methods, and the weight affects the load planning of the drone. These all constitute part of the prior low-altitude delivery demand. For example, refrigerated drugs need to be stored and transported at low temperatures and have higher requirements for delivery speed. Heavier drugs may require drones with larger load capacities. At the same time, drones can be equipped with portable refrigeration or freezing packs, which use a variety of refrigeration media to achieve control of different temperature ranges, not just limited to liquid nitrogen refrigeration (liquid nitrogen refrigeration is usually used in extremely low temperature environments, such as below -180°C). Similar to the intelligent refrigeration equipment described in CN114992505A, or the concept of portable transport refrigeration devices disclosed in CN106560408A, these refrigeration or freezing packs can be set with different refrigeration or freezing partitions to meet the specific temperature control requirements of different medicines.

[0051] Step S112: clustering each priori delivery interaction data corresponding to the same priori low-altitude delivery demand to generate a first clustering sequence corresponding to each priori low-altitude delivery demand.

[0052] Continuing with the example of drug delivery, all prior delivery interaction data related to drug delivery, such as drug orders from different pharmacies, drug orders in different time periods, drug orders of different types and weights, etc., have similar prior low-altitude delivery requirements, namely, ensuring drug efficacy (refrigerated drugs) and timely delivery. These prior delivery interaction data with the same prior low-altitude delivery requirements are classified into one category to form a first cluster sequence corresponding to a prior low-altitude delivery requirement. For drug delivery, whether it is home medication delivered to residential areas or drug replenishment delivered to medical institutions, as long as it is a priori delivery interaction data related to drug delivery, it is clustered into the first cluster sequence corresponding to drug delivery due to its prior low-altitude delivery requirements for safety, timeliness, and special storage conditions.

[0053] Step S113, determining the training value corresponding to each prior delivery interaction data according to the prior low-altitude delivery demand.

[0054] For example, the training value of prior delivery interaction data for drug orders with order weights close to the maximum load of the drone and delivery time before the peak period of drug use is higher. Because the maximum load is a test of the drone's load capacity planning, and the delivery time before the peak period of drug use is more stringent, which involves more complex delivery scheduling and route planning, and can reflect more key factors of low-altitude delivery needs. For drug delivery, the training value of prior delivery interaction data for emergency drug orders is higher than that for ordinary chronic disease drug orders, because emergency drugs have extremely high requirements for delivery speed and accuracy, involving more urgent prior low-altitude delivery needs, and special route planning and safety measures may be required during the delivery process.

[0055] Step S114, taking the first clustering sequence as a low-altitude distribution demand prediction sub-mode, clustering each low-altitude distribution demand prediction sub-mode corresponding to the same training value, and generating a second clustering sequence corresponding to each training value. The second clustering sequence includes each sample distribution interaction data corresponding to the same low-altitude distribution demand prediction mode.

[0056] For example, the two low-altitude delivery demand prediction sub-modes of general medicine delivery and emergency medicine delivery with high training value may be grouped in a second cluster sequence. Although general medicine and emergency medicine are different delivery items, they have similar delivery demand characteristics such as delivery timeliness, accuracy, and special handling requirements (safe transportation of medicines) under high training value. This second cluster sequence contains the sample delivery interaction data corresponding to the same low-altitude delivery demand prediction mode. For general commodity delivery (such as daily necessities, small office supplies, etc.) and conventional medicine delivery (non-emergency) with medium training value, they will also be grouped in another second cluster sequence due to similar training value and delivery demand characteristics. The sample delivery interaction data in this sequence all reflect relatively less urgent delivery demand situations with relatively low requirements for special environments.

[0057] Step S115: generating the sample delivery interaction data sequence according to each second clustering sequence.

[0058] In this embodiment, the second cluster sequence contains sample delivery interaction data under a specific low-altitude delivery demand prediction mode, and these sample delivery interaction data together constitute a sample delivery interaction data sequence. This sample delivery interaction data sequence will be used for the subsequent construction and training of the low-altitude delivery demand prediction network, providing a comprehensive and representative data basis for accurately predicting low-altitude drone delivery demand, covering delivery interaction data under various low-altitude delivery demand scenarios from emergency to routine, from special items to ordinary items.

[0059] In a possible implementation, the sample delivery interaction data in the sample delivery interaction data sequence is annotated with network learning supervision data, the low-altitude delivery demand prediction network is initialized as a basic low-altitude delivery demand prediction network or a temporary low-altitude delivery demand prediction network, and the step of performing parameter learning on the graph convolution conversion unit of the initialized low-altitude delivery demand prediction network includes:

[0060] The initialized low-altitude delivery demand prediction network is used to predict any input sample delivery interaction data to generate an estimated low-altitude delivery demand corresponding to the any input sample delivery interaction data.

[0061] The network parameter information of the initialized low-altitude delivery demand prediction network is optimized according to the network learning supervision data corresponding to the arbitrarily input sample delivery interaction data and the loss function value of estimating the low-altitude delivery demand until the training termination conditions are met.

[0062] In this embodiment, taking the sample delivery interaction data of medicines delivered to pharmacies in commercial areas as an example, the initialized low-altitude delivery demand prediction network is used to predict the sample delivery interaction data. For the input sample delivery interaction data containing information such as order placement time, medicine type, weight, and pharmacy destination, the network extracts and analyzes these data based on the calculation logic of its internal graph convolution conversion unit to generate an estimated low-altitude delivery demand corresponding to the sample delivery interaction data. This estimated low-altitude delivery demand may include information such as recommended drone models, estimated delivery routes, and estimated delivery duration. For example, based on the weight of the medicine and the geographical location of the pharmacy, the network may estimate that a drone with a medium load capacity is required, flying along a route that avoids high-rise buildings and busy traffic areas, and is expected to be delivered within a specific time.

[0063] Then, the network parameter information of the low-altitude delivery demand prediction network is optimized and initialized according to the network learning supervision data corresponding to any input sample delivery interaction data and the loss function value of the estimated low-altitude delivery demand until the training termination condition is met. The network learning supervision data is data that accurately reflects the actual delivery situation. Taking drug delivery as an example, the network learning supervision data may include the actual drone model used, the actual flight delivery route, and the actual delivery time. The estimated low-altitude delivery demand is compared with the network learning supervision data, and the difference between the two is calculated through a specific loss function to obtain the loss function value. If the loss function value is large, it means that the prediction result of the current network deviates greatly from the actual situation, and the network parameter information needs to be optimized. For example, if the estimated delivery time is much shorter than the actual time, it may be that the parameter settings for path planning or flight speed calculation in the graph convolution conversion unit are unreasonable, so the relevant parameters are adjusted. This optimization process continues, and different sample delivery interaction data are continuously input, repeating the above process of prediction, comparison, calculation of loss function value and optimization parameters until the loss function value meets the preset conditions, such as being less than a set threshold or the change in the loss function value after several consecutive iterations is less than a certain amplitude. At this time, the training termination conditions are met, indicating that the network has learned more accurate parameters and can make more reliable predictions of low-altitude delivery demand.

