A low-altitude aircraft cluster task scheduling method and system based on big data

Abstract

The present application relates to the technical field of low-altitude aircraft scheduling, and particularly relates to a low-altitude aircraft cluster task scheduling method and system based on big data, which comprises the following steps: generating a predicted delivery time length through a time sequence feature hybrid model based on historical data of a low-altitude aircraft distribution big data platform, wherein the time sequence feature hybrid model is constructed based on LSTM and Transformer network architecture; constructing an operational decision model with the optimization objective of minimizing the average of multiple predicted delivery time lengths and the decision variable of order allocation mode; and solving the operational decision model based on a multi-objective search strategy to determine the order allocation mode of the low-altitude aircraft cluster. The present application realizes the balance between instant response and global efficiency in low-altitude aircraft cluster scheduling, improves the adaptability of dynamic delay scheduling strategy, and improves the delivery efficiency and reduces the overtime rate.

Description

Technical Field

[0001] This invention relates to the field of low-altitude aircraft scheduling technology, and in particular to a low-altitude aircraft cluster task scheduling method and system based on big data. Background Technology

[0002] With the rapid development of low-altitude aircraft technology, low-altitude aircraft clusters are increasingly being used in logistics distribution, emergency response and other fields. The "low-altitude economy" has emerged as an new industry and is actively becoming a new growth engine.

[0003] Currently, the booming development of the express delivery and logistics industry brought about by the growth of e-commerce platforms has led to problems such as warehouse congestion and delivery delays in low-altitude aircraft swarms. This is because most current algorithms for operating low-altitude aircraft swarms are based on biomimicry, controlling the swarms by simulating the behavior of biological clusters. For example, the adaptive weighted pigeon colony algorithm is used for low-altitude aircraft route planning, achieving both quality and efficiency; ant colony algorithms are also used for low-altitude aircraft logistics route planning.

[0004] However, the main optimization direction of the above algorithm is to control the point-to-point operation of low-altitude aircraft. If it is a multi-point to multi-point operation, it cannot be planned on the local algorithm and a global planning algorithm is required.

[0005] Currently, global planning algorithms face significant challenges in achieving efficient and flexible task scheduling, especially given dynamically changing order demands, complex environmental conditions, and multi-objective optimization requirements. Traditional low-altitude aircraft scheduling methods based on global planning algorithms typically employ static threshold strategies and allocation rules based on a single metric.

[0006] Existing static threshold strategies, such as greedy strategies, involve assigning orders as soon as they appear. Traditional greedy strategies provide a rapid response but may be inefficient. They only represent one factor to consider when modeling an optimization problem involving matching decisions for a given batch of orders. In reality, the goal is not to achieve the optimal outcome in a single decision, but rather to maximize the cumulative benefit of the strategy over a period of time.

[0007] Existing allocation rules based on a single metric, such as wave frequency rules, are problematic. Wave frequency rules trigger scheduling based on a fixed accumulated order volume or optimize only delivery distance. These methods struggle to adapt to fluctuations in order arrival rates, dynamic changes in the delivery environment (such as weather and airspace restrictions), and multi-dimensional optimization needs (such as timeliness, energy consumption, and timeout risks). This can easily lead to uneven resource allocation, response delays, or local optima.

[0008] Therefore, the industry currently tends to use delayed scheduling strategies to eliminate the drawbacks of the low efficiency of greedy strategies in overall scheduling and allocation, and to eliminate the drawbacks of uneven resource allocation based on allocation rules with single indicators by combining intelligent operations optimization with machine learning models.

[0009] However, traditional delay scheduling strategies are generally divided into quantitative and timed strategies. Quantitative strategies involve making order assignment decisions after a specific number of orders have been generated, while timed strategies involve making order assignment decisions at fixed time intervals. Although quantitative and timed strategies have slightly slower response times, they can improve efficiency by accumulating orders to find the optimal time for order assignment. However, quantitative strategies may result in longer time delays and slower response times.

[0010] Therefore, how to implement low-altitude aircraft order allocation decisions and dynamic delay scheduling strategies that combine machine learning models is a technical problem that needs to be solved. Summary of the Invention

[0011] To address this, the present invention provides a low-altitude aircraft cluster task scheduling method and system based on big data. By using a hybrid model of time-series features based on LSTM and Transformer network architecture, the delivery time predicted by the machine learning model is used for low-altitude aircraft order allocation decisions and dynamic delay scheduling strategies. This achieves a balance between immediate response and global efficiency in low-altitude aircraft cluster scheduling, improves the adaptability of dynamic delay scheduling strategies, enhances delivery efficiency, and reduces timeout rates.

[0012] To achieve the above objectives, this invention proposes a low-altitude aircraft swarm task scheduling method based on big data, comprising:

[0013] Based on the comparison between the current order accumulation and the dynamic initial allocation threshold, as well as the comparison between the historical delivery time and the dynamic time threshold, determine whether to execute the low-altitude aircraft cluster task scheduling strategy.

[0014] If the task scheduling strategy is executed, the historical data of the low-altitude aircraft delivery big data platform will be used to generate a predicted delivery time through a time-series feature fusion model. The time-series feature fusion model is built on a network architecture based on Transformer and LSTM. The Transformer is used to extract global features, and the LSTM is used to extract time-series trend features.

[0015] An operations research decision model is constructed with the optimization objectives of minimizing the average of multiple predicted delivery times, minimizing the average flight distance per order, and minimizing the timeout rate, and with the order allocation method as the decision variable, and with completion time constraints constructed based on the predicted delivery times.

[0016] The operation research decision model is solved based on a multi-objective search strategy to determine the order allocation method of the low-altitude aircraft cluster, and the dynamic initial allocation threshold and the dynamic duration threshold for the next execution are determined based on the predicted delivery time.

[0017] Furthermore, the process of generating predicted delivery times includes:

[0018] The historical data is preprocessed and feature-stitched to generate fused features. Historical order arrival rates and historical delivery times are standardized and feature-stitched to generate time-series features.

[0019] The fused features are processed through a Transformer cross-feature unit to perform multi-head attention calculations, generating a global feature for order duration.

[0020] The time-series features are used to generate time-series trend features through an LSTM time-series feature encoding unit;

[0021] The global features of order duration and the time-series trend features are used to generate predicted delivery duration through a feature fusion mapping unit.

[0022] Furthermore, the process of generating global features for order duration includes:

[0023] The fused features are processed by a multi-head attention computation subunit to perform multi-head attention computation and multi-head concatenation, thereby generating multi-head features.

[0024] The multi-head features are residually connected and normalized with the fused features through a single residual connection subunit to generate residual features;

[0025] The residual features are used to calculate the global correlation of features through feedforward network subunits to generate globally correlated features;

[0026] The global correlation feature is connected to the residual feature through a secondary residual connection subunit and normalized to generate a global feature of order duration.

