A logistics order allocation optimization method, device and medium based on deep learning

Abstract

The application discloses a logistics order distribution optimization method and device based on deep learning and a medium, relates to the technical field of deep learning, and comprises the following steps: a deep learning neural network is used to decouple and generate dynamic characteristics representing scheduling rules, so as to construct a spatiotemporal manifold space in which orders and distribution are cooperatively evolved. In the space, an order is defined as a gravitational source, a delivery man is defined as a moving particle, a candidate distribution matrix is generated by calculating the motion force of the particle through a gravitational field model, and screening is performed in combination with logistics business constraint rules. On this basis, a virtual global particle is introduced to iteratively optimize the order access sequence, a globally optimal distribution path is obtained, the path structure is divided into multiple delivery man execution tasks according to the path structure, and finally, an order distribution result is obtained. The application can comprehensively consider order demand changes, distribution capacity states and complex road network environments, realize overall optimization and distribution of logistics orders, improve distribution efficiency and improve the scheduling capability of a logistics system.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to a method, device and medium for optimizing logistics order allocation based on deep learning. Background Technology

[0002] With the rapid development of e-commerce, on-demand delivery, and large-item logistics, logistics and distribution systems face challenges such as a significant increase in order volume, wide delivery areas, and complex road network environments. Traditional logistics order allocation methods typically rely on rule-based or simple heuristic algorithms for dispatching orders, such as allocating orders based on the shortest distance, the nearest delivery person, or fixed area divisions. However, these methods often fail to simultaneously consider multiple factors such as changes in order demand, differences in delivery personnel service capabilities, and road network conditions, resulting in low delivery efficiency, unreasonable delivery routes, and insufficient utilization of delivery resources.

[0003] Existing methods typically only use deep learning to predict order demand or estimate delivery time, while still relying on traditional optimization algorithms in the actual order allocation process. This makes it difficult to fully characterize the complex coupling relationship between order demand, delivery personnel behavior, and road network structure. Furthermore, in complex road network environments, delivery routes often have multiple reachable paths. Traditional algorithms struggle to comprehensively consider the traffic impedance, real-time traffic conditions, and delivery time requirements of different paths, resulting in a lack of overall optimization capabilities in order allocation. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a deep learning-based logistics order allocation optimization method to solve the problem that existing logistics order allocation methods are difficult to optimize globally by comprehensively considering order demand, delivery capacity, and complex road network factors.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a logistics order allocation optimization method based on deep learning, which includes performing cross-domain feature extraction on logistics data from heterogeneous regions, using a deep learning neural network to decouple and generate dynamic features representing scheduling rules, and constructing a spatiotemporal manifold space for the co-evolution of orders and delivery.

[0008] After the deep learning neural network, a feature definition unit is connected. The feature definition unit includes taking orders as the gravitational source, establishing a gravitational field in the spatiotemporal manifold space based on real-time order information, defining the deliveryman as a moving particle driven by the dynamic features, calculating the gravitational force of the particle in the gravitational field to generate a candidate allocation matrix, and using logistics order business constraint rules to enforce constraints on the candidate allocation matrix.

[0009] Following the feature definition unit, an iterative optimization unit is introduced. The iterative optimization unit includes, under constraints, driving particles to generate delivery tasks in the spatiotemporal manifold space through the candidate allocation matrix, and outputting the final order allocation result through iterative optimization.

[0010] As a preferred embodiment of the deep learning-based logistics order allocation optimization method described in this invention, the logistics data of the heterogeneous region refers to historical logistics information in different delivery regions, including order information, delivery information, and road network information;

[0011] When performing the cross-domain feature extraction, the preprocessed logistics data in different delivery areas are uniformly encoded so that order information, delivery information and road network information are represented in a unified data structure.

[0012] As a preferred embodiment of the deep learning-based logistics order allocation optimization method of the present invention, wherein: the decoupling generation of dynamic features representing scheduling rules using a deep learning neural network includes:

[0013] Deep learning neural networks are used to extract features from order information, delivery information, and road network information respectively;

[0014] By using a cross-attention mechanism to perform joint feature fusion on the order information, delivery person information, and road network information, the coupling relationship between the three is decoupled and reconstructed into a unified dynamic tensor;

[0015] The dynamic tensor is mapped to the metric operator of the spatiotemporal manifold space, and the scalar curvature of the road network at various points in the manifold space is adjusted through the metric operator, so that the road network traffic impedance and delivery personnel service habits can be mapped to the geometric deformation in the manifold space, thereby constructing a manifold structure that reflects the coupling relationship between order demand and delivery capacity.

[0016] Based on the geometric deformation of the manifold space, in the spatiotemporal manifold space, the adjusted scalar curvature at each point of the road network is divided by the unadjusted scalar curvature to generate a non-uniformly distributed gravitational conduction coefficient. This allows the intensity of the order's effect on the delivery particle to be adjusted by the spatial geometric characteristics, thereby forming a dynamic feature that can guide the delivery particle to avoid high-resistance regions and converge towards high-value order nodes.