[0064] In a possible implementation, the current low-altitude distribution demand prediction network is any one of a basic low-altitude distribution demand prediction network, a temporary low-altitude distribution demand prediction network, an iterative low-altitude distribution demand prediction network, and a target low-altitude distribution demand prediction network. The graph convolution conversion unit of the current low-altitude distribution demand prediction network is used to perform graph convolution conversion processing on any input sample distribution interaction data to generate a basic graph convolution conversion result. The classification conversion unit of the current low-altitude distribution demand prediction network is used to perform classification conversion on the basic graph convolution conversion result to generate a classification conversion result.

[0065] The graph convolution conversion unit includes a plurality of graph convolution conversion subunits, the classification conversion unit includes a plurality of classification conversion subunits, the graph convolution conversion subunits and the classification conversion subunits correspond to each other, the basic graph convolution conversion result includes the basic graph convolution conversion subresults generated by each graph convolution conversion subunit, and the generation step of the classification conversion result includes:

[0066] The basic graph convolution conversion sub-results generated by each graph convolution conversion sub-unit are loaded into the corresponding classification conversion sub-unit to generate the classification conversion sub-results generated by each classification conversion sub-unit.

[0067] The classification conversion result is generated according to each classification conversion sub-result.

[0068] In this embodiment, taking the basic low-altitude delivery demand prediction network as an example, when processing any input sample delivery interaction data, it is assumed that the input sample delivery interaction data is related to the delivery of medicines to a pharmacy in a commercial area, which includes order information, medicine information, pharmacy location information, and delivery time requirements. Multiple graph convolution conversion subunits in the graph convolution conversion unit begin to perform graph convolution conversion processing on these data.

[0069] For example, there is a graph convolution conversion subunit that specializes in processing order-related information. It can perform graph convolution conversion processing on data such as drug type, weight, and order placement time in the order to generate basic graph convolution conversion sub-results. This result is a feature representation of order-related information under the graph structure, which reflects certain characteristics of the order in the entire distribution interaction system, such as the urgency of the order (comprehensively judged by the order placement time, drug efficacy requirements, and pharmacy business preparation time), and the drug's requirements for delivery conditions (judged based on the drug type and weight).

[0070] At the same time, there is also a graph convolution conversion sub-unit that processes delivery-related information. It performs graph convolution conversion on data such as the pharmacy’s geographical location (including the coordinates of the commercial building, floor height, etc.), delivery time requirements, and other data, and also generates corresponding basic graph convolution conversion sub-results. This result reflects the relevant characteristics of the delivery destination and the constraints of the delivery time.

[0071] These basic graph convolution conversion sub-results generated by each graph convolution conversion sub-unit will be loaded into the corresponding classification conversion sub-unit. Each graph convolution conversion sub-unit has a corresponding classification conversion sub-unit. This correspondence is pre-set based on the network structure and is intended to perform classification conversions on different types of features. For example, after the basic graph convolution conversion sub-results generated by the graph convolution conversion sub-unit that processes order-related information are loaded into the corresponding classification conversion sub-unit, this classification conversion sub-unit will generate classification conversion sub-results about the order type classification according to its own conversion logic. The orders may be classified into different types such as emergency refrigeration orders and ordinary refrigeration orders. This classification result is obtained by converting the basic graph convolution conversion sub-results based on the order-related information through the classification conversion sub-unit.

[0072] Similarly, the classification conversion subunit corresponding to the graph convolution conversion subunit that processes delivery-related information will generate distribution-related classification conversion subresults based on the loaded basic graph convolution conversion subresults, such as classifying delivery routes into emergency delivery routes, regular delivery routes, etc. according to geographical features and time constraints.

[0073] Finally, the classification conversion result is generated based on each classification conversion sub-result. The classification conversion sub-results generated by all classification conversion sub-units, such as order type classification results, delivery path classification results, etc., are integrated to generate the classification conversion result according to the pre-set integration logic. This classification conversion result is a comprehensive description of the entire sample delivery interaction data after graph convolution conversion and classification conversion. It contains the overall classification information about the order and the comprehensive planning of delivery, which can provide a more comprehensive and accurate basis for the prediction of low-altitude drone delivery demand.

[0074] The same processing flow is followed in the temporary low-altitude distribution demand prediction network, the iterative low-altitude distribution demand prediction network, and the target low-altitude distribution demand prediction network. Regardless of the sample distribution interaction data being processed, the basic graph convolution conversion sub-results are generated through each graph convolution conversion sub-unit in the graph convolution conversion unit, and then these results are loaded into the corresponding classification conversion sub-unit to generate classification conversion sub-results. Finally, the classification conversion results are generated based on all the classification conversion sub-results, thereby achieving effective prediction of low-altitude distribution demand.

[0075] In a possible implementation, the arbitrary input sample delivery interaction data includes order-related interaction data and delivery-related interaction data, the graph convolution conversion unit includes a graph convolution conversion subunit corresponding to the order-related interaction data and a graph convolution conversion subunit corresponding to the delivery-related interaction data, and the step of generating the basic graph convolution conversion result includes:

[0076] The order-related interaction data is processed by graph convolution conversion using the graph convolution conversion subunit corresponding to the order-related interaction data to generate order interaction graph features.

[0077] The delivery-related interaction data is processed by graph convolution conversion using the graph convolution conversion subunit corresponding to the delivery-related interaction data to generate delivery interaction graph features.

[0078] The basic graph convolution conversion result is generated according to the delivery interaction graph features and the order interaction graph features.

[0079] In this embodiment, first, when processing order-related interaction data, taking the sample delivery interaction data of delivering refrigerated medicines to pharmacies in commercial areas as an example, the order-related interaction data covers information such as medicine types, weight, order placement time, customer special requirements, etc. The graph convolution conversion subunit corresponding to the order-related interaction data performs graph convolution conversion processing on these data. For example, for the types of medicines, different types of medicines have different requirements during the distribution process. For example, refrigerated medicines may require special refrigeration measures and shorter delivery time, while the relative insulation time of medium and low temperature storage medicines may be slightly longer. The graph convolution conversion subunit converts the information related to these types of medicines into a specific order interaction graph feature through graph convolution conversion processing. The order placement time is also an important order-related interaction data, which reflects the customer's expectations for the timeliness of delivery. The graph convolution conversion subunit will combine this time information with other time-related factors in the entire distribution system (such as pharmacy business preparation time, drug refrigeration period, etc.) for graph convolution conversion processing, so that the order interaction graph feature can reflect the urgency or looseness of the order in the time dimension. Special customer requirements (such as requiring delivery within a specific time period) will also be included in the scope of graph convolution transformation processing, ultimately forming an order interaction graph feature that comprehensively reflects the characteristics of order-related interaction data.