[0027] Furthermore, the delivery time trend strength, delivery time periodic fluctuation coefficient, delivery time sudden fluctuation coefficient, and time-series delivery time feature are weighted and biased through a multi-layer fully connected layer with ReLU activation function set in the feature fusion mapping unit to generate biased predicted delivery time.

[0028] The global feature of order duration and the biased predicted delivery duration are used to generate a spliced ​​feature through the feature splicing layer of the feature fusion mapping unit;

[0029] The spliced ​​features are fused and mapped through a fully connected network with a ReLU activation function set in the feature fusion mapping unit to generate a predicted delivery time.

[0030] Furthermore, the construction process of the temporal feature fusion model includes:

[0031] A prediction error term is constructed based on the mean squared error of the training predicted values ​​and the actual values ​​of delivery time;

[0032] A waiting cost term is constructed based on the training prediction value of delivery time and the optimal waiting time.

[0033] A comprehensive loss function is constructed based on the weighted sum of the prediction error term and the waiting cost term, and the comprehensive loss function is used for the optimization training of the time-series feature hybrid model.

[0034] In the above scheme, the implicit correlation between global delivery route environment features, time period features, and low-altitude aircraft order allocation status is mined through the multi-head attention mechanism of Transformer cross feature units to identify their impact on delivery time. The trend and periodic features of local historical order arrival rate and delivery time are extracted through LSTM. The global and local features are integrated through the bias prediction mechanism, for example, more attention is paid to local time series changes during morning and evening peak periods, thus achieving better business adaptability in delivery time prediction.

[0035] Furthermore, the process of constructing completion time constraints based on predicted delivery duration includes:

[0036] The completion time constraint is constructed based on the pickup and delivery times of two orders executed sequentially by the low-altitude aircraft in the order allocation method and the predicted delivery time of the earlier order.

[0037] Furthermore, the process of solving the operations research decision-making model based on a multi-objective search strategy to determine the order allocation method for the low-altitude aircraft cluster includes:

[0038] After randomly removing the removal order and delivery time of the first low-altitude aircraft from the initial solution, the removal order is assigned to the second low-altitude aircraft with the highest idle rate to generate a temporary solution;

[0039] Calculate the fitness score of the temporary solution;

[0040] If the fitness score is greater than or equal to the score threshold, then the temporary solution is used as a candidate output solution, and the temporary solution is iteratively calculated.

[0041] If the fitness score is less than the score threshold, the temporary solution is iteratively calculated to generate candidate output solutions using the Metropolis criterion, and then the temporary solution is iteratively calculated.

[0042] When the number of iterations of the temporary solution equals the maximum number of iterations, the candidate output solution is used as the order allocation method.

[0043] Furthermore, the process of calculating the fitness score of the temporary solution includes:

[0044] Calculate operator selection weights based on the total score of the order operator iteration using the decay parameter and temporary solution;

[0045] The operator selection probability is calculated based on the ratio of the operator selection weight to the total operator selection weight;

[0046] The fitness score is calculated based on the operator selection probability, the total score of the order operator iteration, and the previous score of the order operator.

[0047] Furthermore, the process of determining the dynamic initial allocation threshold and the dynamic duration threshold for the next execution based on the predicted delivery time includes:

[0048] The product of the predicted delivery time and the current order density is used as the dynamic initial allocation threshold;

[0049] Calculate the ratio of total delivery time to total order arrival rate, and calculate the dynamic time threshold based on the minimum of the predicted delivery time and the ratio;

[0050] The process of determining whether to execute the task scheduling strategy of the low-altitude aircraft cluster includes: if the current accumulated order amount is less than the dynamic initial allocation threshold, or the historical delivery time is less than the dynamic duration threshold, then the task scheduling strategy of the low-altitude aircraft cluster is executed.

[0051] This invention also provides a system for scheduling low-altitude aircraft swarm missions using a big data-based method, comprising:

[0052] The acquisition module is used to determine whether to execute the task scheduling strategy of the low-altitude aircraft cluster based on the comparison results of the current accumulated order volume and the dynamic initial allocation threshold, as well as the comparison results of the historical delivery time and the dynamic time threshold.

[0053] The delivery time prediction module, connected to the acquisition module, is used to generate a predicted delivery time from the historical data of the low-altitude aircraft delivery big data platform through a time-series feature fusion model if the task scheduling strategy is executed. The time-series feature fusion model is built on a network architecture based on Transformer and LSTM fusion, where Transformer is used to extract global features and LSTM is used to extract time-series trend features.

[0054] The operations research decision model construction module is connected to the delivery time prediction module and is used to construct an operations research decision model with the optimization objectives of minimizing the average of multiple predicted delivery times, minimizing the average flight distance per order, and minimizing the timeout rate, the order allocation method as the decision variable, and the completion time constraint based on the predicted delivery time.

[0055] The solution module is connected to the operations research decision model construction module and the acquisition module. It is used to solve the operations research decision model based on a multi-objective search strategy, determine the order allocation method of the low-altitude aircraft cluster, and determine the dynamic initial allocation threshold and the dynamic duration threshold for the next execution based on the predicted delivery time.

[0056] The above scheme further improves the feasibility, solution efficiency, and scenario adaptability of low-altitude aircraft scheduling by using refined constraint modeling, adaptive search algorithm optimization, and dynamic threshold collaborative feedback.

[0057] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0058] 1. By using a hybrid model of temporal features based on LSTM and Transformer network architecture, the delivery time predicted by the machine learning model is used for order allocation decisions and dynamic delay scheduling strategies for low-altitude aircraft. This achieves a balance between immediate response and global efficiency in low-altitude aircraft cluster scheduling, improves the adaptability of dynamic delay scheduling strategies, enhances delivery efficiency, and reduces timeout rates.

[0059] 2. By leveraging the multi-head attention mechanism of Transformer cross-feature units, we can uncover the implicit correlations between global delivery route environmental features, time period features, and low-altitude aircraft order allocation status, and identify their impact on delivery time. We can also extract the trend and periodic features of local historical order arrival rates and delivery times using LSTM, and fuse global and local features through a bias prediction mechanism. For example, we can pay more attention to local time-series changes during morning and evening peak hours, thus achieving better business adaptability in delivery time prediction.

[0060] 3. By using refined constraint modeling, adaptive search algorithm optimization, and dynamic threshold collaborative feedback, the feasibility, solution efficiency, and scenario adaptability of low-altitude aircraft scheduling are further improved. Attached Figure Description

[0061] Figure 1 This is a flowchart illustrating the low-altitude aircraft swarm task scheduling method and system based on big data, according to an embodiment of the present invention.