[0017] As a preferred solution of the logistics order allocation optimization method based on deep learning according to the present invention, wherein: the gravitational field includes:

[0018] In the space-time manifold space, each order to be allocated is mapped into a gravitational source unit with time-varying mass, and according to the execution position corresponding to the order, a potential energy well is formed at the local coordinates of the manifold space;

[0019] The gravity generated by the gravity source unit of each order is expressed as: ;

[0020] where t represents the numerical value of the order waiting time; m is the order mass; represents the distance of the deliveryman arriving at the gravity source unit on the i-th path; represents the i-th path; represents the path average gravitational conduction coefficient;

[0021] The gravitational field is generated by representing the gravity of each gravity source unit.

[0022] As a preferred solution of the logistics order allocation optimization method based on deep learning according to the present invention, wherein: the moving gravitational force includes, for any deliveryman particle, calculating the gravity of the particle under the action of each gravity source unit in each reachable path state respectively, laying all the gravities on the corresponding paths, and vectorially adding all the laid gravities at each position of the road network, so as to form the direction and magnitude of the guiding tendency of the path to the particle at any position of the road network.

[0023] A candidate allocation matrix is constructed with the orders to be allocated as the horizontal axis and the deliverymen as the vertical axis; each intersection point in the matrix represents the matching tendency between the order and the deliveryman; by calculating the vector addition result of all the gravities of each reachable path between the gravity source unit represented by the order and the particle represented by the deliveryman and dividing it by the path length, the matching tendency under each reachable path is obtained, so that each element in the candidate allocation matrix is in a quantum state.

[0024] As a preferred solution of the logistics order allocation optimization method based on deep learning according to the present invention, wherein: the logistics order business constraint rules include the time period traffic restrictions, road traffic restrictions, vehicle load restrictions and delivery time limit restrictions in the road network.

[0025] As a preferred solution of the logistics order allocation optimization method based on deep learning according to the present invention, wherein: driving the particles to generate delivery tasks in the space-time manifold space includes simplifying all the particles into a virtual global particle and using the virtual global particle to complete the logistics orders in the road network.

[0026] Furthermore, by using the shortest path in the road network after scalar curvature adjustment as the fitness function, the order sequence is iteratively optimized to obtain the optimal single-particle delivery solution.

[0027] The single-particle delivery scheme is divided equally among each delivery person according to the path length adjusted by scalar curvature, resulting in a continuous path equal to the number of delivery persons K.

[0028] The obtained K consecutive paths are randomly assigned to delivery personnel to obtain the final order allocation result.

[0029] As a preferred embodiment of the deep learning-based logistics order allocation optimization method of the present invention, the iterative optimization includes: updating the candidate allocation matrix in real time according to the position of the virtual global particle; eliminating the corresponding gravitational source unit after each order is completed; and normalizing each element in the matrix to obtain the selection probability of the next position of the virtual global particle, thereby realizing the selection of the next order position.

[0030] After completing the iterative optimization, repeat the iterative optimization a preset number of L times to obtain L optimization results;

[0031] The L optimization results are returned to the neural network, which uses each optimization result to calculate the reconstruction loss and obtains the optimization result with the minimum loss, which is taken as the optimal single-particle delivery scheme.

[0032] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the deep learning-based logistics order allocation optimization method described in the first aspect of the present invention.

[0033] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the deep learning-based logistics order allocation optimization method described in the first aspect of the present invention.

[0034] The beneficial effects of this invention are as follows: By utilizing deep learning to uniformly model heterogeneous regional logistics data, this invention can simultaneously characterize the correlation between changes in order demand, delivery personnel service capabilities, and road network traffic conditions, thereby improving the accuracy of order allocation decisions. Based on the construction of a spatiotemporal manifold space, orders and delivery personnel are abstracted as gravitational sources and moving particles, respectively, allowing the impact of order demand on delivery behavior to be dynamically expressed within the road network structure, thus improving the rationality of delivery route selection. Simultaneously, by using virtual global particles to optimize the overall order access sequence, and then assigning the resulting delivery routes to multiple delivery personnel, path duplication and detours can be reduced while ensuring delivery timeliness, improving the utilization rate of delivery resources. Compared to traditional rule-based or locally optimized order allocation methods, this invention can obtain more reasonable order allocation results in complex road networks and multi-order scenarios, thereby improving overall logistics and delivery efficiency. Attached Figure Description

[0035] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a flowchart of a deep learning-based optimization method for logistics order allocation.

[0037] Figure 2 This is a framework diagram of deep learning in a deep learning-based logistics order allocation optimization method. Detailed Implementation

[0038] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0039] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0040] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0041] Reference Figure 1 and Figure 2 This is one embodiment of the present invention, which provides a deep learning-based logistics order allocation optimization method, including the following steps:

[0042] S1: By performing cross-domain feature extraction on logistics data from heterogeneous regions, and using deep learning neural networks to decouple and generate dynamic features that represent scheduling patterns, a spatiotemporal manifold space for the co-evolution of orders and delivery is constructed.

[0043] The logistics data in the heterogeneous regions refers to historical logistics information in different delivery areas, including but not limited to order data, delivery person status data, road network structure data, traffic status data, and historical dispatch data.