[0080] Next, the graph convolution conversion subunit corresponding to the delivery-related interaction data performs graph convolution conversion processing on the delivery-related interaction data to generate delivery interaction graph features. The delivery-related interaction data includes the delivery starting point (such as the location of the drug warehouse), the destination (the precise location of the pharmacy), the geographic information on the delivery route (surrounding buildings, terrain, and other factors that may affect the flight of the drone), and the environmental information during the delivery period (such as weather conditions, airspace control during the period, etc.). The graph convolution conversion subunit corresponding to the delivery-related interaction data performs graph convolution conversion on these data. For example, for the geographic coordinate information of the delivery starting point and destination, the graph convolution conversion subunit will analyze the distance and relative position relationship between the two points (whether it is in an area with high-rise buildings, whether it is close to a no-fly zone, etc.), and convert these geographic information into part of the delivery interaction graph features. The geographic information on the delivery route is equally important. The graph convolution conversion subunit will identify obstacles on the route (high-rise buildings, bridges, etc.), available flight channels (such as the airspace above an open square), etc., and incorporate this information into the delivery interaction graph features. Environmental information during the delivery period, such as wind speed, wind direction, visibility and other factors in weather conditions, will affect the flight stability and navigation accuracy of the drone. The airspace control situation determines the range of airspace in which the drone can fly. This information becomes part of the delivery interaction graph features after being processed by the graph convolution conversion sub-unit, thereby generating a delivery interaction graph feature that comprehensively reflects the characteristics of delivery-related interaction data.

[0081] Then, the basic graph convolution conversion result is generated based on the delivery interaction graph features and the order interaction graph features. This basic graph convolution conversion result is the result of combining the features of order-related interaction data and delivery-related interaction data. For example, the combination of the drug weight in the order interaction graph features and the information related to the drone load capacity in the delivery interaction graph features can determine whether batch delivery is required or whether a drone with a specific load capacity is selected; the combination of the order placement time in the order interaction graph features and the traffic conditions of the delivery route in the delivery interaction graph features (comprehensively judged by geographic information and airspace control conditions) can determine a more reasonable delivery sequence and time schedule. The basic graph convolution conversion result will be used as the input of the subsequent classification conversion unit, providing an important data basis for low-altitude delivery demand prediction.

[0082] In a possible implementation, the current low-altitude delivery demand prediction network includes an interactive session node extraction unit, the graph convolution conversion unit includes a graph convolution conversion subunit corresponding to the interactive session node extraction unit and a graph convolution conversion subunit corresponding to any input sample delivery interaction data, and the step of generating the basic graph convolution conversion result includes:

[0083] The interactive conversation node extraction unit is used to extract interactive conversation nodes from the arbitrarily input sample delivery interactive data, and to generate delivery-related interactive data corresponding to the arbitrarily input sample delivery interactive data.

[0084] The graph convolution conversion subunit corresponding to the interactive session node extraction unit is used to perform graph convolution conversion processing on the delivery-related interaction data to generate delivery interaction graph features.

[0085] The graph convolution conversion subunit corresponding to the arbitrary input sample delivery interaction data is used to perform graph convolution conversion processing on the arbitrary input sample delivery interaction data to generate order interaction graph features.

[0086] The basic graph convolution conversion result is generated according to the delivery interaction graph features and the order interaction graph features.

[0087] In this embodiment, the delivery of medicines to pharmacies in commercial areas is still taken as an example. The sample delivery interaction data contains various interactive information from the order placement to the delivery completion. The interactive session node extraction unit will extract key interactive node information related to delivery, such as the interactive node between the order placement and the warehouse stocking, which may involve the time when the order information is transmitted to the warehouse management system, the priority of the warehouse stocking, etc.; the interactive node between the completion of the warehouse stocking and the take-off of the drone, including the safety inspection and loading time of the goods in the process of loading the drone; the interactive node between the take-off of the drone and the arrival at the destination, involving the update of navigation information during the flight, the interaction with the ground monitoring system, etc.; the interactive node between the arrival at the destination and the customer receiving the goods, such as the method of notifying the customer to pick up the goods, the time limit for the customer to pick up the goods, etc. The information contained in these interactive session nodes constitutes the delivery-related interactive data.

[0088] Next, the graph convolution conversion subunit corresponding to the interactive session node extraction unit performs graph convolution conversion processing on the delivery-related interactive data to generate delivery interaction graph features. For each interactive session node, such as the interactive node between order placement and warehouse stocking, the graph convolution conversion subunit analyzes factors such as the accuracy of order information transmission, the relationship between warehouse stocking priority and other orders, and converts these factors into part of the delivery interaction graph features. For the interactive node between warehouse stocking completion and drone takeoff, the graph convolution conversion subunit considers factors such as the efficiency of cargo loading and the impact of safety inspection results on subsequent flights, so that the delivery interaction graph features can fully reflect the characteristics of delivery-related interaction data at the interactive session node level.

[0089] At the same time, the graph convolution conversion subunit corresponding to any input sample delivery interaction data performs graph convolution conversion processing on any input sample delivery interaction data to generate order interaction graph features. This process is similar to the previous processing of order-related interaction data separately, but this time it starts from the perspective of the entire sample delivery interaction data and comprehensively considers all order-related information, including the interactive relationship between order-related interaction data and delivery-related interaction data at the order level. For example, the relationship between the order placement time and the warehouse stocking completion time will affect the overall timeliness of the order, and this relationship will be incorporated into the generation process of the order interaction graph features.

[0090] Finally, the basic graph convolution conversion result is generated based on the delivery interaction graph features and the order interaction graph features. This result combines the features of the delivery-related interaction data extracted from the interactive session nodes and the order-related features in the entire sample delivery interaction data. For example, the interaction node information between the drone takeoff and arrival at the destination (such as flight time, emergency handling during flight, etc.) in the delivery interaction graph features is combined with the order timeliness requirements in the order interaction graph features to more accurately determine the key time nodes and resource allocation plans in the entire delivery process. This basic graph convolution conversion result provides more comprehensive and detailed input data for subsequent classification conversion units, which helps to improve the accuracy and effectiveness of low-altitude delivery demand forecasting.