[0062] Figure 2 This is a schematic diagram of the structure of the low-altitude aircraft swarm task scheduling method and system based on big data, according to an embodiment of the present invention.

[0063] Figure 3 This is a schematic diagram of the overall system architecture of the low-altitude aircraft swarm task scheduling method and system based on big data, according to an embodiment of the present invention.

[0064] Figure 4 This is a schematic diagram illustrating the system order information and low-altitude aircraft information flow of the low-altitude aircraft cluster task scheduling method and system based on big data, as described in an embodiment of the present invention.

[0065] Figure 5 This is a comparative diagram of multiple order allocation methods for the low-altitude aircraft cluster task scheduling method and system based on big data, according to an embodiment of the present invention.

[0066] Figure 6 This is a schematic diagram of the operation and decision-making model of the low-altitude aircraft cluster task scheduling method and system based on big data, according to an embodiment of the present invention. Detailed Implementation

[0067] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0068] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0069] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0070] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0071] like Figures 1 to 6 As shown, this invention provides a low-altitude aircraft cluster task scheduling method and system based on big data. By using a time-series feature hybrid model based on LSTM and Transformer network architecture, the delivery time predicted by the machine learning model is used for low-altitude aircraft order allocation decisions and dynamic delay scheduling strategies. This achieves a balance between immediate response and global efficiency in low-altitude aircraft cluster scheduling, improves the adaptability of dynamic delay scheduling strategies, enhances delivery efficiency, and reduces timeout rates.

[0072] like Figures 1 to 6 As shown, this embodiment proposes a low-altitude aircraft swarm task scheduling method based on big data, including:

[0073] Based on the comparison between the current order accumulation and the dynamic initial allocation threshold, as well as the comparison between the historical delivery time and the dynamic time threshold, determine whether to execute the low-altitude aircraft cluster task scheduling strategy.

[0074] If the task scheduling strategy is executed, the historical data of the low-altitude aircraft delivery big data platform will be used to generate a predicted delivery time through a time-series feature fusion model. The time-series feature fusion model is built on a network architecture based on Transformer and LSTM. The Transformer is used to extract global features, and the LSTM is used to extract time-series trend features.

[0075] An operations research decision model is constructed with the optimization objectives of minimizing the average of multiple predicted delivery times, minimizing the average flight distance per order, and minimizing the timeout rate, and with the order allocation method as the decision variable, and with completion time constraints constructed based on the predicted delivery times.

[0076] The operation research decision model is solved based on a multi-objective search strategy to determine the order allocation method of the low-altitude aircraft cluster, and the dynamic initial allocation threshold and the dynamic duration threshold for the next execution are determined based on the predicted delivery time.

[0077] It should be noted that the historical data includes historical order arrival rate, historical delivery time, time period characteristics, delivery route environmental characteristics, current order accumulation, and current low-altitude aircraft order allocation status.

[0078] It is understood that the low-altitude aircraft delivery big data platform is an instant delivery big data platform, and the low-altitude aircraft cluster task scheduling method described in this embodiment is based on, for example, Figure 3 The overall architecture shown is constructed in which the instant delivery big data platform integrates the machine learning framework TensorFlow. In TensorFlow, the temporal feature hybrid model based on LSTM and Transformer network architecture described in this embodiment is set up, and the operations research optimization part is set up by programming the operations research decision model.

[0079] Furthermore, such as Figure 6 As shown, the order allocation method includes allocating a single set of low-altitude aircraft orders, the order execution order of low-altitude aircraft, and the pickup and delivery times of low-altitude aircraft. The process of constructing the operations research decision model also includes:

[0080] The completion time constraint is constructed based on the pickup and delivery times of two orders executed sequentially by the low-altitude aircraft and the predicted delivery time of the earlier order.

[0081] It should be noted that in the operations research decision model described in this embodiment, the independent variables are obtained from the real-time delivery big data platform and operations research optimization platform of the current time period, which are the fixed values ​​input to the operations research decision model. The order allocation method, for example, aircraft 1 is responsible for orders 1 and 3, and aircraft 2 is responsible for orders 2, 4 and 5, which corresponds to the values ​​of each decision variable. For example, if aircraft 1 is responsible for order 1, the arrival time of order 1 at the pickup point can be determined according to the predicted delivery time of aircraft 1. For example, the pickup time of order 1 allocated to aircraft 1 (pos1 to pos2) is 08:15, and the delivery time is 08:15 + predicted delivery time = 08:28. Decision variables refer to variables that can be controlled and adjusted in the operations research decision model. The purpose of the operations research decision model described in this embodiment is to solve or optimize the decision variables under the constraints of the constraints, with the optimization objective as the optimization direction. The specific solution optimization algorithm adopts a multi-objective search strategy.

[0082] Specifically, such as Figure 6 As shown, the optimization objectives of the operations research decision-making model include:

[0083] Objective 1: Minimize average time per unit:

[0084] ;

[0085] In the formula, the minimized To optimize the direction, the number of low-altitude aircraft, m, used to determine the minimum average time per order delivery task for the entire cluster is minimized. It is a low-altitude aircraft Execution order of assigned tasks The largest predicted delivery time , Right now The average delivery time per low-altitude aircraft. It's understandable that this is not due to the number of low-altitude aircraft, as the delivery routes of multiple orders may overlap or be close together. The more orders, the shorter the average delivery time. Instead, the same low-altitude aircraft is needed to perform multiple delivery tasks with similar delivery routes.

[0086] For example, suppose there are delivery locations k21, k22, and k23, and a pickup location k1. The distance between k21 and k1 is 4km, and the distance between k21 and k22 is 7km. The delivery path from k23 to k1 overlaps with the delivery path from k22 to k1, with the overlapping distances being 5km between k22 and k1, 5km between k23 and k22, and 8km between k23 and k1. If the number of low-altitude aircraft m is 1, then the average delivery time per order is... If m is 2, then the average time per unit is If m is 3, then the average time per unit is Q represents the time per kilometer (the assumption that the predicted delivery time is proportional to the time per kilometer is for ease of understanding Objective 1; in reality, the predicted delivery time takes into account factors such as weather conditions, route complexity, and airspace control levels in this embodiment, and is not necessarily proportional to Q). Therefore, when m is 2, Objective 1 minimizes the average delivery time per unit. Thus, Objective 1 can balance the relationship between delivery efficiency and the number of low-altitude aircraft, avoiding excessively low delivery efficiency (e.g., when m is 1 above), and k1 also avoids an excessive number of low-altitude aircraft and excessively high costs (e.g., when m is 3 above).