[0044] The order data is used to characterize the basic attribute information of orders to be delivered, including order pickup location, order delivery location, order generation time, order priority, order weight or volume, and promised delivery time. The delivery person status data is used to characterize the real-time status of delivery capacity, including the delivery person's current location, remaining capacity, current number of tasks, historical fulfillment records, and service area preferences. The road network structure data describes the connectivity of roads within the delivery area, including road node information, road connection topology, and the length of each road segment. The traffic status data describes the real-time traffic conditions, including road speed, congestion level, and temporary traffic control information. The historical scheduling data characterizes the changes in order demand and capacity supply in the logistics system over historical time periods, including historical order generation density, delivery trajectory records, and historical order completion time distribution.

[0045] By uniformly encoding and fusing the logistics data from the aforementioned heterogeneous regions, cross-domain feature extraction can be performed on data from different regions, at different time scales, and of different types under the same data structure, thereby providing basic input data for subsequent decoupling of dynamic features and construction of spatiotemporal manifold space.

[0046] When performing the cross-domain feature extraction, the preprocessed logistics data in different delivery areas are uniformly encoded so that order information, delivery information and road network information are represented in a unified data structure.

[0047] Furthermore, deep learning neural networks are used to extract features from order information, delivery information, and road network information respectively; and a cross-attention mechanism is used to perform joint feature fusion on the order information, delivery information, and road network information, decoupling the coupling relationship between the three and reconstructing it into a unified dynamic tensor. The dynamic tensor is essentially a set of multi-dimensional feature matrices used to describe the dynamic relationship between order nodes, delivery personnel nodes, and road network nodes, and its values ​​are generated by the deep learning model based on historical data and real-time status.

[0048] In this embodiment, the neural network can employ a deep learning-based multimodal feature extraction network, such as a fusion neural network structure composed of an order information encoding network, a delivery information encoding network, and a road network structure encoding network. Specifically, the order information encoding network extracts features related to changes in order demand, the delivery information encoding network extracts features related to delivery personnel location, service capacity, and historical behavior, and the road network structure encoding network extracts features related to road topology and traffic conditions. These three types of features are then fused using a cross-attention mechanism to generate a dynamic tensor representing the dynamic relationship between order demand, delivery capacity, and the road network environment.

[0049] In other alternative embodiments, the neural network may also be implemented using other deep learning models, such as convolutional neural networks (CNN), recurrent neural networks (RNN), long short-term memory networks (LSTM), graph neural networks (GNN), or Transformer structures.

[0050] In the process of logistics order allocation, there are complex coupling relationships between order demand, delivery personnel capacity, and road network structure. For example, the number of orders and their spatial distribution affect the delivery personnel's movement paths, while the delivery personnel's real-time location and service capabilities influence the order allocation strategy. Road network structure and traffic conditions directly constrain delivery efficiency. Traditional methods typically process these factors separately, making it difficult to accurately characterize the dynamic relationships among them. Therefore, this paper utilizes deep learning neural networks to extract features from order information, delivery information, and road network information separately, and uses a cross-attention mechanism to jointly fuse features from different sources. This establishes the relationship between order demand, delivery capacity, and road network structure in a unified feature space, providing a foundational data representation for subsequent construction of spatiotemporal manifold space and gravitational field models.

[0051] Among them, order information is used to characterize order demand and time-related pressures, delivery information is used to characterize individual delivery personnel's behavioral characteristics and service capacity status, and road network information is used to characterize road topology and real-time traffic conditions.

[0052] The dynamic tensor is mapped to a metric operator in the spatiotemporal manifold space, and the scalar curvature of the road network at various points in the manifold space is adjusted through the metric operator; so that the road network traffic impedance and delivery personnel service habits can be mapped to geometric deformation in the manifold space, thereby constructing a manifold structure that reflects the coupling relationship between order demand and delivery capacity.

[0053] Based on the geometric deformation of the manifold space, in the spatiotemporal manifold space, the adjusted scalar curvature at each point of the road network is divided by the unadjusted scalar curvature to generate a non-uniformly distributed gravitational conduction coefficient. This allows the intensity of the effect of orders (gravitational sources) on delivery personnel particles to be adjusted by the spatial geometric characteristics, thereby forming a dynamic feature that can guide delivery personnel particles to avoid high-resistance regions and converge towards high-value order nodes.

[0054] The key point is that by mapping the dynamic tensor to a metric operator in a spatiotemporal manifold, the multidimensional characteristics of the logistics system can be transformed into geometric properties in the manifold, thus expressing the dynamic relationship between order demand, delivery capacity, and the road network environment in a unified space. By adjusting the scalar curvature in the manifold, the difficulty of traversing different road segments and the behavioral characteristics of delivery personnel can be represented as geometric deformations. This allows delivery route planning to consider not only spatial distance but also the complexity of the actual delivery environment. Based on this geometric deformation, a gravitational transmission coefficient is further generated, enabling the propagation intensity of order gravity in the road network to be adaptively adjusted according to road conditions. This guides delivery personnel particles to tend towards high-value order areas and avoid high-impedance road segments during the delivery process, improving the rationality of order allocation and route selection.