[0091] In a possible implementation, the delivery-related interaction data includes at least one interaction session node, the delivery interaction graph feature includes an interaction session vector corresponding to each interaction session node, and generating the basic graph convolution conversion result according to the delivery interaction graph feature and the order interaction graph feature includes:

[0092] Step A110 , assigning focus weights to the interaction conversation vectors corresponding to the respective interaction conversation nodes according to the order interaction graph features, and generating focus weight information corresponding to the respective interaction conversation nodes.

[0093] Step A120, performing a standardized conversion on the focus weight information corresponding to each interactive conversation node to generate an interactive weight coefficient corresponding to each interactive conversation node.

[0094] In this embodiment, it is assumed that after the focus weight distribution, the focus weight information obtained by the three interactive conversation nodes is 0.96, 0.7 and 1.2 respectively. In order to enable these weights to work better together in subsequent calculations, a standardized conversion is required. A common method of standardized conversion is to divide each focus weight information by the sum of all focus weight information. Here, the sum is 0.96 + 0.7 + 1.2 = 2.86. Then the interaction weight coefficient corresponding to the first interactive conversation node is 0.96 / 2.86 ≈ 0.336, the interaction weight coefficient corresponding to the second interactive conversation node is 0.7 / 2.86 ≈ 0.245, and the interaction weight coefficient corresponding to the third interactive conversation node is 1.2 / 2.86 ≈ 0.42.

[0095] Step A130: generating optimized delivery interaction graph features according to the interaction session vectors and interaction weight coefficients corresponding to the interaction session nodes.

[0096] Take the interactive conversation node between order placement and warehouse stocking mentioned above as an example. Its interactive conversation vector is combined with the interactive weight coefficient to obtain the optimized vector element. Assuming that the time delay element of the order delivery to the warehouse in the interactive conversation vector is 5 minutes (this is just for the convenience of illustration, the actual situation may be a more complex numerical representation), the corresponding interactive weight coefficient is 0.336, then the optimized element value is 5 * 0.336 = 1.68 minutes (the calculation result here is only indicative, and the actual calculation may be different according to the specific vector and coefficient representation method). This calculation is performed for each element in the interactive conversation vector of each interactive conversation node, so as to obtain the optimized distribution interaction graph feature, which comprehensively considers the importance weight of each interactive conversation node in the entire distribution system.

[0097] Step A140, generating the basic graph convolution conversion result according to the order interaction graph features and the optimized delivery interaction graph features.

[0098] In this embodiment, the information such as the type and weight of the medicine in the order interaction graph feature is combined with the information of each interactive conversation node in the optimized delivery interaction graph feature. For example, the weight of the medicine in the order interaction graph feature is 10 kg, and the interactive conversation node between the take-off and arrival of the drone at the destination in the optimized delivery interaction graph feature contains the load capacity information of the drone. Combining these two pieces of information can determine whether the drone needs to be transported in batches or adjust the flight attitude and other decision information. These comprehensive information constitutes the basic graph convolution conversion result, which provides an important data basis for the subsequent classification conversion unit and helps to more accurately predict the demand for low-altitude delivery.

[0099] In a possible implementation, step A110 includes:

[0100] Step A111 , feature decomposition is performed on the order interaction graph features to generate a plurality of sub-feature sets with different semantic information, and each interaction session vector is vectorized to determine the interaction session information represented by each vector element.

[0101] In this embodiment, taking the delivery of medicines to pharmacies in commercial areas as an example, the order interaction graph features include information such as the type of medicine, weight, order placement time, and special customer requirements, which together constitute a complex order interaction graph feature. Feature decomposition of the order interaction graph features will generate multiple sub-feature sets with different semantic information. For example, information related to the type of medicine can constitute a sub-feature set, which may include semantic information such as refrigeration requirements and fragility of low-temperature refrigerated medicines; information related to the order placement time can constitute another sub-feature set, which reflects the timeliness semantic information of the order. At the same time, the interactive session vectors corresponding to each interactive session node are vectorized to determine the interactive session information represented by each vector element. Assuming that there is an interactive session node between order placement and warehouse stocking in the distribution link, the vector elements of its interactive session vector may include interactive session information such as the time delay of the order being delivered to the warehouse and the initial response time after the warehouse receives the order.

[0102] Step A112, for each interactive conversation node, based on semantics and logical relationships, the interactive conversation vector corresponding to the interactive conversation node is matched with each sub-feature set implementing the order interaction graph feature to obtain an association analysis result, which records the degree of association and association type between the interactive conversation vector of each interactive conversation node and each sub-feature set implementing the order interaction graph feature.

[0103] Taking the interactive conversation node between order placement and warehouse stocking as an example, the time delay of the order being delivered to the warehouse in the interactive conversation vector has a strong correlation with the order placement time sub-feature set in the order interaction graph feature. The type of this correlation is time logic correlation, and the degree of correlation may be high, because the time delay of the order being delivered to the warehouse directly affects the timeliness of the order placement time. However, the initial response time of the warehouse after receiving the order in the interactive conversation vector may have a weak correlation with the drug type sub-feature set, and the correlation type is indirect impact correlation, because the drug type may affect the complexity of warehouse stocking, thereby indirectly affecting the initial response time. Through such correlation matching, the correlation analysis results of the degree of correlation and the type of correlation between the interactive conversation vector of each interactive conversation node and each sub-feature set of the order interaction graph feature are obtained.

[0104] Step A113, based on the association analysis result, assign a corresponding initial weight value to the association relationship between the interaction session vector of each interaction session node and each sub-feature set of the order interaction graph feature to obtain an initial weight assignment result.

[0105] For example, for the interactive conversation node between order placement and warehouse stocking, the correlation between its interactive conversation vector and the sub-feature set of order placement time may be assigned a higher initial weight value, such as 0.8, due to the high correlation degree; while the correlation with the sub-feature set of drug types may be assigned a lower initial weight value, such as 0.2, due to the low correlation degree. Such initial weight allocation is performed on the correlation between each interactive conversation node and each sub-feature set of the order interaction graph feature, thereby obtaining the entire initial weight allocation result.

[0106] Step A114, evaluates the initial weight distribution result to determine the weight adjustment factor, and generates a set of weight adjustment factors, wherein the position of each interactive session node in the distribution link, the frequency of interaction with other interactive session nodes, and the potential impact on the final distribution result are specifically analyzed, the importance of each interactive session node in the entire distribution interaction system is evaluated, and the weight adjustment factor is determined according to the importance of each interactive session node.

