Intelligent order allocation method and system considering rider load balance

Abstract

The invention belongs to the technical field of intelligent logistics, and discloses an intelligent order distribution method and system considering rider load balancing, and the method comprises the steps: obtaining the real-time position data, historical order data and dynamic traffic data of a rider in a distribution region, and constructing a load balancing distribution scheme according to the data; performing rider cluster division on the distribution area, generating a load distribution thermodynamic diagram, and calculating a load balancing index based on the load distribution thermodynamic diagram; based on the load balancing index and the load balancing distribution scheme, distribution instruction generation is carried out on to-be-distributed orders, a hierarchical distribution sequence is formed through priority ranking, and an optimal order distribution strategy is obtained after matching optimization; according to the configuration instruction, order pushing is adjusted, the execution state is monitored, the load efficiency is evaluated, an adjustment and optimization parameter is obtained, and the adjustment and optimization parameter is fed back to the distribution scheme; according to the method, the dynamic optimal configuration of the distribution network is realized, the load of a rider is effectively balanced, and the distribution efficiency and the service quality are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent distribution, more particularly, the present application relates to an intelligent order allocation method and system considering rider load balancing. BACKGROUND

[0002] Take-out delivery service is an important part of modern life consumption, and order allocation efficiency directly affects platform operating costs and user experience. Traditional order allocation mainly relies on manual scheduling or simple nearest allocation principle. These methods not only require a large number of dispatchers, but also lead to uneven distribution of delivery resources, slow response during peak hours, and large income gap among riders, which seriously affects delivery efficiency and platform operating costs.

[0003] With the development of information technology, intelligent allocation systems based on geographic information systems and simple algorithms have emerged one after another, but these systems generally have defects such as static allocation rules, low adaptability, and single optimization target, and cannot cope with complex and variable delivery environments.

[0004] Data-driven intelligent decision-making as a modern operational optimization method has been widely used in the field of logistics and distribution in recent years. When combined with big data analysis and artificial intelligence algorithms, it can extract spatio-temporal patterns from historical delivery data, predict order distribution and traffic conditions. Studies have shown that there are significant differences in order characteristics, rider behavior patterns, and traffic conditions in different regions and time periods, and these differences will have a systematic impact on delivery efficiency and rider load balancing.

[0005] Existing technologies are difficult to effectively capture and integrate multi-dimensional dynamic data, leading to unreasonable allocation of delivery resources. Order peak prediction accuracy is insufficient, with a high error rate, making it difficult for the platform to schedule rider resources in advance to respond to demand fluctuations. Traditional allocation algorithms lack the ability to identify individual differences among riders, and cannot make personalized allocations based on rider speed characteristics, path selection preferences, and work time patterns. Regional division is too general and does not take into account road network topology and traffic dynamics, leading to serious load imbalance between adjacent regions. Existing systems also lack real-time feedback optimization mechanisms and cannot dynamically adjust allocation strategies based on execution conditions, especially in abnormal situations such as order surges, extreme weather, or traffic congestion, where the system response is slow. Lack of quantitative models to evaluate the matching degree of rider ability and order characteristics results in low order allocation efficiency and high delivery costs.

[0006] In view of the above problems, the present application provides an intelligent order allocation method and system considering rider load balancing. SUMMARY

[0007] In order to overcome the above-mentioned defects of the prior art, in order to achieve the above-mentioned purpose, the present application provides the following technical solutions: An intelligent order allocation method considering rider load balancing, comprising: Step S1: Obtain real-time position data of riders in the delivery area, order generation historical data and traffic dynamic data; construct a rider path optimization model according to the real-time position data of the riders and the traffic dynamic data to obtain path optimization prediction results; construct an order demand prediction model according to the order generation historical data and the path optimization prediction results to obtain demand peak prediction data; generate a preliminary allocation strategy based on the demand peak prediction data to obtain a load balancing allocation scheme; Step S2: Divide the delivery area into rider clusters to obtain rider cluster data; perform load space-time analysis according to the rider cluster data and the path optimization prediction results to obtain a load distribution heat map; calculate the load balancing index of the load distribution heat map; Step S3: Generate allocation instructions for the to-be-allocated orders based on the load balancing index and the load balancing allocation scheme to obtain an allocation instruction set; perform real-time priority sorting on the allocation instruction set to obtain a hierarchical allocation sequence; perform matching optimization according to the hierarchical allocation sequence and the load balancing index to obtain an optimal order allocation strategy; generate rider task parameters based on the optimal order allocation strategy to obtain task parameter configuration instructions; Step S4: Adjust the order pushing of each rider according to the task parameter configuration instructions to obtain rider task real-time configuration data; perform execution state monitoring on the rider task real-time configuration data to obtain task execution state data; perform load efficiency evaluation according to the task execution state data to obtain load efficiency evaluation results; perform feedback tuning based on the load efficiency evaluation results to obtain tuning parameters; apply the tuning parameters to the load balancing allocation scheme.