[0087] Objective 2: Minimize average flight distance per unit:

[0088] ;

[0089] In the formula, Pickup location to delivery location The navigation distance for a given low-altitude aircraft j (from 1 to m). The decision variable represents the order in which low-altitude aircraft j is assigned tasks, in order to minimize the average flight distance per unit of the entire cluster. To optimize the direction, reasonable decision variables are determined by solving for the optimal order allocation method. The number of low-altitude aircraft, m, is used to optimize the order allocation routes, reduce costs, and improve efficiency. For example, under the same assumptions, it can be understood that the difference between objective 2 and objective 1 is that the predicted delivery time in the average order consumption of objective 1 takes into account factors such as weather conditions, route complexity, and airspace control level in this embodiment, while objective 2 does not take into account factors such as weather conditions, but only considers the optimal allocation of routes. Objective 1 and objective 2 are used as optimization objectives at the same time, realizing the comprehensive consideration of predicted delivery time and actual route length.

[0090] Objective 3: Minimize the timeout rate.

[0091] ;

[0092] Minimizing the average delivery time per unit means minimizing the average of the predicted delivery times.

[0093] In the formula, As an indicator function, according to the existing definition of an indicator function, when order i satisfies hour, The value is 1 when order i does not satisfy the condition. hour, The value is 0, therefore Indicates n orders The number of timeouts, of which The delivery time of order i, which is the decision variable, is determined by a simple conversion based on the predicted delivery time. , Let the desired delivery time of order i be the independent variable. Therefore, the overall formula represents an optimization direction focused on minimizing the timeout rate. By solving for the optimized order allocation method, the formula is modified accordingly. The value of is used to determine the reasonable allocation of low-altitude aircraft orders for delivery.

[0094] The constraints for modeling the operations research decision-making model include:

[0095] Constraint 1: Delivery of an order must be made after pickup.

[0096] ;

[0097] In the formula, The location of the pickup item is the independent variable. to delivery location Navigation distance, The independent variable is the flight speed of the low-altitude aircraft j. The pickup time of order i represents the decision variable. The arrival time of order i at the delivery point represents the decision variable. This represents the delivery time of order i, which is the independent variable. Let i represent the delivery time of order i, which is the decision variable.

[0098] Constraint 2: Order pickup can only proceed after both the goods have been prepared and the drone has arrived.

[0099] ;

[0100] In the formula, This represents the order placement time of order i, which is the independent variable. This represents the stock preparation time for order i, which is the independent variable. The pickup time of order i represents the decision variable. Let represent the arrival time of order i at the delivery point, which is the decision variable.

[0101] Constraint 3: Time limits for completing multiple tasks assigned to the same drone:

[0102] ;

[0103] In the formula, The delivery time of order i1 represents the decision variable. The delivery time of order i2 represents the decision variable. This represents the distance from the delivery location of order i1 to the pickup location of order i2.

[0104] Constraint 4: The pickup and delivery of an order are completed by the same drone:

[0105] ;

[0106] In the formula, Let order i be the pickup task, which is the independent variable. Let order i be the pickup task, which is the independent variable. The set of tasks assigned to low-altitude aircraft j represents the decision variable.

[0107] Therefore, it is understandable that the above constraints can restrict the range of values ​​for each decision variable.

[0108] The time constraint refers to the time limit for completing multiple tasks assigned to the same low-altitude aircraft.

[0109] The above formulas include the following independent variable: n: the number of orders for the low-altitude aircraft cluster; The time when order i was placed; The desired delivery time for order i; :The preparation time for order i; The flight speed of the low-altitude aircraft j: Delivery time for order i; The current position of low-altitude aircraft j; Order i pickup task; :Order i Delivery task; : All task collections; : Order i Pickup location; : Order i Delivery location; Pickup location to delivery location Navigation distance.

[0110] Understandably, all independent variables are readily available. Figure 3 The timely delivery big data platform and operations optimization platform shown illustrate the process for obtaining specific low-altitude aircraft status (number of low-altitude aircraft, flight speed, delivery time, current location) and order information (order quantity, order placement time, etc.) as follows: Figure 4 As shown.

[0111] The above formulas include the following decision variables: : The set of tasks assigned to low-altitude aircraft j; m: The number of low-altitude aircraft in the low-altitude aircraft cluster used to perform order delivery tasks; The order of missions assigned to low-altitude aircraft J; The arrival time of order i at the pickup point; The pickup time for order i; The arrival time of order i at the delivery point; : Delivery time of order i, where the difference between the arrival time at the pickup point and the arrival time at the delivery point is the delivery time.

[0112] Furthermore, the process of determining the dynamic initial allocation threshold and the dynamic duration threshold for the next execution based on the predicted delivery time includes:

[0113] The product of the predicted delivery time and the current order density is used as the dynamic initial allocation threshold;

[0114] Calculate the ratio of total delivery time to total order arrival rate, and calculate the dynamic time threshold based on the minimum of the predicted delivery time and the ratio;

[0115] The process of determining whether to execute the task scheduling strategy of the low-altitude aircraft cluster includes: if the current accumulated order amount is less than the dynamic initial allocation threshold, or the historical delivery time is less than the dynamic duration threshold, then the task scheduling strategy of the low-altitude aircraft cluster is executed.

[0116] Specifically, the process of calculating the dynamic duration threshold is as follows:

[0117] ;

[0118] In the formula, Indicates the dynamic duration threshold. Indicates the adjustment factor. Indicates the predicted delivery time. This represents the total delivery time of all low-altitude aircraft in the low-altitude aircraft cluster. This indicates the total order fulfillment rate.

[0119] If the dynamic start allocation threshold and dynamic duration threshold are fixed values, then traditional quantitative and timing strategies are employed. Traditional quantitative strategies include... Figure 5 As shown in (B), if a delayed scheduling strategy is implemented, then once there exists an order i that is forcibly satisfied... That is, when the pickup time of order i is greater than the desired delivery time of order i, optimization objective 3 is lost. Minimizing the timeout rate increases with the number of orders (n) in the low-altitude aircraft cluster, leading to longer order allocation delays and slower response times. Traditional timing strategies, such as... Figure 5 As shown in (C), while ensuring optimization objective 1 Under the condition of minimum average time per unit, the number n may vary greatly in different time periods, which will affect the set of tasks assigned to low-altitude aircraft j. Significant fluctuations lead to an imbalance in the allocation of low-altitude aircraft resources, resulting in high operational pressure on low-altitude aircraft during peak periods and long standby times during off-peak periods, making it difficult to distinguish between peak and off-peak order periods.

[0120] In addition, such as Figure 5 The greedy algorithm shown in (A) makes it difficult to achieve high order allocation efficiency while taking into account the overall situation when determining the order allocation.