[0107] In the distribution link, the position of each interactive conversation node is different, and its importance is also different. For example, the interactive conversation node between the drone take-off and arrival at the destination is at the core of the entire distribution process, because this node is directly related to whether the distribution can be successfully completed, and this node has a high frequency of interaction with other interactive conversation nodes, such as the interactive conversation node between the completion of warehouse stocking and the drone take-off, and the interactive conversation node between the arrival at the destination and the customer receiving the goods. Moreover, this node has a huge potential impact on the final delivery result, and any abnormality during the flight may lead to delivery failure. Therefore, the importance of this interactive conversation node is high, and a larger weight adjustment factor is determined based on its importance, such as 1.5. Although the interactive conversation node between order placement and warehouse stocking is important, its impact on the final delivery result is not as direct as the interactive conversation node between the drone take-off and arrival at the destination. Its importance is slightly lower, and a smaller weight adjustment factor is determined, such as 1.2. By analyzing the position of each interactive conversation node in the distribution link, the interaction frequency with other interactive conversation nodes, and the potential impact on the final distribution result, the importance of each interactive conversation node in the entire distribution interaction system is evaluated, thereby determining the weight adjustment factor of each interactive conversation node and generating a set of weight adjustment factors.

[0108] Step A115, using a set of weight adjustment factors to adjust the initial weight distribution result to obtain a final weight distribution result, wherein the final weight distribution result includes focus weight information corresponding to each interactive session node, wherein, specifically for each interactive session node, the initial weight value between the interactive session node and each sub-feature set of the order interaction graph feature is multiplied by the corresponding weight adjustment factor to obtain the focus weight information corresponding to the interactive session node.

[0109] In this embodiment, the initial weight distribution result is adjusted using a set of weight adjustment factors to obtain the final weight distribution result, that is, the focus weight information corresponding to each interactive session node. For the interactive session node between order placement and warehouse stocking, the initial weight value between it and the order placement time sub-feature set is 0.8, and the corresponding weight adjustment factor is 1.2. The initial weight value is multiplied by the weight adjustment factor to obtain an adjusted weight value of 0.96, which is the focus weight information between the interactive session node and the order placement time sub-feature set. Such adjustments are made to the association relationship between each interactive session node and each sub-feature set of the order interaction graph feature to obtain the final weight distribution result, which contains the focus weight information corresponding to each interactive session node.

[0110] In a possible implementation, the current sample delivery interaction data is subjected to graph convolution conversion processing by the current graph convolution conversion subunit to generate a graph convolution conversion result corresponding to the current sample delivery interaction data, including:

[0111] The current graph convolution conversion subunit is used to perform forward graph convolution conversion processing on the current sample delivery interaction data to generate a forward graph convolution conversion result.

[0112] The current graph convolution conversion subunit is used to perform backward graph convolution conversion processing on the current sample delivery interaction data to generate a backward graph convolution conversion result.

[0113] A graph convolution conversion result corresponding to the current sample delivery interaction data is generated based on the forward graph convolution conversion result and the backward graph convolution conversion result.

[0114] In this embodiment, taking the common low-altitude drone delivery scenario of delivering medicines to pharmacies in commercial areas as an example, the current sample delivery interaction data contains many aspects of information, such as the detailed content of the order (type of medicines, weight, order placement time, etc.), delivery-related information (warehouse location, pharmacy location, geographical and environmental information on the delivery route, etc.) and other factors that may affect delivery (such as special customer requirements, airspace control conditions, etc.).

[0115] First, the current graph convolution conversion subunit is used to perform forward graph convolution conversion processing on the current sample delivery interaction data to generate a forward graph convolution conversion result. In this process, the graph convolution conversion subunit will process the current sample delivery interaction data according to the pre-set forward graph convolution algorithm. For the drug type information in the order, the graph convolution conversion subunit will gradually transfer information and extract features from the previous nodes (such as information nodes related to the source of drug supply, if any) to the subsequent nodes (such as nodes related to warehouse stocking, drone delivery, etc.) based on its internal convolution kernel and related weight parameters. Assuming that the drug is a low-temperature refrigerated drug, it has high refrigeration requirements and strong delivery timeliness. The forward graph convolution conversion will propagate and convert the relevant refrigeration requirements, delivery priority and other information along the forward direction of the delivery process based on this characteristic of low-temperature refrigerated drugs. For example, at the node related to warehouse stocking, the forward graph convolution conversion result may be reflected in the characteristics that the warehouse needs to give priority to the order and needs special refrigeration equipment for stocking. For the order placement time, the forward graph convolution transformation will take into account the time constraints in the entire delivery process, such as the time sequence from the order placement to the warehouse preparation completion, the drone take-off, and the arrival at the destination. The urgency of the order placement time and other information will be gradually transmitted to the nodes related to the subsequent delivery links, generating a forward graph convolution transformation result containing these time feature transmission results.

[0116] Next, the current graph convolution conversion subunit is used to perform backward graph convolution conversion processing on the current sample delivery interaction data to generate a backward graph convolution conversion result. Unlike the forward graph convolution conversion, the backward graph convolution conversion is to transfer information and extract features from the post-order node to the forward node. Taking the delivery of refrigerated medicines to pharmacies as an example, starting from the end point of delivery, that is, the location of the pharmacy and the link where the customer receives the goods, the backward graph convolution conversion subunit will pass back information such as the business preparation time of the pharmacy and the receiving time required by the customer. For example, if the pharmacy requires the goods to be received before a specific time, the backward graph convolution conversion will reversely pass this time limit information to the drone delivery link, affecting the drone's flight speed, route planning and other decisions, and will be reflected in the graph convolution conversion result as constraint features related to flight speed and route. For the warehouse stocking link, the backward graph convolution conversion will reversely determine the latest time for warehouse stocking, the quantity and type accuracy of stocking, etc. according to the final delivery requirements, such as the time requirement for arriving at the pharmacy, and generate a backward graph convolution conversion result containing these reverse information transfer results.

[0117] Finally, the graph convolution conversion result corresponding to the current sample delivery interaction data is generated based on the forward graph convolution conversion result and the backward graph convolution conversion result. This comprehensive graph convolution conversion result integrates all the information of the forward and backward graph convolution conversions. For example, the information about the drug refrigeration requirements and order placement time in the forward graph convolution conversion result is combined with the information about the pharmacy receiving time and the latest warehouse stocking time in the backward graph convolution conversion result. If the forward graph convolution conversion result shows that the refrigerated drugs need to be delivered in a short time and the order placement time is early, and the backward graph convolution conversion result shows that the pharmacy requires receiving the goods at a specific early time and the warehouse stocking time is also early, then the comprehensive graph convolution conversion result may reflect the characteristics that the entire delivery process needs to be highly compactly arranged, the drone needs to choose the shortest and safest route, fly at a faster speed, and the warehouse needs to complete high-quality stocking in the shortest time. This graph convolution conversion result comprehensively reflects the feature representation of the current sample delivery interaction data after graph convolution conversion, provides an important data basis for subsequent classification conversion and other operations, and helps to accurately predict the demand for low-altitude drone delivery.