[0055] Scalar curvature is a scalar value describing the degree of local geometric curvature in a manifold space. In this embodiment, scalar curvature is used to characterize the traffic characteristics at different locations in the road network. When the traffic conditions or delivery efficiency of a certain road segment are good, the scalar curvature change in that area is small; when a certain road segment experiences traffic congestion, road restrictions, or low delivery efficiency, the scalar curvature is adjusted to cause a larger geometric deformation in the manifold space. In this way, the traffic impedance in the real road network is mapped to the curvature change in the manifold space, allowing the delivery path to automatically bypass high-impedance areas geometrically, thus reflecting the influence of the actual road network environment during the route planning process.

[0056] The gravitational conduction coefficient is used to describe the propagation ability of order gravity in the road network, and its value is calculated by the ratio between the adjusted scalar curvature and the unadjusted scalar curvature. This coefficient reflects the enhancement or attenuation effect of different road segments on the propagation of order gravity. When the traffic conditions of a certain road segment are good, the gravitational conduction coefficient is large, indicating that the attracting effect of order demand on delivery staff particles can be easily propagated through this path; when the traffic conditions of a certain road segment are poor or there are restrictions, the gravitational conduction coefficient is small, indicating that the propagation of order gravity in this area is inhibited. By introducing the gravitational conduction coefficient, the attracting effect of order demand on delivery staff can show a non-uniform distribution in the road network, so that the movement trend of delivery staff particles more conforms to the actual delivery environment, and finally a delivery decision-making mechanism that can reflect the comprehensive influence of order demand, delivery capacity and road network structure is formed.

[0057] S2: After the neural network of the degree learning, a feature definition unit is connected: taking the order as the gravity source, according to the real-time order information, a gravitational field is established in the space-time manifold space; at the same time, the delivery staff is defined as a moving particle driven by the dynamic characteristics, and the gravitational force of the particle in the gravitational field is calculated to generate a candidate allocation matrix.

[0058] Specifically, in the space-time manifold space, each to-be-allocated order is mapped to a gravitational source unit with time-varying mass, and according to the execution position corresponding to the order, a potential energy well is formed at the local coordinates of the manifold space. The distribution characteristics of order demand in space are transformed into a computable potential energy structure, so that the attracting effect of order demand on delivery staff particles can be dynamically expressed in the manifold space. Enable the delivery staff particles to be guided by the order gravity during the movement process, and thus gradually gather towards the location of the order.

[0059] Referring to the formula form of universal gravitation and combining with the diversity of reachable paths in the road network, the gravity generated by the gravity source unit of each order is expressed as: .

[0060] Replace the constant term in the original universal gravitation formula with: in the road network, representing the path average gravitational conduction coefficient of , so that the order gravity can reflect the traffic conditions and gravity propagation ability of different paths. When the traffic conditions of a certain path are good or the traffic resistance is small, its corresponding gravitational conduction coefficient is large, thus enhancing the attracting effect of the order on the delivery staff particles; when there is congestion or traffic restriction on the path, the gravity propagation on this path is inhibited by reducing the gravitational conduction coefficient. And according to the number n of reachable paths for the delivery staff to reach the gravity source unit in the road network, make present as a quantum state representing n reachable paths, so that F presents as n states, expressed as: .

[0061] Where t represents the order waiting time, which negatively reflects the quality of fulfilling logistics responsibilities; m is the order quality, representing the order importance calculated by a preset model; This represents the distance the delivery person travels to the gravitational source unit along the i-th path; Let i represent the i-th path, where i ∈ (0, n). The gravitational field is generated by representing the gravitational force of each gravitational source unit.

[0062] It's important to note that in the logistics order allocation process, there are usually multiple accessible paths for delivery personnel to reach a given order location. These paths differ in terms of travel distance, traffic conditions, and road restrictions. Traditional order allocation methods based on single-path distance or simple time estimations often fail to comprehensively reflect the actual delivery costs in a multi-path environment, easily leading to unreasonable path selection or reduced order allocation efficiency. Therefore, this embodiment, referencing the expression of the universal gravitation formula, abstracts order demand as a gravitational source and, combined with the diversity of accessible paths in the road network, performs path-based modeling of the order's gravitational force. This allows the attractive effect of the order on the delivery personnel particle to be dynamically expressed under different path conditions.

[0063] Considering that multiple orders along the same route are a crucial factor in delivery, releasing multiple states of the gravitational source simultaneously makes it easier to identify the "route-related" element, thus serving as a guiding tendency for particles within the network. For any delivery particle, the gravitational force exerted on it by each gravitational source unit under each reachable path state is calculated. All gravitational forces are then laid out along the corresponding paths. At each location in the network, all laid-out gravitational forces are vector-summed, forming the direction and magnitude of the path's guiding tendency for the particle at any position in the network. In other words, this involves expanding the quantum state, with each state acting simultaneously to exert a force on the particle, thereby generating a guiding tendency.