[0121] Therefore, this embodiment controls the dynamic initial allocation threshold and dynamic duration threshold through machine learning, i.e. Figure 5 (D) The order begins to be assigned a dividing line, which can ensure that the order is assigned a dividing line. That is, when the pickup time of order i is later than the desired delivery time of order i, machine learning is used to control the number of orders n. The maximum delivery time is used as the historical delivery time and then fed into the time series feature hybrid model for the next batch of orders for learning and optimization. The data of the next batch is then input into the time series feature hybrid model to obtain the execution time of the next batch of orders, i.e., the predicted delivery time. This can achieve a balance between immediate response and global efficiency, allocate orders through a dynamic starting allocation threshold during peak periods, and allocate orders through a dynamic duration threshold by increasing the dynamic starting allocation threshold during off-peak periods, thereby shortening the waiting time window and preventing order backlog.

[0122] Furthermore, the temporal feature fusion model includes a Transformer cross-feature unit, an LSTM temporal feature encoding unit, and a feature fusion mapping unit. The process of generating the predicted delivery time includes:

[0123] The route environment features, time period features, and current low-altitude aircraft order allocation status are preprocessed and feature splicing to generate fused features. The historical order arrival rate and historical delivery time are standardized and feature splicing to generate time-series features.

[0124] The fused features are processed through a Transformer cross-feature unit to perform multi-head attention calculations, generating a global feature for order duration.

[0125] The time-series features are used to generate time-series trend features through an LSTM time-series feature encoding unit;

[0126] The global features of order duration and the time-series trend features are used to generate predicted delivery duration through a feature fusion mapping unit.

[0127] Understandably, the Transformer cross-feature unit can capture cross-feature correlations, such as the impact of weather on a specific time period, and the LSTM temporal feature encoding unit can be adjusted based on historical order arrival rates and other temporal information. Preferably, it is fine-tuned and optimized every 15 minutes using 200 data points updated by the low-altitude aircraft delivery big data platform.

[0128] Specifically, the route environment features include weather conditions (sunny / rainy / snowy, etc.), and the time period features include off-peak / peak periods. Both are preprocessed using a one-hot coding algorithm to generate embedding vectors. The current low-altitude aircraft order allocation status, including the current cumulative order volume, low-altitude aircraft idle rate, and the remaining response time of the first order, are preprocessed using standardization to generate a time series. The embedding vectors are independently embedded into low-dimensional vectors, and the low-dimensional vectors and time series are concatenated to generate fused features.

[0129] More specifically, the route environmental characteristics include four main categories: weather conditions, route complexity, airspace control level, and electromagnetic interference intensity. Weather conditions include weather type (sunny / rainy / snowy / foggy), wind speed, visibility, and precipitation intensity. These are discretized and embedded to generate a 16-dimensional vector, including a 4-dimensional vector corresponding to sunny / rainy / snowy / foggy conditions, and 4-dimensional vectors each corresponding to the maximum, average, variance, and duration of exceeding the aircraft speed limit threshold for wind speed / visibility / precipitation intensity. Route complexity includes the route altitude change rate and route obstacle density, which are normalized using MinMax and logarithmic transform to generate a 2-dimensional vector. Airspace control levels range from 0 to 9, which are one-hot encoded to generate a 10-dimensional vector. Electromagnetic interference intensity includes high-altitude electromagnetic interference intensity (0-100%) and low-altitude electromagnetic interference intensity (0-100%), which are processed using Sigmoid to generate 2-dimensional vectors. It is understood that the obstacles in the line obstacle density include buildings and power grid equipment. Low-altitude aircraft fly at an altitude of less than 100 meters and will encounter more buildings and power grid equipment during the order delivery process.

[0130] The time period features include hourly segments, weekday / weekend markers, holiday markers, and seasonal features. Specifically, the hourly segments are encoded using sine and cosine cycles to generate a 4-dimensional vector (including morning delivery peak, noon delivery peak, evening delivery peak, and weekdays). The weekday / weekend segments are encoded using one-hot encoding to generate a 4-dimensional vector. The holiday markers are encoded using binary encoding to generate a 2-dimensional vector. The seasonal features are mapped and encoded to generate an 8-dimensional vector (including spring, summer, autumn, winter, and four transitional periods).

[0131] The low-altitude aircraft is preferably a drone. The current low-altitude aircraft order allocation status includes the proportion of idle drones, average battery remaining, current load rate, and order waiting queue. The proportion of idle drones is Z-score normalized to generate a 2D vector, the average battery remaining is discretized and embedded to generate a 12D vector, the current load rate is compressed using arctangent transformation to generate a 2D vector, and the order waiting queue is logarithmically normalized to generate a 2D vector.

[0132] In summary, after adjusting the vector dimensions by embedding the above vectors, feature concatenation is performed to generate 128-dimensional fused features. The embedding process is implemented by defining multiple trainable embedding layers in the temporal feature fusion model.

[0133] Furthermore, the Transformer cross-feature unit includes a multi-head attention computation subunit, a first-order residual connection subunit, a feedforward network subunit, and a second-order residual connection subunit. The process of generating global features of order duration includes:

[0134] The fused features are processed by a multi-head attention computation subunit to perform multi-head attention computation and multi-head concatenation, thereby generating multi-head features.

[0135] The multi-head features are residually connected and normalized with the fused features through the first residual connection subunit to generate residual features;

[0136] The residual features are globally correlated through the feedforward network subunit to generate globally correlated features.

[0137] The global correlation feature is connected to the residual feature through the secondary residual connection subunit and normalized to generate the global feature of order duration.

[0138] Preferably, the multi-head attention computation subunit is configured with 4 heads, a model dimension of 32, a feedforward network hidden layer dimension ffn dim of 64, an overfitting prevention parameter dropout rate of 0.1, and a stacking layer number of transformerlayers of 2.

[0139] Specifically, the primary residual connection subunit is represented as:

[0140] ;

[0141] In the formula, Representing residual characteristics, Indicates residual connection, Indicates fusion features, This indicates a random deactivation operation. This indicates a multi-headed characteristic.

[0142] Specifically, the feedforward network subunits implement nonlinear mapping of the feature space through the ReLU activation function.

[0143] Specifically, the secondary residual connection subunit is represented as:

[0144] ;

[0145] In the formula, This represents the global feature of order duration. Indicates residual connection, Representing residual characteristics, This indicates a random deactivation operation. This indicates a globally related feature.