[0008] Further, the process of obtaining the load balancing allocation scheme comprises: Collect historical trajectory data of riders and perform space-time pattern analysis to obtain a rider behavior feature library; and collect real-time traffic data and weather influence data to obtain environmental interference factor data; Obtain delivery network area division data; construct a path network according to the delivery network area division data to obtain path network topology data; collect order generation time series data through an order management center to obtain historical order data; Construct a rider path optimization model based on the rider behavior feature library and the environmental interference factor data to obtain path behavior pattern data; extract dynamic features from the path behavior pattern data to obtain path dynamic features; Perform spatial path prediction according to the path behavior pattern data and the path dynamic features to obtain a path coverage probability atlas; divide the path coverage probability atlas into time periods to obtain segmented path density distribution data; mapping the segmented path density distribution data to the path network topology data to obtain order load prediction data; constructing an order demand prediction model according to historical order data and the order load prediction data to obtain demand fluctuation prediction data; performing peak detection and abnormality analysis on the demand fluctuation prediction data to obtain a demand burst index; performing rider resource distribution analysis based on the path network topology data to obtain a rider resource capability graph; and performing available capacity evaluation on the rider resource capability graph to obtain a rider available matrix; generating a load balancing allocation scheme according to the demand burst index and the rider available matrix.

[0009] Further, the load balancing index acquisition process includes: performing geographic gridding division on the delivery area, and performing density clustering analysis according to real-time position data of the riders to obtain rider cluster data; performing spatio-temporal load vector decomposition according to the rider cluster data and the path optimization prediction result to extract a task backlog trend in the cluster; performing thermal visualization rendering on the task backlog trend in the cluster to generate a load distribution thermal map; performing entropy calculation and variance balance degree measurement on the load distribution thermal map to obtain a load balancing index, including a cluster-to-cluster load deviation rate and an overall spatio-temporal balance score.

[0010] Further, the task parameter configuration instruction acquisition process includes: obtaining to-be-allocated orders, and performing state analysis and priority marking on the to-be-allocated orders to obtain an order state matrix; and performing order-rider matching degree calculation based on the load balancing index and the load balancing allocation scheme to obtain a matching fitness matrix; generating allocation decision according to the order state matrix and the matching fitness matrix to obtain an initial allocation instruction set; performing timeliness evaluation on the initial allocation instruction set according to the segmented path density distribution data to obtain instruction timeliness level data; performing real-time priority sorting on the allocation instruction set based on the instruction timeliness level data to obtain a hierarchical allocation sequence; performing order matching simulation according to the hierarchical allocation sequence and the load balancing index to obtain a matching prediction result; and performing conflict detection and resolution on the matching prediction result to obtain a conflict resolution strategy; generating an optimal order allocation strategy based on the matching prediction result and the conflict resolution strategy; and converting the optimal order allocation strategy to generate a task parameter configuration instruction.

[0011] Further, the order push adjustment of each rider according to the task parameter configuration instruction acquisition process includes: The task parameter configuration instruction is analyzed and processed to obtain a task parameter set; the task parameter set is used for parameter updating and pushing of each jockey to obtain jockey parameter updating data; the jockey parameter updating data is verified for pushing effectiveness to obtain a verification result; The jockey task performance is monitored according to the verification result to obtain task performance data; the task performance data is collected in real time to obtain task execution state data; Efficiency evaluation data is obtained by calculating efficiency indicators according to the task execution state data; delivery quality indicators are obtained by evaluating the delivery quality of the task execution state data; Based on the efficiency evaluation data and the delivery quality indicators, a comprehensive load performance evaluation result is obtained; The load performance evaluation result is analyzed for improvement potential to obtain improvement direction data; parameter tuning strategies are generated according to the improvement direction data to obtain tuning parameters; the tuning parameters are fed back to the load balancing distribution scheme.

[0012] Further, the jockey path optimization model is constructed based on the jockey behavior feature library and the environmental interference factor data to obtain path behavior pattern data, including: The jockey behavior feature library is analyzed for jockey type classification to obtain jockey type data; and the interference sensitivity is extracted according to the environmental interference factor data to obtain a path interference index; Based on the jockey type data, a reinforcement learning path optimization network is constructed to obtain a path optimization base model; and the path optimization base model is adjusted in parameters adaptively using the path interference index to obtain a jockey path optimization model, and the path behavior pattern data is generated based on the jockey path optimization model.