[0064] By uniformly calculating the effects of various gravitational sources under different path conditions and performing vector superposition in the road network space, a guidance direction and intensity distribution reflecting the order distribution characteristics can be formed in the overall space. In this way, the movement trend of delivery personnel particles can be adjusted based on comprehensive guidance information in space, making it easier to form a reasonable delivery sequence during path selection. Simultaneously, this mechanism also enables the continuous propagation of the impact of order demand on delivery behavior in space, allowing subsequent order matching relationships to simultaneously consider order location distribution, path accessibility, and overall delivery efficiency, thereby improving the overall rationality of order allocation and path planning.

[0065] A candidate allocation matrix is ​​constructed using the orders to be allocated as the horizontal axis and delivery personnel as the vertical axis. Each intersection in the matrix represents the matching tendency between orders and delivery personnel. The matching tendency for each reachable path is obtained by calculating the vector sum of all gravitational forces between the gravitational source unit represented by the order and the particle represented by the delivery personnel, divided by the path length. This ensures that each element in the candidate allocation matrix is ​​a quantum state, with the dimension of the quantum state matching the number of reachable paths. In reality, this matrix has only one vertical axis, but it is updated in real time.

[0066] Furthermore, it's important to understand that after obtaining the gravitational forces under each path state, these forces need to be propagated along the corresponding paths within the road network, spreading from the order location to the entire network space. In other words, each order's gravitational source creates an "influence region" within its reachable path. In this way, every node in the road network can receive gravitational influences from multiple orders, thus forming a spatial distribution structure. This process essentially constructs a potential field-like spatial distribution model within the road network, enabling order demand to continuously propagate throughout space, rather than only having a localized impact at the order location.

[0067] S3: Apply mandatory constraints to the candidate allocation matrix using logistics order business constraint rules.

[0068] After generating the candidate allocation matrix, it is necessary to screen the candidate allocation relationships for legality in accordance with the business constraint rules in the actual logistics and distribution environment, so as to avoid generating order allocation results that do not meet the transportation conditions or delivery requirements.

[0069] Specifically, in this embodiment, the logistics order business constraint rules include time-of-day traffic restrictions, road traffic restrictions, vehicle load limits, and delivery time limits in the road network.

[0070] First, based on time-limited traffic restrictions in the road network, it is determined whether delivery vehicles are allowed to enter a certain road segment during a specific time period. If a road segment has traffic restrictions during delivery time, the candidate allocation relationship on the corresponding path is eliminated or its matching weight is reduced. Second, based on road traffic restriction rules, it is determined whether the vehicle type, length, height, or weight of the delivery vehicle meets the road traffic requirements. If the delivery vehicle does not meet the road traffic requirements, the candidate allocation relationship on the corresponding path is eliminated.

[0071] Furthermore, based on vehicle load limits, the remaining load capacity of the delivery driver's vehicle is matched with the weight of the order goods. When the order weight exceeds the vehicle's remaining load capacity, the allocation relationship between the order and the corresponding delivery driver is prohibited.

[0072] Finally, based on delivery time constraints, the time required for delivery personnel to complete order delivery under current road network conditions is estimated. If the estimated delivery time exceeds the order's stipulated fulfillment time limit, the allocation relationship is considered invalid.

[0073] By filtering and correcting the candidate allocation matrix item by item through the above constraint rules, an effective order allocation relationship that meets the actual logistics and transportation conditions can be obtained.

[0074] S4: After the feature definition unit, an iterative optimization unit is introduced: Under constraints, the candidate allocation matrix drives the particles to generate delivery tasks in the spatiotemporal manifold space, and the final order allocation result is output through iterative optimization.

[0075] Furthermore, all particles are simplified into a single virtual global particle, which is then used to complete logistics orders within the road network.

[0076] Furthermore, by using the shortest path in the road network after scalar curvature adjustment as the fitness function, the order sequence is iteratively optimized to obtain the optimal single-particle delivery scheme.

[0077] The single-particle delivery scheme is divided equally among each delivery person according to the path length adjusted by scalar curvature, resulting in a continuous path equal to the number of delivery persons K; the uneven parts are randomly merged into adjacent paths.

[0078] The obtained K consecutive paths are randomly assigned to delivery personnel to obtain the final order allocation result.

[0079] Specifically, in actual logistics and delivery processes, the number of orders usually far exceeds the number of delivery personnel. If route planning and order allocation are performed simultaneously for multiple delivery personnel, the combination space often increases dramatically, making the optimization process overly complex and difficult to achieve stable optimization results in a short time. Therefore, this invention first abstracts multiple delivery personnel particles, simplifying them into a single virtual global particle. This virtual particle then completes all order tasks sequentially within the road network according to the order access order. In this way, the original multi-delivery personnel collaborative scheduling problem can be transformed into a single-particle path optimization problem, significantly reducing the computational complexity of the optimization process and making the solution for the order access order more stable and efficient.

[0080] After optimizing the single-particle path, the overall delivery path is divided according to the path length of the road network adjusted by scalar curvature, ensuring that the number of path segments corresponds to the number of delivery personnel. Since the spatial distribution of different orders is uneven, path lengths may vary. Therefore, a small number of unevenly divided sections are allowed during path division, and these are adjusted by incorporating them into adjacent paths, thus ensuring a relatively balanced workload for each delivery person. Finally, the resulting multiple continuous paths are assigned to each delivery person for execution, transforming the original global delivery plan into actual delivery tasks for multiple delivery personnel. This design ensures both the overall path optimization effect and the feasibility of multi-delivery task allocation, thereby improving the overall efficiency of logistics order allocation and path planning.