[0146] Furthermore, the time-series trend features include delivery time trend strength, delivery time periodic fluctuation coefficient, delivery time sudden fluctuation coefficient, and time-series delivery time features. The feature fusion mapping unit includes a multi-layer fully connected layer, a feature splicing layer, and a fully connected network. The process of generating the predicted delivery time includes:

[0147] The delivery time trend strength, delivery time periodic fluctuation coefficient, delivery time sudden fluctuation coefficient, and time-series delivery time features are weighted and biased by a multi-layer fully connected layer with ReLU activation function to generate biased predicted delivery time.

[0148] The global feature of order duration and the biased predicted delivery duration are used to generate a spliced ​​feature through a feature splicing layer;

[0149] The spliced ​​features are fused and mapped through a fully connected network with a ReLU activation function to generate a predicted delivery time.

[0150] Specifically, the historical order arrival rate is the number of orders delivered by low-altitude aircraft per hour, and the historical delivery duration is the actual dispatch interval of the past 5 windows (corresponding to the morning delivery peak, noon delivery peak, evening delivery peak, morning-noon normal, and noon-evening normal). The forget gate of the LSTM temporal feature encoding unit can be used to discard data during the order dispatch trough period. The cell state update is used to retain the long-term patterns of the trough and peak periods and update the short-term changes, thereby enabling the output gate to output effective information of the temporal trend features.

[0151] Specifically, the process of standardizing and concatenating the historical order arrival rate and historical delivery time to generate time-series features is as follows: the historical order arrival rate, historical delivery time, and order change gradient are respectively processed by Z-score standardization to reflect the order density, historical delivery efficiency, and order change trend in the current period. The time periods of the five windows are sinusoidally encoded, and the data after Z-score standardization / sinusoidal period encoding are concatenated to generate time-series features.

[0152] Preferably, the LSTM temporal feature encoding unit has 64 neurons, the input sequence length is 24 time steps, each time step has 2 features, and the dropout ratio is 0.1. Specifically, neurons 0 to 15 are used to capture the trend strength of delivery time as a short-term fluctuation, neurons 16 to 31 are used to capture the burst fluctuation coefficient of delivery time as a daily cycle pattern, neurons 32 to 47 are used to capture the periodic fluctuation coefficient of delivery time as a daily cycle pattern, and neurons 48 to 63 are used to capture the temporal feature delivery time characteristic as a response to abnormal events.

[0153] Preferably, the first fully connected layer of the multi-layer fully connected network has 64 neurons and an output dimension of 64; the second fully connected layer has 32 neurons and an output dimension of 32; and the third fully connected layer has 1 neuron for primary regression prediction to generate a biased predicted delivery time. The fully connected network also has 1 neuron for secondary regression prediction to generate a predicted delivery time.

[0154] Furthermore, the construction process of the temporal feature fusion model includes:

[0155] A prediction error term is constructed based on the mean squared error of the training predicted values ​​and the actual values ​​of delivery time;

[0156] A waiting cost term is constructed based on the training prediction value of delivery time and the optimal waiting time.

[0157] A comprehensive loss function is constructed based on the weighted sum of the prediction error term and the waiting cost term, and the comprehensive loss function is used for the optimization training of the time-series feature hybrid model.

[0158] Specifically, the comprehensive loss function is:

[0159] ;

[0160] In the formula, This represents two weighting coefficients. Training prediction value representing delivery time and the true value The mean square error, This indicates a waiting cost item.

[0161] The waiting cost item is as follows:

[0162] ;

[0163] In the formula, Indicates the waiting cost item. For the sample size, The training prediction value represents the delivery time. This indicates the optimal waiting time. For example, if the training prediction value is 5 minutes, then the optimal waiting time is 3 minutes, thereby penalizing overly conservative prediction results.

[0164] In the above scheme, the implicit correlation between global delivery route environment features, time period features, and low-altitude aircraft order allocation status is mined through the multi-head attention mechanism of Transformer cross feature units to identify their impact on delivery time. The trend and periodic features of local historical order arrival rate and delivery time are extracted through LSTM. The global and local features are integrated through the bias prediction mechanism, for example, more attention is paid to local time series changes during morning and evening peak periods, thus achieving better business adaptability in delivery time prediction.

[0165] Furthermore, the process of solving the operations research decision-making model based on a multi-objective search strategy to determine the order allocation method for the low-altitude aircraft cluster includes:

[0166] After randomly removing the removal order and delivery time of the first low-altitude aircraft from the initial solution, the removal order is assigned to the second low-altitude aircraft with the highest idle rate to generate a temporary solution;

[0167] Calculate the fitness score of the temporary solution;

[0168] If the fitness score is greater than or equal to the score threshold, then the temporary solution is used as a candidate output solution, and the temporary solution is iteratively calculated.

[0169] If the fitness score is less than the score threshold, the temporary solution is iteratively calculated to generate candidate output solutions using the Metropolis criterion, and then the temporary solution is iteratively calculated.

[0170] When the number of iterations of the temporary solution equals the maximum number of iterations, the candidate output solution is used as the order allocation method.

[0171] The initial solution is a randomly generated order allocation method and its corresponding decision variables under the condition of meeting the constraints. For example, aircraft 1 is responsible for orders 1 and 3, aircraft 2 is responsible for orders 2, 4 and 5, and the pickup time of order 1 (pos1 to pos2) is 08:15, and the delivery time is 08:15 + predicted delivery time = 08:28.

[0172] Understandably, adopting the Metropolis criterion can effectively improve the efficiency of multi-objective search strategies and global search capabilities.

[0173] Furthermore, the process of calculating the fitness score of the temporary solution includes:

[0174] Calculate operator selection weights based on the total score of the order operator iteration using the decay parameter and temporary solution;

[0175] The operator selection probability is calculated based on the ratio of the operator selection weight to the total operator selection weight;

[0176] The fitness score is calculated based on the operator selection probability, the total score of the order operator iteration, and the previous score of the order operator.

[0177] Specifically, the steps for solving the operations research decision-making model based on a multi-objective search strategy include:

[0178] Step S1: Randomly remove the removal orders and pickup / delivery times of a low-altitude aircraft from the initial solution, and assign these removed orders to the low-altitude aircraft with the highest idle rate to generate a temporary solution. For example: Suppose that in the initial solution, aircraft 1 is responsible for orders 1 and 3, and aircraft 2 is responsible for orders 2, 4, and 5. Remove order 1 from aircraft 1 and assign it to aircraft 2, which has a higher idle rate, to generate a new temporary solution. Temporary solution: aircraft 1 is responsible for order 3, and aircraft 2 is responsible for orders 1, 2, 4, and 5.

[0179] Step S2: Calculate the fitness score. The fitness score is used to evaluate the quality of the temporary solution, taking into account factors such as mission completion efficiency and aircraft load balancing.