[0013] Further, the path network topology data is represented by a weighted directed graph structure, the network nodes in the graph structure include distribution sites, merchants, customer addresses and transportation hubs, the connection relationships between nodes include distribution paths and logical associations; and the connection weights between nodes are composed of several indicators, including distance, estimated delivery time, road grade indicators.

[0014] Further, the load performance evaluation result is obtained by: The efficiency evaluation data is normalized in multiple dimensions to obtain normalized efficiency indicators; The delivery quality indicators are mapped for user feedback to obtain satisfaction evaluation data; A load performance evaluation framework is constructed, the normalized efficiency indicators and the satisfaction evaluation data are dynamically weighted and fused to obtain a comprehensive performance index, and the load performance evaluation result is generated according to the comprehensive performance index.

[0015] Further, the initial allocation instruction set is evaluated in time effectiveness according to the segmented path density distribution data, to obtain instruction time effectiveness level data, including: The rider's expected delivery path is decomposed into road segments, and the path is mapped to the segmented path density distribution data to obtain a congestion impact index; Based on the rider type data and historical behavior patterns, the rider movement efficiency is predicted in combination with the path characteristics to obtain a rider movement speed estimate; According to the historical data and real-time state of the merchant, the order pickup waiting time is calculated, and the order processing time is estimated by combining the path length and the rider movement speed to obtain an order processing time prediction value; The expected completion time of the allocation instruction is calculated by combining the congestion impact index, the rider movement speed estimate, and the order processing time prediction value, to obtain time effectiveness prediction data; The time effectiveness prediction data is compared with the order expected delivery time, and the instruction is classified in time effectiveness according to the time difference to obtain instruction time effectiveness level data.

[0016] An intelligent order allocation system considering rider load balancing, comprising: A load allocation module; for obtaining rider real-time position data, order generation historical data and traffic dynamic data in the delivery area; constructing a rider path optimization model according to the rider real-time position data and the traffic dynamic data to obtain a path optimization prediction result; constructing an order demand prediction model according to the order generation historical data and the path optimization prediction result to obtain demand peak prediction data; generating a preliminary allocation strategy based on the demand peak prediction data to obtain a load balancing allocation scheme; An index acquisition module; for dividing the delivery area into rider clusters to obtain rider cluster data; performing load space-time analysis based on the rider cluster data and the path optimization prediction result to obtain a load distribution heat map; calculating the load balancing index based on the load distribution heat map; An order allocation module; for generating allocation instructions for the to-be-allocated orders based on the load balancing index and the load balancing allocation scheme to obtain an allocation instruction set; performing real-time priority sorting on the allocation instruction set to obtain a hierarchical allocation sequence; matching and optimizing the hierarchical allocation sequence and the load balancing index to obtain an optimal order allocation strategy; generating rider task parameters based on the optimal order allocation strategy to obtain task parameter configuration instructions; A feedback optimization module; for adjusting the order pushing of each rider according to the task parameter configuration instructions to obtain rider task real-time configuration data; monitoring the execution state of the rider task real-time configuration data to obtain task execution state data; evaluating the load efficiency based on the task execution state data to obtain a load efficiency evaluation result; performing feedback tuning based on the load efficiency evaluation result to obtain tuning parameters; applying the tuning parameters to the load balancing allocation scheme.

[0017] The technical effects and advantages of the intelligent order distribution method and system considering rider load balancing of the present application are as follows: The present application realizes accurate modeling and optimized scheduling of the dynamic matching relationship between riders and orders in the distribution network. The present application integrates the originally dispersed rider positions, order distribution, and traffic conditions and other data into a systematic load balancing decision basis. The present application changes the response mode of the take-out order distribution, and changes passive response into active prediction and advance scheduling, greatly reducing the distribution delay rate and operation cost. In actual application, the distribution platform no longer needs to maintain excessive rider resources to respond to order fluctuations, but can accurately allocate resources according to the load balancing prediction results, thereby improving the overall efficiency of the distribution system. The closed-loop feedback optimization feature of the present application realizes continuous adaptive adjustment of the distribution strategy, avoiding the problems of uneven load and resource waste in the traditional method. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 FIG. 1 is a schematic diagram of an intelligent order distribution method considering rider load balancing of the present application; Figure 2 FIG. 2 is a schematic diagram of an intelligent order distribution system considering rider load balancing of the present application. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0020] Embodiment 1 Please refer to Figure 1 The intelligent order distribution method considering rider load balancing described in the present embodiment includes the following steps: Step S1: Obtain the real-time position data of riders in the distribution area, order generation historical data, and traffic dynamic data; and perform path prediction based on the real-time position data of riders and the traffic dynamic data to obtain a path coverage probability atlas; and perform order demand prediction based on the order generation historical data to obtain demand peak prediction data; and generate a preliminary distribution strategy based on the same to obtain a load balancing distribution scheme; It needs to be further explained that, in the specific implementation process, the process of obtaining the load balancing distribution scheme includes the following steps: Collect historical trajectory data of riders and perform spatiotemporal pattern analysis to obtain a rider behavior feature library; and collect real-time traffic data and weather influence data to obtain environmental interference factor data; Obtain distribution network area division data; construct a path network based on the distribution network area division data to obtain path network topology data; collect order generation time sequence data through an order management center to obtain historical order data; Construct a rider path optimization model based on the rider behavior feature library and the environmental interference factor data to obtain path behavior pattern data; extract dynamic features from the path behavior pattern data to obtain path dynamic features; Perform spatial path prediction based on the path behavior pattern data and the path dynamic features to obtain a path coverage probability atlas; divide the path coverage probability atlas into time segments to obtain segmented path density distribution data; Map the segmented path density distribution data to the path network topology data to obtain order load prediction data; construct an order demand prediction model based on the historical order data and the order load prediction data to obtain demand fluctuation prediction data; Perform peak detection and anomaly analysis on the demand fluctuation prediction data to obtain a demand burst index; Perform rider resource distribution analysis based on the path network topology data to obtain a rider resource capability atlas; perform available capacity evaluation on the rider resource capability atlas to obtain a rider available matrix; Generate a load balancing allocation scheme based on the demand burst index and the rider available matrix.