[0081] It should be noted that the iterative optimization includes updating the candidate allocation matrix in real time according to the position of the virtual global particle, eliminating the corresponding gravitational source unit after each order is completed, and normalizing each element in the matrix as the selection probability of the next position of the virtual global particle to realize the selection of the next order position.

[0082] After completing the iterative optimization, repeat the iterative optimization a preset number of times L to obtain L optimization results.

[0083] The L optimization results are returned to the neural network, which uses each optimization result to calculate the reconstruction loss and obtains the optimization result with the minimum loss, which is taken as the optimal single-particle delivery scheme.

[0084] This step does not directly obtain the path result after one optimization. Instead, it generates multiple candidate single-particle delivery schemes through multiple randomized iterations. Then, the aforementioned neural network is used to determine the consistency of each candidate scheme, thereby selecting the optimal scheme that satisfies both the current road network conditions and historical good delivery patterns. Its core is not "finding an additional external scoring standard," but rather directly reusing the feature discrimination ability formed by the neural network during historical learning, using the neural network's own reconstruction loss as the evaluation criterion for the quality of the path scheme.

[0085] From an iterative perspective, the virtual global particle doesn't consistently choose the order with the highest current gravity at each step. Instead, it first constructs the selection probability of the next order position based on the normalized results of each element in the candidate allocation matrix, and then selects the next position according to this probability. Since the corresponding gravity source unit is eliminated after each order is completed, and the candidate allocation matrix is ​​updated synchronously, the selection probability of subsequent orders continuously changes. This makes the entire optimization process exhibit significant sequence dependence and path dependence. In other words, the choice in the previous step directly affects the gravity distribution and candidate relationship in the next step. Therefore, the same batch of orders may result in significantly different access orders under different initial choices or different random sampling conditions. Consequently, the result obtained in a single optimization is often only a locally optimal result formed under the current random sampling path and cannot guarantee that it is the globally optimal or the result that best conforms to the actual delivery pattern.

[0086] From the perspective of the selection process, the re-optimization process, repeated L times as preset, is essentially aimed at generating multiple different access order samples in the candidate solution space, thereby expanding the search range and reducing the impact of randomness caused by a single probability selection. Especially when there are a large number of orders, complex road network conditions, and strong sequential or competitive relationships between orders, different iterative trajectories may correspond to completely different delivery structures. If only one optimization is performed, it is easy to be affected by initial probability fluctuations and fall into local path patterns; however, by re-optimizing multiple times, multiple candidate single-particle delivery solutions with different structures can be obtained, providing a sufficient basis for comparison for subsequent optimization.

[0087] Furthermore, these L optimization results are fed back into the neural network, and the network is used to calculate the reconstruction loss of each result relative to the historical optimal delivery feature distribution. The neural network then uses its "empirical distribution standard" formed during historical learning to judge the rationality of the current optimization result. Because the neural network has already learned which delivery structures are more in line with efficient, stable, and low-cost delivery patterns from historical orders, delivery personnel, road networks, and scheduling results during the training phase, the smaller its reconstruction loss, the closer the current optimization result's distribution in the feature space is to the historically optimal solution, and therefore the more practically feasible and stable it is. In this way, solution selection no longer relies on a single, manually designed external indicator, but is directly judged by the feature structure already learned within the neural network. This ensures that the final selected single-particle delivery solution considers both the current road network and order status, while maintaining consistency with historically optimal scheduling patterns.

[0088] The reconstruction loss of each result relative to the historical optimal delivery feature distribution is calculated using a neural network; the optimal single-particle delivery scheme is selected as the result with the smallest reconstruction loss, i.e., the feature space distribution that best conforms to the historical pattern.

[0089] What needs to be said is that instead of arbitrarily choosing a standard score after optimization, we directly reuse the neural network's "values" (i.e., the loss it uses to self-correct when learning historical records) to select paths.

[0090] From an overall logical perspective, this step does not involve artificially designing an additional scoring function to judge the quality of paths after path optimization is completed. Instead, it directly reuses the error evaluation mechanism already formed during the training phase of the deep learning neural network. During the training process, the neural network continuously adjusts itself using historical delivery data and measures the difference between the current prediction and the actual delivery results using a loss function. As training progresses, the neural network gradually learns a set of implicit delivery patterns, forming a distribution structure of "high-quality delivery plans" in the feature space. Therefore, this loss function is not only an optimization objective during the training process but also implicitly reflects which delivery paths better conform to the scheduling patterns that performed well in historical data.