[0118] In a possible implementation, the first graph convolution conversion result includes a first graph convolution conversion sub-result generated by each graph convolution conversion sub-unit of the temporary low-altitude distribution demand prediction network, and the second graph convolution conversion result includes a second graph convolution conversion sub-result generated by each graph convolution conversion sub-unit of the iterative low-altitude distribution demand prediction network. Step S150 includes:

[0119] Step S151, loading each second graph convolution conversion sub-result corresponding to the previously learned sample delivery interaction data into the corresponding classification conversion sub-unit in the iterative low-altitude delivery demand prediction network to generate corresponding prediction classification conversion sub-results.

[0120] In this embodiment, a typical low-altitude drone delivery scenario of delivering refrigerated medicines to pharmacies in commercial areas is taken as an example. The sample delivery interaction data involved contains rich information, such as order information (drug type, weight, order placement time, etc.), delivery-related information (warehouse location, pharmacy location, geographical and environmental conditions of the delivery route, etc.) and other factors affecting delivery (customer special requirements, airspace control conditions, etc.).

[0121] First, assume that the previously learned sample delivery interaction data is the data accumulated from the previous multiple delivery of refrigerated drugs to the pharmacy. For the second graph convolution conversion sub-result generated by the graph convolution conversion sub-unit that processes order-related information (such as drug type and weight, etc.), it is loaded into the classification conversion sub-unit that specializes in processing order type classification in the iterative low-altitude delivery demand prediction network. This classification conversion sub-unit processes the input second graph convolution conversion sub-result according to its own algorithm and parameters to generate a predicted classification conversion sub-result about the order type. For example, if the weight of the drug in the input second graph convolution conversion sub-result is light and it is a common room temperature drug, the classification conversion sub-unit may generate a predicted classification conversion sub-result of a common refrigerated order type. For the second graph convolution conversion sub-result generated by the graph convolution conversion sub-unit that processes delivery-related information (such as warehouse location and pharmacy location, etc.), after being loaded into the corresponding classification conversion sub-unit of the delivery path classification, it may generate a predicted classification conversion sub-result of a regular delivery path type based on the relative position of the warehouse and the pharmacy, the surrounding geographical environment and other factors, indicating that the delivery path is relatively simple and has no special geographical obstacles or restrictions.

[0122] Step S152, obtaining the graph convolution conversion subunit and the classification conversion subunit that establish the mapping relationship as the target graph convolution conversion subunit and the target classification conversion subunit.

[0123] In the structural design of the low-altitude delivery demand prediction network, there is a pre-set mapping relationship between the graph convolutional conversion subunit and the classification conversion subunit. For example, a mapping relationship is established between the graph convolutional conversion subunit that processes order-related information such as drug type and weight and the classification conversion subunit that classifies the order type, and they are determined as the target graph convolutional conversion subunit and the target classification conversion subunit. This mapping relationship is constructed based on the logic of delivery demand prediction and is intended to ensure the correct transmission and processing of data from graph convolutional conversion to classification conversion.

[0124] Step S153, calculating the network learning error based on the first graph convolution conversion sub-result generated by the target graph convolution conversion sub-unit corresponding to the same sample delivery interaction data and the predicted classification conversion sub-result generated by the target classification conversion sub-unit, and generating the network learning error corresponding to each classification conversion sub-unit.

[0125] Continuing with the example of refrigerated drug delivery, for the target graph convolutional conversion subunit and the target classification conversion subunit for order type classification, if the first graph convolutional conversion sub-result generated by the target graph convolutional conversion subunit based on the temporary low-altitude delivery demand prediction network for processing the sample delivery interaction data indicates that the order is a high-value refrigerated order (possibly because the drug is a special drug and is heavy), and the predicted classification conversion sub-result generated by the target classification conversion subunit based on the iterative low-altitude delivery demand prediction network is a common refrigerated order type, then there is a classification error. This difference is calculated by a specific error calculation method (such as a cross entropy loss function, etc.) to obtain the network learning error corresponding to the classification conversion subunit for order type classification. For the target graph convolutional conversion subunit and the target classification conversion subunit for delivery path classification, if the first graph convolutional conversion sub-result indicates that a special delivery path is required due to a special geographical environment (such as a newly built high-rise building or temporary traffic control), and the predicted classification conversion sub-result is a regular delivery path type, the network learning error corresponding to the classification conversion subunit for delivery path classification is also obtained by the error calculation method.

[0126] Step S154, optimizing the corresponding classification conversion subunits in the iterative low-altitude distribution demand prediction network according to each network learning error until the training termination condition is met, thereby generating an optimized temporary low-altitude distribution demand prediction network.

[0127] Finally, for example, for the classification conversion subunit for order type classification, if the calculated network learning error is large, it means that its classification result is inaccurate, and the parameters of the classification conversion subunit need to be adjusted. For example, the weight parameters used to judge the order type inside the classification conversion subunit may be adjusted to increase the weights of features related to drug value, weight, etc., so as to improve the recognition ability of high-value refrigerated orders. For the classification conversion subunit for distribution path classification, the weight parameters related to factors such as geographical environment and traffic control are adjusted according to the calculated network learning error to more accurately judge special distribution paths. This optimization process is repeated continuously, and different prior learning sample distribution interaction data is input each time to calculate the network learning error and adjust the parameters of the classification conversion subunit. When the network learning error meets certain conditions, such as the overall error is less than the preset threshold or the error change after multiple consecutive iterations is less than a certain amplitude, the training termination condition is met. At this time, the iterative low-altitude distribution demand prediction network becomes the optimized temporary low-altitude distribution demand prediction network after optimization. This optimized temporary low-altitude distribution demand forecasting network can more accurately predict low-altitude distribution demand, improve distribution efficiency and accuracy, and reduce various problems that may be caused by inaccurate distribution demand forecasts, such as drug expiration and delivery delays.

[0128] Figure 2The hardware structure of the low-altitude UAV delivery demand prediction system 100 based on the AI ​​prediction model for implementing the above-mentioned low-altitude UAV delivery demand prediction method based on the AI ​​prediction model provided by an embodiment of the present invention is shown. Figure 2 As shown, the low-altitude drone delivery demand prediction system 100 based on the AI ​​prediction model may include a processor 110, a machine-readable storage medium 120, a bus 130, and a communication unit 140.