[0021] In this embodiment, rider historical trajectory data is collected and analyzed in space-time mode to obtain a rider behavior feature library; the collection methods include GPS positioning data, mobile trajectory records, distribution time efficiency data, and rider working state records, covering the main distribution period in the past 6 months; the collected rider historical trajectory data is analyzed by K-means clustering algorithm, and the distribution path and key stopping points are identified and extracted; in the clustering process, the clustering center position is optimized by iteration, and the corresponding distribution trajectory points are classified into the same cluster; the number of clustering clusters is usually set to 8 to 15 to represent the number of main distribution areas, and the convergence threshold is set to 0.01 to represent the minimum threshold of the change of the clustering center; through clustering analysis, high-frequency sections, key distribution areas, and load intensity change patterns in the distribution path are identified, and a rider behavior feature library is constructed based thereon. At the same time, traffic environment data and real-time change data are collected, including road congestion index, weather conditions, holiday identification, time period characteristics, and regional heat, etc. dynamic indicators, to form environmental interference factor data.

[0022] The delivery network area division data is obtained, including network deployment information such as delivery station location coordinates, coverage range, delivery radius, and rider distribution; path network topology data is obtained by constructing a path network according to the delivery network area division data; the path network topology data is represented by a weighted directed graph structure, wherein the network nodes in the graph structure include delivery stations, merchants, customer addresses, and transportation hubs, the connection relationship between nodes includes delivery paths and logical associations, and the connection weight between nodes is composed of multiple indicators such as distance, estimated delivery time, road level, etc.; it can be used to describe the spatial and logical connection relationship of the delivery network, and provide spatial reference for subsequent load allocation; the logical relationship includes site-area association, merchant-site association, and merchant-customer flow association, etc. The order generation time series data is collected through the order management center to obtain historical order data; including order generation quantity, order type distribution, delivery address distribution and other indicators at each time point; and time series analysis is performed on the historical order data to identify daily order fluctuation patterns, periodic variation rules and sudden order events; A rider path optimization model is constructed based on the rider behavior characteristic library and environmental interference factor data; the DQN network is used as the backbone network to construct the rider path optimization model, and the decision dependence relationship of the rider delivery behavior and the influence weight of the environmental factors are captured; the model input includes rider historical trajectory data and environmental interference factor data, and the output is the predicted path selection behavior; and the DQN network can be used to process decision sequence data with long-term reward dependence, and can effectively capture the path selection mode and environmental response characteristics in the rider delivery behavior; the path behavior mode data of different types of riders is generated through the rider path optimization model, including predicted path selection probability, delivery time estimation, speed distribution, and stay mode, etc.