[0180] Specifically, step S2 includes:

[0181] Step S21, the process of calculating the operator selection weights is as follows:

[0182] ;

[0183] In the formula, This is the decay parameter, preferably 0.5, used to balance the weights of historical performance and current performance. It is the number of times the order operator i is used. This represents the selection weight of the order operator i at the current iteration number t. This represents the selection weight of order operator i in the previous iteration number t-1. This represents the total score of the order operator i after the t-th iteration. Therefore, through the decay parameter... Operator selection weights can be used in historical performance and current performance By finding a balance between these factors, the weights of the operator selection can be updated during the iteration process.

[0184] Step S22, the process of calculating the operator selection probability is as follows:

[0185] ;

[0186] In the formula, This represents the probability of choosing operator i at the current iteration number t. This represents the selection weight of the order operator i at the current iteration number t. Let r represent the total operator selection weight of all order operators i in the current iteration number t (the iteration number is denoted as r). Therefore, it is possible to determine the selection probability of order operator i based on its proportion in the overall set of selection weights.

[0187] Step S23, the process of calculating the fitness score is as follows:

[0188] ;

[0189] In the formula, This represents the fitness score in the t-th iteration. This represents the probability of choosing operator i at the current iteration number t. This represents the score of order operator i in the previous iteration, which is t-1. This represents the total score of order operator i after the t-th iteration (the number of iterations is denoted as r). Therefore, this adaptive score is combined with the historical score. and current score This allows the algorithm to make more stable decisions in dynamic environments, avoiding over-reliance on current data or ignoring historical experience. The fitness score measures the quality of a solution (order allocation method). A higher score indicates a better performance of the solution (order allocation method).

[0190] Here, the order operator i is the basic unit used to describe task allocation during the optimization process. Its form can include order allocation rules and time constraints. For example, the order allocation rules are: aircraft 1 is responsible for orders 1 and 3, and aircraft 2 is responsible for orders 2, 4 and 5. The time constraints are: the pickup time for order 1 is 08:15, and the delivery time is 08:28.

[0191] The selection weight of an operator determines the probability that the operator will be selected when generating a new solution, and is a value between 0 and 1. The fitness score is used to evaluate the quality of a solution (order allocation method); the higher the score, the better the solution, and is a value between 0 and 1.

[0192] Step S3: Evaluation and Iteration of Temporary Solution. Specifically, step S3 includes:

[0193] Step S31: Determining the fitness score: If the fitness score is greater than or equal to the score threshold (preferably 0.9), the temporary solution is determined to meet the optimization requirements. The temporary solution is then used as a candidate output solution, and steps S1 to S2 are iteratively calculated on the temporary solution. Otherwise, the temporary solution is iterated using the Metropolis criterion to generate candidate output solutions, and then steps S1 to S2 are iteratively calculated on the temporary solution. The Metropolis criterion is a probabilistic acceptance criterion. It calculates the difference in fitness scores between the temporary solution and the current optimal solution, and uses a probability function to determine whether to accept the worse solution. This is used to accept the worse solution during the optimization process to avoid getting trapped in local optima. The Metropolis criterion is a standard setting for multi-objective search strategies and will not be elaborated further here.

[0194] Step S4: When the number of iterations of the temporary solution equals the maximum number of iterations (preferably 50), stop the iteration and use the candidate output solution as the final order allocation method.

[0195] It is understandable that, given the forms of the initial and provisional solutions, the forms of the candidate output solutions and the current optimal solution in the iterative evaluation process (steps S1 to S3) are the same and can be derived.

[0196] Therefore, the order operator i that performs better and meets the requirements will have a higher score and a higher weight, thus increasing the number of times it is selected and making the order allocation process more accurate.

[0197] like Figure 2 As shown, this embodiment also provides a system for scheduling low-altitude aircraft swarm missions using a big data-based method, comprising:

[0198] The acquisition module is used to determine whether to execute the task scheduling strategy of the low-altitude aircraft cluster based on the comparison results of the current accumulated order volume and the dynamic initial allocation threshold, as well as the comparison results of the historical delivery time and the dynamic time threshold.

[0199] The delivery time prediction module, connected to the acquisition module, is used to generate a predicted delivery time by using the historical order arrival rate, historical delivery time, time period characteristics, delivery route environmental characteristics, current order accumulation, and current low-altitude aircraft order allocation status of the low-altitude aircraft delivery big data platform through a time series feature hybrid model if the task scheduling strategy is executed. The time series feature hybrid model is built based on LSTM and Transformer network architecture.

[0200] The operations research decision model construction module is connected to the delivery time prediction module and is used to construct an operations research decision model with the optimization objectives of minimizing the average of multiple predicted delivery times, minimizing the average flight distance per order, and minimizing the timeout rate, the order allocation method as the decision variable, and the completion time constraint based on the predicted delivery time.

[0201] The solution module is connected to the operations research decision model construction module and the acquisition module. It is used to solve the operations research decision model based on a multi-objective search strategy, determine the order allocation method of the low-altitude aircraft cluster, and determine the dynamic initial allocation threshold and the dynamic duration threshold for the next execution based on the predicted delivery time.

[0202] The above scheme further improves the feasibility, solution efficiency, and scenario adaptability of low-altitude aircraft scheduling by using refined constraint modeling, adaptive search algorithm optimization, and dynamic threshold collaborative feedback.

[0203] In this embodiment, a hybrid model of temporal features based on LSTM and Transformer network architecture is used to apply the delivery time predicted by the machine learning model to low-altitude aircraft order allocation decisions and dynamic delay scheduling strategies. This achieves a balance between immediate response and global efficiency in low-altitude aircraft cluster scheduling, improves the adaptability of dynamic delay scheduling strategies, enhances delivery efficiency, and reduces timeout rates. Through the multi-head attention mechanism of Transformer cross-feature units, implicit correlations between global delivery route environment features, time period features, and low-altitude aircraft order allocation status are mined to identify their impact on delivery time. LSTM is used to extract local historical order arrival rates and delivery time trend and periodic features. A bias prediction mechanism fuses global and local features, for example, focusing more on local temporal changes during morning and evening peak periods, achieving better business adaptability in delivery time prediction. Refined constraint modeling, adaptive search algorithm optimization, and dynamic threshold collaborative feedback further improve the feasibility, solution efficiency, and scenario adaptability of low-altitude aircraft scheduling.