[0129] 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 data and / or instructions that the low-altitude drone delivery demand prediction system 100 based on the AI ​​prediction model uses to execute or use to complete the exemplary method described in the present invention.

[0130] During the specific implementation process, one or more processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the low-altitude drone delivery demand prediction method based on the AI ​​prediction model in the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0131] The specific implementation process of the processor 110 can be found in the various method embodiments executed by the low-altitude drone delivery demand prediction system 100 based on the above-mentioned AI prediction model. The implementation principles and technical effects are similar, and this embodiment will not be repeated here.

[0132] 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 low-altitude UAV delivery demand prediction method based on the AI ​​prediction model as mentioned above is implemented.

[0133] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, various features are sometimes combined into one embodiment, drawing or description thereof.

Claims

1. A method for predicting low-altitude UAV delivery demand based on AI prediction model, characterized in that: The method comprises: Acquire a sample delivery interaction data sequence in a low-altitude UAV delivery process; the sample delivery interaction data sequence includes sample delivery interaction data corresponding to a plurality of low-altitude delivery demand prediction modes; Obtain candidate sample delivery interaction data corresponding to the candidate low-altitude delivery demand prediction mode, perform parameter learning on the graph convolution conversion unit of the basic low-altitude delivery demand prediction network according to the candidate sample delivery interaction data, and generate a temporary low-altitude delivery demand prediction network; Acquire iterative sample delivery interaction data corresponding to the next low-altitude delivery demand prediction mode, perform parameter learning on the graph convolution conversion unit of the temporary low-altitude delivery demand prediction network according to the iterative sample delivery interaction data, and generate an iterative low-altitude delivery demand prediction network; Loading the previously learned sample delivery interaction data of the iterative sample delivery interaction data into the graph convolution conversion unit of the temporary low-altitude delivery demand prediction network and the iterative low-altitude delivery demand prediction network, respectively, to generate corresponding first graph convolution conversion results and second graph convolution conversion results; According to the first graph convolution conversion result and the second graph convolution conversion result corresponding to the same sample delivery interaction data, parameter learning is performed on the classification conversion unit of the iterative low-altitude delivery demand prediction network. After generating an optimized temporary low-altitude delivery demand prediction network, return to the step of obtaining the iterative sample delivery interaction data corresponding to the next low-altitude delivery demand prediction mode until the training termination condition is met, and generate a target low-altitude delivery demand prediction network according to the temporary low-altitude delivery demand prediction network corresponding to the training termination condition; Obtaining target delivery interaction data during low-altitude UAV delivery, loading the target delivery interaction data into a target low-altitude delivery demand prediction network, and generating a target low-altitude delivery demand prediction result corresponding to the target delivery interaction data; The graph convolution conversion unit of the basic low-altitude distribution demand prediction network, the temporary low-altitude distribution demand prediction network, the iterative low-altitude distribution demand prediction network or the target low-altitude distribution demand prediction network is used to perform graph convolution conversion processing on any input sample distribution interaction data to generate a basic graph convolution conversion result; the classification conversion unit of the basic low-altitude distribution demand prediction network, the temporary low-altitude distribution demand prediction network, the iterative low-altitude distribution demand prediction network or the target low-altitude distribution demand prediction network is used to perform classification conversion on the basic graph convolution conversion result to generate a classification conversion result.

2. The method for predicting low-altitude UAV delivery demand based on AI prediction model according to claim 1 is characterized in that: The method of obtaining a sample delivery interaction data sequence in the low-altitude UAV delivery process includes: Acquire a plurality of prior delivery interaction data in a low-altitude UAV delivery process; the prior delivery interaction data is annotated with a prior low-altitude delivery demand; Clustering each priori delivery interaction data corresponding to the same priori low-altitude delivery demand to generate a first clustering sequence corresponding to each priori low-altitude delivery demand; Determine the training value corresponding to each prior delivery interaction data based on the prior low-altitude delivery demand; The first clustering sequence is used as a low-altitude distribution demand prediction sub-mode, and each low-altitude distribution demand prediction sub-mode corresponding to the same training value is clustered to generate a second clustering sequence corresponding to each training value; the second clustering sequence includes each sample distribution interaction data corresponding to the same low-altitude distribution demand prediction mode; The sample delivery interaction data sequence is generated according to each second clustering sequence.

3. The low-altitude UAV delivery demand prediction method based on AI prediction model according to claim 1 is characterized in that: The sample delivery interaction data in the sample delivery interaction data sequence is annotated with network learning supervision data. When the basic low-altitude delivery demand prediction network or the temporary low-altitude delivery demand prediction network is used as the initialized low-altitude delivery demand prediction network, the step of performing parameter learning on the graph convolution conversion unit of the initialized low-altitude delivery demand prediction network includes: Using the initialized low-altitude delivery demand prediction network to predict any input sample delivery interaction data, and generate an estimated low-altitude delivery demand corresponding to the any input sample delivery interaction data; The network parameter information of the initialized low-altitude delivery demand prediction network is optimized according to the network learning supervision data corresponding to the arbitrarily input sample delivery interaction data and the loss function value of estimating the low-altitude delivery demand until the training termination conditions are met.

4. The method for predicting low-altitude UAV delivery demand based on AI prediction model according to claim 1 is characterized in that: The graph convolution conversion unit includes a plurality of graph convolution conversion subunits, the classification conversion unit includes a plurality of classification conversion subunits, the graph convolution conversion subunits correspond to the classification conversion subunits respectively, the basic graph convolution conversion result includes the basic graph convolution conversion subresults generated by each graph convolution conversion subunit, and the generation step of the classification conversion result includes: Load the basic graph convolution conversion sub-results generated by each graph convolution conversion sub-unit into the corresponding classification conversion sub-unit to generate the classification conversion sub-results generated by each classification conversion sub-unit; The classification conversion result is generated according to each classification conversion sub-result.

5. The method for predicting low-altitude UAV delivery demand based on AI prediction model according to claim 4 is characterized in that: The arbitrary input sample delivery interaction data includes order-related interaction data and delivery-related interaction data, the graph convolution conversion unit includes a graph convolution conversion subunit corresponding to the order-related interaction data and a graph convolution conversion subunit corresponding to the delivery-related interaction data, and the step of generating the basic graph convolution conversion result includes: Using the graph convolution conversion subunit corresponding to the order-related interaction data to perform graph convolution conversion processing on the order-related interaction data to generate order interaction graph features; Using the graph convolution conversion subunit corresponding to the delivery-related interaction data to perform graph convolution conversion processing on the delivery-related interaction data to generate delivery interaction graph features; The basic graph convolution conversion result is generated according to the delivery interaction graph features and the order interaction graph features.