[0023] The path behavior mode data is dynamically characterized to obtain path dynamic characteristics; the spatial pattern, time pattern, and load variation characteristics of the delivery activity are analyzed; and through multi-scale feature extraction, the variation characteristics of the delivery activity at different spatio-temporal scales are identified, including hourly, daily, and weekly changes, to obtain path dynamic characteristics; The path behavior mode data and path dynamic characteristics are used for spatial path prediction to obtain a path coverage probability atlas; the Monte Carlo simulation is used to simulate the rider group delivery trajectory in the delivery activity, and the path coverage probability atlas is generated by collecting and evaluating the rider group delivery trajectory, representing the probability distribution of different areas being covered by riders at different time points; The path coverage probability graph is divided into time windows, and the path density distribution of each region in each time window is calculated, which reflects the expected rider coverage frequency per unit area per unit time and is a key indicator for quantifying the strength of delivery capacity. Through time window division, segmented path density distribution data is obtained, which can be used to describe the spatio-temporal evolution process and dynamic change characteristics of delivery activities. The segmented path density distribution data is mapped to the path network topology data to establish a mapping relationship between delivery capacity and order load. Through the corresponding mapping relationship, the support degree of rider delivery capacity to order processing demand is quantified to obtain order load prediction data. An order demand prediction model is constructed based on historical order data and order load prediction data, and a gradient boosting decision tree method is used to predict the order demand of network nodes. The order demand prediction model uses ensemble learning of multiple decision trees, each tree is optimized based on the residual of the previous tree, and has strong nonlinear fitting ability and feature importance recognition ability. Through the model, demand fluctuation prediction data is generated to describe the change trend and fluctuation law of order demand of each node in the network in the future time period.

[0024] The demand fluctuation prediction data is subjected to peak value identification and abnormality analysis, and a peak value detection method based on statistical threshold is used to identify order demand mutation points. The influence degree of the identified mutation points is analyzed, and the demand burst index is calculated, which comprehensively considers the order demand change amplitude, delivery density influence and node importance, providing a quantitative indicator for order allocation priority. The higher the burst index, the higher the priority of order allocation at that time point to prevent the decline of delivery service quality.

[0025] Based on the path network topology data, the rider resource distribution is analyzed, and the rider resources are abstracted as distributed resource pools, each resource pool containing three types of delivery capacity, time availability and skill level. By analyzing the scheduling relationship and response delay between resource pools, a rider resource capacity map is generated to describe the rider distribution and scheduling capacity of the entire network. The available capacity of the rider resource capacity map is evaluated, and the available resource amount of each resource pool at different time points is calculated. According to the resource availability and scheduling constraints, a rider availability matrix is generated, which comprehensively describes the feasibility and constraint conditions of rider scheduling in the network.

[0026] A load balancing allocation scheme is generated according to the demand burst index and the rider availability matrix. A target function is constructed, which includes minimizing delivery cost, minimizing delivery delay risk and maximizing rider utilization rate. Restricted by the rider availability matrix and network constraints, the target function is solved by genetic algorithm to generate a load balancing allocation scheme, including allocation schedule, rider scheduling path and allocation strategy, which realizes dynamic optimization configuration of order allocation and effectively responds to the spatio-temporal changes of delivery demand.

[0027] It needs to be further explained that in the specific implementation process, the rider path optimization model is constructed based on the rider behavior feature library and the environmental interference factor data, and the path behavior mode data is obtained, including: The rider type data is obtained by performing rider type classification analysis on the rider behavior feature library; and the path interference index is obtained by performing interference sensitivity extraction according to the environmental interference factor data. The rider path optimization model is obtained by constructing the reinforcement learning path optimization network based on the rider type data, and performing self-adaptive parameter adjustment on the path optimization basic model by using the path interference index, and the path behavior mode data is generated based on the rider path optimization model.

[0028] In this embodiment, when performing rider type classification analysis on the rider behavior feature library, a multi-dimensional feature clustering and behavior pattern recognition method is adopted; first, five dimensions of key features including delivery speed feature, path selection preference, working time mode, service area preference and emergency response capability are extracted from the rider behavior feature library, wherein the delivery speed feature is calculated by analyzing the historical delivery data of the rider, including average delivery speed, speed variance and speed adaptability of different road sections; the path selection preference is obtained by statistically analyzing the decision tendency of the rider in the multi-path selection scene, including the selection frequency of the shortest distance path, the time optimal path and the congestion avoidance path; the working time mode analyzes the active time period distribution, the continuous working time length and the rest interval law of the rider; the service area preference analyzes the adaptability and delivery efficiency difference of the rider to different area types; the emergency response capability evaluates the processing capability of the rider in the case of sudden orders, special weather or traffic anomalies; Based on the extracted key features, and by combining a hierarchical clustering method to classify the riders, the riders with similar features are classified into the same type through clustering iteration optimization, and each type corresponds to different rider working modes, to obtain a classification result, including high-efficiency riders, stable riders, flexible riders and professional riders; wherein the high-efficiency riders are characterized by fast delivery speed, optimized path selection and concentrated working time; the stable riders are characterized by stable delivery quality, fixed service area and strong emergency handling capability; the flexible riders are characterized by strong adaptability, flexible working time and multi-region service capability; the professional riders are characterized by strong special order processing capability, high skill level and high customer satisfaction; through classification analysis, the rider type data is obtained.