[0091] Based on this characteristic, this embodiment, after completing multiple path optimizations, does not directly evaluate based on path length, time cost, or manually set weight functions. Instead, it re-inputs each optimized single-particle delivery scheme into the neural network and evaluates the difference between the scheme and historical optimal delivery structures by calculating its reconstruction loss in the feature space. If a scheme has a small reconstruction loss, it indicates that the scheme's distribution in the feature space is closer to the efficient delivery pattern learned by the neural network from historical data. Conversely, if the loss is large, it indicates that although the scheme may perform well in the current path calculation, its overall delivery structure deviates significantly from historical experience, potentially leading to instability or uneven resource allocation in actual execution.

[0092] This method essentially uses the "scheduling experience" formed by the neural network during training on historical data as an implicit evaluation criterion, directly applying it to screen candidate paths. Compared to traditional methods, traditional path optimization typically requires constructing complex scoring functions, such as weighted combinations of factors like distance, time, and load balancing. This approach not only requires extensive human experience but also struggles to fully reflect the complex relationships in a real delivery environment. This embodiment, however, utilizes the reconstruction loss of the neural network as an evaluation criterion, enabling the path selection process to directly inherit the overall scheduling patterns learned by the neural network from historical data, thereby avoiding subjective biases introduced by manually designed evaluation functions.

[0093] Furthermore, this mechanism has a significant advantage: the evaluation criteria remain consistent with the feature extraction model. Since candidate path generation, feature representation, and loss calculation are all based on the same neural network structure, the path evaluation process and the feature modeling process reside in the same feature space, thus avoiding inconsistencies in metrics between different models. This means that path selection not only considers the path's length or time but also implicitly takes into account multi-dimensional factors such as order distribution structure, delivery personnel behavior patterns, and the road network environment, making the final selected delivery solution more consistent with the operational patterns of real logistics systems.

[0094] In this embodiment, the reconstruction loss is used to measure the difference between the distribution of the optimized single-particle delivery scheme in the feature space and the distribution of historical best delivery features. Specifically, during the training phase, the neural network learns the feature representations of order demand, deliveryman behavior, and road network status based on historical delivery records, forming a corresponding feature space structure. When evaluating each optimization result, the corresponding delivery route scheme is input into the neural network, its feature representation is reconstructed, and the degree of difference between the reconstructed result and the historical best delivery features is calculated. This difference can be calculated using mean squared error (MSE) or a similar distance metric to obtain the corresponding reconstruction loss value. When the reconstruction loss of a certain optimization result is small, it indicates that the distribution of that route scheme in the feature space is closer to the distribution characteristics of historical high-quality delivery schemes, and therefore it is regarded as the optimal single-particle delivery scheme.

[0095] In other feasible embodiments, the reconstruction loss can also be calculated using other forms of loss functions, such as mean absolute error (MAE), cross-entropy loss, or distance metrics based on feature distribution differences, such as Kullback-Leibler divergence or Wasserstein distance, to measure the degree of deviation between the candidate delivery scheme and historical delivery patterns in the feature space.

[0096] This embodiment also provides a computer device applicable to the logistics order allocation optimization method based on deep learning, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the logistics order allocation optimization method based on deep learning as proposed in the above embodiment.

[0097] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0098] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the deep learning-based logistics order allocation optimization method proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0099] In summary, this invention constructs a deep learning-based cross-domain feature extraction model to uniformly model the complex relationships between order demand, delivery personnel capacity, and the road network environment. Based on this, it utilizes dynamic tensors to construct a spatiotemporal manifold space, enabling the multidimensional features of the logistics system to be transformed into geometric attributes within a unified space. This allows the dynamic coupling relationships between order demand distribution, delivery capacity changes, and road conditions to be expressed within the overall space. Furthermore, by establishing an order gravitational field within this spatiotemporal manifold space, orders are abstracted as gravitational sources, and delivery personnel as moving particles. This allows order demand to propagate through the road network in the form of gravity. The gravity is then path-based, combining multi-path states to form a continuous delivery guidance structure within the road network space, ensuring that the movement trends of delivery personnel particles simultaneously consider order distribution characteristics and path reachability. Subsequently, logistics business constraint rules are used to screen the candidate allocation matrices for validity, ensuring that the generated order allocation relationships meet the requirements of road conditions, vehicle load capacity, and delivery timeliness. Building upon this foundation, a virtual global particle is introduced to iteratively optimize the order access sequence. Multiple candidate delivery schemes are generated through repeated randomization optimization. Then, the reconstruction loss generated during neural network training is used as an evaluation criterion to screen the candidate schemes based on feature consistency, thereby selecting the optimal delivery route that best conforms to historical efficient delivery patterns. This method enables collaborative optimization of order allocation and route planning in complex road networks and multi-order scenarios, improving the rationality of logistics delivery decisions and the overall system scheduling efficiency.

[0100] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A deep learning-based optimization method for logistics order allocation, characterized in that, include: By performing cross-domain feature extraction on logistics data from heterogeneous regions, and using deep learning neural networks to decouple and generate dynamic features representing scheduling patterns, a spatiotemporal manifold space for the co-evolution of orders and delivery is constructed. After the deep learning neural network, a feature definition unit is connected. The feature definition unit includes taking orders as the gravitational source, establishing a gravitational field in the spatiotemporal manifold space based on real-time order information, defining the deliveryman as a moving particle driven by the dynamic features, calculating the gravitational force of the particle in the gravitational field to generate a candidate allocation matrix, and using logistics order business constraint rules to enforce constraints on the candidate allocation matrix. Following the feature definition unit, an iterative optimization unit is introduced. The iterative optimization unit includes, under constraints, driving particles to generate delivery tasks in the spatiotemporal manifold space through the candidate allocation matrix, and outputting the final order allocation result through iterative optimization.