6. The method for predicting low-altitude UAV delivery demand based on AI prediction model according to claim 4 is characterized in that: The basic low-altitude distribution demand prediction network, the temporary low-altitude distribution demand prediction network, the iterative low-altitude distribution demand prediction network or the target low-altitude distribution demand prediction network includes an interactive session node extraction unit, the graph convolution conversion unit includes a graph convolution conversion subunit corresponding to the interactive session node extraction unit and a graph convolution conversion subunit corresponding to any input sample distribution interaction data, and the step of generating the basic graph convolution conversion result includes: Extracting interactive conversation nodes from the arbitrarily input sample delivery interactive data using the interactive conversation node extraction unit to generate delivery-related interactive data corresponding to the arbitrarily input sample delivery interactive data; Using the graph convolution conversion subunit corresponding to the interactive session node extraction unit to perform graph convolution conversion processing on the delivery-related interaction data to generate delivery interaction graph features; Using the graph convolution conversion subunit corresponding to the arbitrary input sample delivery interaction data, the arbitrary input sample delivery interaction data is subjected to graph convolution conversion processing to generate order interaction graph features; The basic graph convolution conversion result is generated according to the delivery interaction graph features and the order interaction graph features.

7. The method for predicting low-altitude UAV delivery demand based on an AI prediction model according to claim 5 or 6, characterized in that: The delivery-related interaction data includes at least one interaction session node, the delivery interaction graph feature includes an interaction session vector corresponding to each interaction session node, and generating the basic graph convolution conversion result according to the delivery interaction graph feature and the order interaction graph feature includes: According to the order interaction graph features, focusing weights are assigned to the interaction session vectors corresponding to the respective interaction session nodes to generate focusing weight information corresponding to the respective interaction session nodes; Performing standardized conversion on the focus weight information corresponding to each interactive conversation node to generate an interactive weight coefficient corresponding to each interactive conversation node; Generate optimized delivery interaction graph features according to the interaction session vectors and interaction weight coefficients corresponding to each interaction session node; Generating the basic graph convolution conversion result according to the order interaction graph features and the optimized delivery interaction graph features; The step of assigning focus weights to the interaction session vectors corresponding to the respective interaction session nodes according to the order interaction graph features to generate focus weight information corresponding to the respective interaction session nodes comprises: Decomposing the order interaction graph features to generate a plurality of sub-feature sets with different semantic information, and vectorizing each interaction session vector to determine the interaction session information represented by each vector element; For each interactive conversation node, based on semantics and logical relationships, the interactive conversation vector corresponding to the interactive conversation node is matched with each sub-feature set implementing the order interaction graph feature to obtain an association analysis result, which records the degree of association and the type of association between the interactive conversation vector of each interactive conversation node and each sub-feature set implementing the order interaction graph feature; Based on the association analysis result, a corresponding initial weight value is assigned to the association relationship between the interaction session vector of each interaction session node and each sub-feature set of the order interaction graph feature to obtain an initial weight assignment result; The initial weight distribution result is evaluated to determine the weight adjustment factor, and a weight adjustment factor set is generated, wherein the position of each interactive session node in the delivery link, the interaction frequency with other interactive session nodes, and the potential impact on the final delivery result are specifically analyzed, the importance of each interactive session node in the entire delivery interaction system is evaluated, and the weight adjustment factor is determined according to the importance of each interactive session node; The initial weight distribution result is adjusted using a set of weight adjustment factors to obtain a final weight distribution result, wherein the final weight distribution result includes focus weight information corresponding to each interactive session node, wherein specifically for each interactive session node, the initial weight value between the interactive session node and each sub-feature set of the order interaction graph feature is multiplied by the corresponding weight adjustment factor to obtain the focus weight information corresponding to the interactive session node.

8. The method for predicting low-altitude UAV delivery demand based on an AI prediction model according to claim 5 or 6, characterized in that: The current sample delivery interaction data is subjected to graph convolution conversion processing by the current graph convolution conversion subunit to generate a graph convolution conversion result corresponding to the current sample delivery interaction data, including: Using the current graph convolution conversion subunit to perform forward graph convolution conversion processing on the current sample delivery interaction data to generate a forward graph convolution conversion result; Using the current graph convolution conversion subunit to perform backward graph convolution conversion processing on the current sample delivery interaction data to generate a backward graph convolution conversion result; A graph convolution conversion result corresponding to the current sample delivery interaction data is generated based on the forward graph convolution conversion result and the backward graph convolution conversion result.

9. The method for predicting low-altitude UAV delivery demand based on AI prediction model according to claim 4 is characterized in that: The first graph convolution conversion result includes the first graph convolution conversion sub-result generated by each graph convolution conversion sub-unit of the temporary low-altitude distribution demand prediction network, and the second graph convolution conversion result includes the second graph convolution conversion sub-result generated by each graph convolution conversion sub-unit of the iterative low-altitude distribution demand prediction network. The first graph convolution conversion result and the second graph convolution conversion result corresponding to the same sample distribution interaction data are used to perform parameter learning on the classification conversion unit of the iterative low-altitude distribution demand prediction network to generate an optimized temporary low-altitude distribution demand prediction network, including: Loading each second graph convolution conversion sub-result corresponding to the previously learned sample delivery interaction data into the corresponding classification conversion sub-unit in the iterative low-altitude delivery demand prediction network to generate corresponding prediction classification conversion sub-results; Acquire the graph convolutional transformation subunit and the classification transformation subunit that establish the mapping connection as the target graph convolutional transformation subunit and the target classification transformation subunit; The network learning error is calculated based on the first graph convolution conversion sub-result generated by the target graph convolution conversion sub-unit corresponding to the same sample delivery interaction data and the predicted classification conversion sub-result generated by the target classification conversion sub-unit, and the network learning error corresponding to each classification conversion sub-unit is generated; According to the learning errors of each network, the corresponding classification conversion subunits in the iterative low-altitude distribution demand prediction network are optimized until the training termination conditions are met to generate an optimized temporary low-altitude distribution demand prediction network.

10. A low-altitude drone delivery demand prediction system based on AI prediction model, characterized in that: The low-altitude UAV delivery demand prediction system based on the AI ​​prediction model 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 low-altitude UAV delivery demand prediction method based on the AI ​​prediction model as described in any one of claims 1 to 9 above.

Images