[0029] When extracting the interference sensitivity according to the environmental interference factor data, a multi-factor influence evaluation framework is used to quantify the specific influence degree of different environmental factors on each type of rider. The environmental interference factors mainly include five core dimensions: traffic congestion status, weather change influence, time period busy degree, regional complexity, and emergency frequency. For traffic congestion sensitivity analysis, by comparing the delivery performance differences of riders in three road conditions: smooth, general congestion, and severe congestion, the sensitivity of different types of riders to traffic conditions is identified. Efficiency priority riders are most sensitive to congestion and actively seek alternative paths. Robust conservative riders have moderate sensitivity to congestion and prefer to wait for congestion to ease. Flexible adaptive riders can quickly respond to changes in congestion. Experience-oriented riders rely on past experience to judge the impact of congestion. For weather influence sensitivity, by analyzing the delivery efficiency changes of riders under different conditions such as sunny, rainy, and severe weather, the influence weight of weather factors is quantified. Time period busy degree sensitivity is evaluated by comparing the delivery performance differences between peak and off-peak periods. By integrating the sensitivity analysis results of each dimension, a weighted comprehensive calculation is used to calculate the path interference index of each rider type, which reflects the expected change degree of rider path selection behavior under different environmental conditions.

[0030] When constructing the reinforcement learning path optimization network based on rider type data, a DQN network architecture is used to combine type-specific features for personalized modeling. The network design uses a hierarchical structure, including a state perception layer, a feature extraction layer, a type adaptation layer, a decision reasoning layer, and an action output layer. The state perception layer is responsible for receiving and processing input information, including current geographic location, target delivery address, real-time traffic conditions, weather conditions, time information, and other multi-dimensional state features. The feature extraction layer uses a deep neural network structure to extract high-level abstract representations of state features through multiple layers of nonlinear transformation, identifying complex pattern association relationships. The type adaptation layer is the core innovation point of this network, which sets up special parameter configurations and weight adjustment strategies for different types of riders based on rider type data. The network parameters of efficiency priority riders are biased towards fast path search and time optimization. The parameter configuration of robust conservative riders focuses on path safety and reliability evaluation. The network of flexible adaptive riders has stronger environmental adaptability and path adjustment flexibility. The network of experience-oriented riders relies more on historical path data and experience weights. The decision reasoning layer uses a dual-network structure, including a main decision network and a target evaluation network, to generate the optimal path selection strategy through comparative learning and value evaluation. The action output layer generates specific path selection suggestions based on the decision results, including path priority ranking and alternative scheme recommendation.

[0031] When the path interference index is used to adaptively adjust the parameters of the path optimization base model, dynamic weight adjustment and environmental adaptive learning mechanism are adopted to realize real-time optimization of the model. According to the real-time calculation of the path interference index, the key parameters in the network are dynamically adjusted, including the learning rate, the exploration strategy and the reward weight, etc. When the environmental interference index is high, the system increases the learning rate of the model to speed up the adaptation to the new environmental conditions, and at the same time increases the proportion of exploration behavior to encourage the model to try new path selection strategies. When the environment is relatively stable, the system reduces the learning rate to maintain the stability of the model and reduce unnecessary parameter adjustment. The adjustment of the exploration strategy adopts an intelligent balance mechanism to seek the optimal balance point between making full use of existing knowledge and exploring new possibilities. The dynamic adjustment of the reward mechanism considers the impact of environmental interference on the distribution effect. When the interference is small, the system gives higher performance rewards. When the interference is large, the system adjusts the reward standard accordingly to avoid unreasonable punishment. Through this adaptive adjustment mechanism, the rider path optimization model can continuously learn and optimize, and maintain high path selection accuracy and adaptability under different environmental conditions.

[0032] When the rider path optimization model generates path behavior pattern data, a large-scale simulation experiment and behavior pattern statistical analysis method are used for systematic data generation. Through the Monte Carlo simulation technology, a large number of path selection simulation experiments are conducted for each rider type under different environmental conditions, and thousands of independent simulations are conducted for each type of environmental combination to ensure the reliability and representativeness of the statistical results. During the simulation process, the system records the detailed information of each path selection, including the specific path selected, the decision time, the expected delivery time, the actual execution effect and other key data. Through statistical analysis of the simulation results, the path behavior pattern characteristics of different rider types are extracted and identified. The path selection preference is quantified by statistically analyzing the frequency and conditions of each type of path being selected, forming a path selection probability distribution map. The delivery time pattern is identified by analyzing the time distribution characteristics under different path conditions, identifying the regularity and variability of time prediction. The path length preference is formed by statistical analysis of the rider's selection tendency for paths of different distances, forming a length preference distribution model. The stop behavior pattern is identified by spatial analysis technology to identify the high-frequency stop area and stop time distribution of the rider; all these behavior pattern characteristics are structured, organized and coded to form a standardized path behavior pattern data set, which contains complete information such as rider type identification, environmental condition parameters, behavior feature vector, and spatiotemporal distribution characteristics, providing accurate and reliable behavior modeling basis data for subsequent path prediction, order allocation and system optimization.