2. The deep learning-based logistics order allocation optimization method as described in claim 1, characterized in that: The logistics data in the heterogeneous regions refers to historical logistics information in different delivery regions, including order information, delivery information, and road network information; When performing the cross-domain feature extraction, the preprocessed logistics data in different delivery areas are uniformly encoded so that order information, delivery information and road network information are represented in a unified data structure.

3. The deep learning-based logistics order allocation optimization method as described in claim 2, characterized in that, The dynamic characteristics of the decoupling generation of scheduling rules using deep learning neural networks include: Deep learning neural networks are used to extract features from order information, delivery information, and road network information respectively; By using a cross-attention mechanism to perform joint feature fusion on the order information, delivery person information, and road network information, the coupling relationship between the three is decoupled and reconstructed into a unified dynamic tensor; The dynamic tensor is mapped to the metric operator of the spatiotemporal manifold space, and the scalar curvature of the road network at various points in the manifold space is adjusted through the metric operator, so that the road network traffic impedance and delivery personnel service habits can be mapped to the geometric deformation in the manifold space, thereby constructing a manifold structure that reflects the coupling relationship between order demand and delivery capacity. Based on the geometric deformation of the manifold space, in the spatiotemporal manifold space, the adjusted scalar curvature at each point of the road network is divided by the unadjusted scalar curvature to generate a non-uniformly distributed gravitational conduction coefficient. This allows the intensity of the order's effect on the delivery particle to be adjusted by the spatial geometric characteristics, thereby forming a dynamic feature that can guide the delivery particle to avoid high-resistance regions and converge towards high-value order nodes.

4. The deep learning-based logistics order allocation optimization method as described in claim 3, characterized in that, The gravitational field includes: In the spatiotemporal manifold space, each order to be assigned is mapped as a gravitational source unit with time-varying mass, and a potential energy trap is formed at the local coordinates of the manifold space according to the execution position corresponding to the order. The gravitational force generated by the gravitational source unit of each order is represented as: ; Where t represents the order waiting time; m is the order quality; This represents the distance the delivery person travels to the gravitational source unit along the i-th path; This represents the i-th path; Representing a path Average gravitational conductivity; The gravitational field is generated by representing the gravitational force of each gravitational source unit.

5. The deep learning-based logistics order allocation optimization method as described in claim 4, characterized in that, The kinetic force includes: For any deliveryman particle, calculate the gravitational force of the particle under the action of each gravitational source unit in each reachable path state, lay all the gravitational forces on the corresponding path, and at each position of the road network, add all the laid gravitational forces vectorively to form the direction and magnitude of the path guiding the particle at any position of the road network. Construct a candidate allocation matrix with the orders to be allocated as the horizontal axis and the deliverymen as the vertical axis; each intersection in the matrix represents the matching tendency between the order and the deliveryman; by calculating the vector sum of all gravitational forces on each reachable path between the gravitational source unit represented by the order and the particle represented by the deliveryman, and dividing the result by the path length, the matching tendency under each reachable path is obtained, so that each element in the candidate allocation matrix is ​​a quantum state.

6. The deep learning-based logistics order allocation optimization method as described in claim 5, characterized in that: The constraints on logistics order operations include time-based traffic restrictions, road traffic restrictions, vehicle weight restrictions, and delivery time restrictions within the road network.

7. The deep learning-based logistics order allocation optimization method as described in claim 6, characterized in that: The process of driving particles to generate delivery tasks in the spatiotemporal manifold space involves simplifying all particles into a virtual global particle and using the virtual global particle to complete logistics orders in the road network. Furthermore, by using the shortest path in the road network after scalar curvature adjustment as the fitness function, the order sequence is iteratively optimized to obtain the optimal single-particle delivery solution. The single-particle delivery scheme is divided equally among each delivery person according to the path length adjusted by scalar curvature, resulting in a continuous path equal to the number of delivery persons K. The obtained K consecutive paths are randomly assigned to delivery personnel to obtain the final order allocation result.

8. The deep learning-based logistics order allocation optimization method according to claim 7, characterized in that, The iterative optimization includes: The candidate allocation matrix is ​​updated in real time based on the position of the virtual global particle. The corresponding gravitational source unit is eliminated after each order is completed. The normalized element in the matrix is ​​used as the selection probability of the next position of the virtual global particle to realize the selection of the position of the next order. After completing the iterative optimization, repeat the iterative optimization a preset number of L times to obtain L optimization results; The L optimization results are returned to the neural network, which uses each optimization result to calculate the reconstruction loss and obtains the optimization result with the minimum loss, which is taken as the optimal single-particle delivery scheme.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the deep learning-based logistics order allocation optimization method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the deep learning-based logistics order allocation optimization method according to any one of claims 1 to 8.

Images