[0204] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0205] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for scheduling low-altitude aircraft swarm missions based on big data, characterized in that, include: Based on the comparison between the current order accumulation and the dynamic initial allocation threshold, as well as the comparison between the historical delivery time and the dynamic time threshold, determine whether to execute the low-altitude aircraft cluster task scheduling strategy. If the task scheduling strategy is executed, the historical data of the low-altitude aircraft delivery big data platform will be used to generate a predicted delivery time through a time-series feature fusion model. The time-series feature fusion model is built on a network architecture based on Transformer and LSTM. The Transformer is used to extract global features, and the LSTM is used to extract time-series trend features. An operations research decision model is constructed with the optimization objectives of minimizing the average of multiple predicted delivery times, minimizing the average flight distance per order, and minimizing the timeout rate, and with the order allocation method as the decision variable, and with completion time constraints constructed based on the predicted delivery times. The operation research decision model is solved based on a multi-objective search strategy to determine the order allocation method of the low-altitude aircraft cluster, and the dynamic initial allocation threshold and the dynamic duration threshold for the next execution are determined based on the predicted delivery time.

2. The low-altitude aircraft swarm task scheduling method based on big data according to claim 1, characterized in that, The process of generating predicted delivery times includes: The historical data is preprocessed and feature-stitched to generate fused features. Historical order arrival rates and historical delivery times are standardized and feature-stitched to generate time-series features. The fused features are processed through a Transformer cross-feature unit to perform multi-head attention calculations, generating a global feature for order duration. The time-series features are used to generate time-series trend features through an LSTM time-series feature encoding unit; The global features of order duration and the time-series trend features are used to generate predicted delivery duration through a feature fusion mapping unit.

3. The low-altitude aircraft swarm task scheduling method based on big data according to claim 2, characterized in that, The process of generating global features for order duration includes: The fused features are processed by a multi-head attention computation subunit to perform multi-head attention computation and multi-head concatenation, thereby generating multi-head features. The multi-head features are residually connected and normalized with the fused features through a single residual connection subunit to generate residual features; The residual features are used to calculate the global correlation of features through feedforward network subunits to generate globally correlated features; The global correlation feature is connected to the residual feature through a secondary residual connection subunit and normalized to generate a global feature of order duration.

4. The low-altitude aircraft swarm task scheduling method based on big data according to claim 2, characterized in that, The process of generating predicted delivery times includes: The delivery time trend strength, periodic fluctuation coefficient, sudden fluctuation coefficient, and time-series delivery time feature of the time-series trend feature are weighted and biased through a multi-layer fully connected layer with ReLU activation function set by the feature fusion mapping unit to generate biased predicted delivery time. The global feature of order duration and the biased predicted delivery duration are used to generate a spliced ​​feature through the feature splicing layer of the feature fusion mapping unit; The spliced ​​features are fused and mapped through a fully connected network with a ReLU activation function set in the feature fusion mapping unit to generate a predicted delivery time.

5. The low-altitude aircraft swarm task scheduling method based on big data according to claim 1, characterized in that, The construction process of the time-series feature fusion model includes: A prediction error term is constructed based on the mean squared error of the training predicted values ​​and the actual values ​​of delivery time; A waiting cost term is constructed based on the training prediction value of delivery time and the optimal waiting time. A comprehensive loss function is constructed based on the weighted sum of the prediction error term and the waiting cost term, and the comprehensive loss function is used for the optimization training of the time-series feature hybrid model.

6. The low-altitude aircraft swarm task scheduling method based on big data according to claim 1, characterized in that, The process of constructing completion time constraints based on predicted delivery time includes: The completion time constraint is constructed based on the pickup and delivery times of two orders executed sequentially by the low-altitude aircraft in the order allocation method and the predicted delivery time of the earlier order.

7. The low-altitude aircraft swarm task scheduling method based on big data according to claim 6, characterized in that, The process of solving the operations research decision model based on a multi-objective search strategy to determine the order allocation method for the low-altitude aircraft cluster includes: After randomly removing the removal order and delivery time of the first low-altitude aircraft from the initial solution, the removal order is assigned to the second low-altitude aircraft with the highest idle rate to generate a temporary solution; Calculate the fitness score of the temporary solution; If the fitness score is greater than or equal to the score threshold, then the temporary solution is used as a candidate output solution, and the temporary solution is iteratively calculated. If the fitness score is less than the score threshold, the temporary solution is iteratively calculated to generate candidate output solutions using the Metropolis criterion, and then the temporary solution is iteratively calculated. When the number of iterations of the temporary solution equals the maximum number of iterations, the candidate output solution is used as the order allocation method.

8. The low-altitude aircraft swarm task scheduling method based on big data according to claim 7, characterized in that, The process of calculating the fitness score of the temporary solution includes: Calculate operator selection weights based on the total score of the order operator iteration using the decay parameter and temporary solution; The operator selection probability is calculated based on the ratio of the operator selection weight to the total operator selection weight; The fitness score is calculated based on the operator selection probability, the total score of the order operator iteration, and the previous score of the order operator.

9. The low-altitude aircraft swarm mission scheduling method based on big data according to any one of claims 1 to 8, characterized in that, The process of determining the dynamic initial allocation threshold and the dynamic duration threshold for the next execution based on the predicted delivery time includes: The product of the predicted delivery time and the current order density is used as the dynamic initial allocation threshold; Calculate the ratio of total delivery time to total order arrival rate, and calculate the dynamic time threshold based on the minimum of the predicted delivery time and the ratio; The process of determining whether to execute the task scheduling strategy of the low-altitude aircraft cluster includes: if the current accumulated order amount is less than the dynamic initial allocation threshold, or the historical delivery time is less than the dynamic duration threshold, then the task scheduling strategy of the low-altitude aircraft cluster is executed.

10. A system applying the big data-based low-altitude aircraft swarm mission scheduling method as described in any one of claims 1 to 9, characterized in that, include: The acquisition module is used to determine whether to execute the task scheduling strategy of the low-altitude aircraft cluster based on the comparison results of the current accumulated order volume and the dynamic initial allocation threshold, as well as the comparison results of the historical delivery time and the dynamic time threshold. The delivery time prediction module, connected to the acquisition module, is used to generate a predicted delivery time from the historical data of the low-altitude aircraft delivery big data platform through a time-series feature fusion model if the task scheduling strategy is executed. The time-series feature fusion model is built on a network architecture based on Transformer and LSTM fusion, where Transformer is used to extract global features and LSTM is used to extract time-series trend features. The operations research decision model construction module is connected to the delivery time prediction module and is used to construct an operations research decision model with the optimization objectives of minimizing the average of multiple predicted delivery times, minimizing the average flight distance per order, and minimizing the timeout rate, the order allocation method as the decision variable, and the completion time constraint based on the predicted delivery time. The solution module is connected to the operations research decision model construction module and the acquisition module. It is used to solve the operations research decision model based on a multi-objective search strategy, determine the order allocation method of the low-altitude aircraft cluster, and determine the dynamic initial allocation threshold and the dynamic duration threshold for the next execution based on the predicted delivery time.

Images