[0033] Step S2: divide the delivery area into rider clusters to obtain rider cluster data; perform load spatiotemporal analysis on the rider cluster data and the path coverage probability map to obtain a load distribution heat map; calculate the load balance index of the load distribution heat map; It should be further explained that, in the specific implementation process, the acquisition of load balancing metrics includes: The delivery area is divided into geographical grids, and density cluster analysis is performed based on the rider's real-time location data to obtain rider cluster data; Spatiotemporal load vector decomposition is performed based on rider cluster data and path coverage probability map to extract task backlog trends within the cluster. Perform heatmap visualization rendering of task backlog trends within the cluster to generate a load distribution heatmap. Entropy values ​​and variance balance measures are performed on the load distribution heatmap to obtain load balance indices, including inter-cluster load deviation rate and overall spatiotemporal balance score.

[0034] Specifically, when dividing the delivery area into geographical grids, the geographical boundaries of the delivery area are first determined, including the maximum and minimum longitude, maximum and minimum latitude, and the total area and spatial span of the area are calculated. Based on delivery density and service accuracy requirements, a basic grid size is set. The entire delivery service area is then divided into regular rectangular grid units based on the set basic grid size. Each grid unit is uniquely identified by its lower left corner coordinates, and grid numbers are indexed using a two-dimensional index. Simultaneously, a spatial index structure is established for the grid units, supporting location-based fast queries and neighboring grid searches.

[0035] When performing density clustering analysis based on rider real-time location data, the real-time GPS coordinates of all riders are first mapped to corresponding rectangular grid cells, and the number and distribution density of riders within each grid cell are counted. Rider density is calculated by dividing the number of riders within the grid by the grid area, with units of riders per square kilometer. For areas with uneven rider distribution, a kernel density estimation method is used for smoothing to reduce the impact of data noise on the clustering results. Then, density clustering is performed based on the number and distribution density of riders, and the clustering process includes three stages: core point identification, cluster expansion, and noise point processing. In the core point identification stage, the neighborhood rider density of each grid cell is calculated, and grid cells with a density exceeding a threshold are marked as core points. In the cluster expansion stage, clusters are expanded from the core points... Initially, grid cells with reachable density are recursively added to the same cluster to form connected rider cluster regions. In the noise point processing stage, isolated low-density grid cells are merged, assigned to the nearest rider cluster, or marked as independent cells. Through cluster analysis, several rider clusters with different characteristics are identified, each containing geographically proximate rider groups with similar densities. Rider cluster data includes cluster center coordinates, cluster coverage, number of riders within the cluster, cluster density index, and cluster shape characteristics. It should be noted that the cluster center coordinates are calculated by a weighted average of the coordinates of all grid cells within the cluster, with the weight being the rider density of each grid cell. The cluster coverage is determined using the minimum bounding rectangle or convex hull algorithm. The cluster density index reflects the compactness and cohesion of the cluster. Cluster shape characteristics include geometric indices such as aspect ratio, roundness, and dispersion.

[0036] When performing spatiotemporal load vector decomposition based on rider cluster data and path coverage probability maps, a combination of time series decomposition and spatial interpolation is adopted. First, a spatiotemporal load matrix is ​​constructed, where the row dimension of the matrix represents different rider clusters, the column dimension represents different time windows, and the matrix elements represent the load intensity of a specific cluster at a specific time. Load intensity is comprehensively considered based on multiple factors, including the number of orders within the cluster, the complexity of delivery tasks, the effectiveness of route optimization, and rider efficiency. The time dimension decomposition employs a seasonal decomposition method, dividing the load time series into three parts: trend component, periodic component, and random component. The trend component reflects the long-term trend of load changes and is extracted using moving average or multinomial fitting methods. The periodic component reflects the regular fluctuations of load and is identified by Fourier transform or wavelet analysis to determine the main periodic characteristics. The random component reflects unpredictable load fluctuations and its impact is quantified through residual analysis. The spatial dimension decomposition uses principal component analysis to identify the main spatial factors affecting load distribution. Load characteristics from different clusters are used as variables, and dimensionality reduction analysis is used to extract the main spatial load patterns. The first principal component typically reflects the overall load level, while the second principal component reflects the degree of spatial imbalance in load. Through principal component analysis, complex spatial load distribution characteristics can be described using fewer comprehensive indicators.

[0037] The task backlog trend extraction within the cluster is achieved through comprehensive analysis of the spatiotemporal decomposition results; the task backlog trend is defined as the cumulative change pattern of the number of pending orders within the cluster over time; the backlog trend calculation considers three key factors: order generation rate, rider processing capacity, and external environmental influences; the order generation rate is calculated through historical data statistics and demand forecasting models; rider processing capacity is assessed based on the average delivery efficiency and available working time of riders within the cluster; external environmental influences include the combined effects of factors such as traffic conditions, weather changes, and special events.

[0038] Backlog trend prediction employs a time series forecasting model, combining an ARIMA model and machine learning methods. The ARIMA model captures the time dependence and trend characteristics of backlog data, while machine learning methods handle nonlinear relationships and complex interactions. The prediction model outputs the task backlog level of each cluster within a future time window, providing forward-looking information for load balancing decisions. Backlog trend data is categorized into four levels based on severity: normal, mild backlog, moderate backlog, and severe backlog, with different warning thresholds and response strategies corresponding to each level.

[0039] When rendering a heatmap to visualize the backlog trend of tasks within the cluster, multi-level color mapping and smooth interpolation techniques are used. The heatmap rendering process first maps the task backlog data to a color space to establish the correspondence between numerical values ​​and colors. The color mapping uses a gradient color spectrum, typically using blue to represent low backlog, green to represent medium backlog, yellow to represent high backlog, and red to represent severe backlog. The color intensity is proportional to the backlog value to ensure that the visual effect can intuitively reflect the load distribution characteristics.

[0040] The heatmap generation employs a bilinear interpolation algorithm to spatially smooth the discrete grid data. The interpolation process considers the overlap between adjacent grid cells and distance weights to calculate the degree of overlap at interpolation points. The distance weights utilize an inverse distance weighting method, where closer grid cells have a greater impact on the interpolation result. Through interpolation, a continuous and smooth heatmap is generated, avoiding visual breaks between discrete data points.

[0041] Heatmap rendering also includes contour plotting and annotation. Contours are used to identify the boundaries of areas with similar backlog levels, facilitating the identification of spatial patterns in load distribution. Annotations include cluster identifiers, backlog values, and timestamps, providing detailed data support for decision-makers. The heatmap supports dynamic display across multiple time windows, allowing observation of the temporal evolution of load distribution through timeline control. The final generated load distribution heatmap visually displays the task backlog status of different regions at different times, providing visual decision support for load balancing strategy formulation.

[0042] When calculating entropy and measuring variance balance of load distribution heatmaps, a combination of information entropy theory and statistical variance analysis is used. Entropy calculation is used to quantify the uncertainty and complexity of load distribution, while variance balance measurement is used to evaluate the uniformity of load distribution in spatial and temporal dimensions.

[0043] The load distribution entropy value is calculated using the Shannon entropy formula: Where H(X) represents the information entropy value of the load distribution, n represents the total number of rectangular grid cells, and p(xi) represents the load probability of the i-th rectangular grid cell, which is calculated by dividing the load value of the grid cell by the total load value. Variance balance calculation includes two dimensions: spatial variance and temporal variance. Spatial variance reflects the degree of load difference between different regions at the same point in time, while temporal variance reflects the degree of load fluctuation within the same region at different points in time. The inter-cluster load deviation rate is obtained by calculating the relative deviation between the load value of each cluster and the average load value. The overall spatiotemporal equilibrium score comprehensively considers both spatial equilibrium and temporal stability, and the calculation formula is as follows: Balance_score = α × (1 - Spatial_variance) + β × (1 - Temporal_variance); where Balance_score represents the overall spatiotemporal balance score, Spatial_variance represents the standardized spatial variance, Temporal_variance represents the standardized temporal variance, and α and β are the weighting coefficients for space and time, respectively. Typically, α is set to 0.6 and β to 0.4, emphasizing the importance of spatial balance. The balance score ranges from 0 to 1, with higher scores indicating a more balanced load distribution. The cost impact factor represents a penalty term related to delivery costs. Its role is to promote the search for optimal delivery efficiency and load balance during the delivery process while ensuring that delivery costs remain within a preset range. Load balancing metrics also include auxiliary indicators such as peak concentration, distribution stability, and early warning index. Peak concentration measures the degree of concentration in peak load areas and is derived by calculating the spatial clustering of high-load grid cells. Distribution stability assesses the temporal consistency of load distribution patterns and is calculated through correlation analysis of load distributions in adjacent time windows. The early warning index predicts potential future load imbalance risks based on backlog trends and historical anomaly patterns.

[0044] Step S3: Generate allocation instructions for the orders to be allocated based on load balancing metrics and load balancing allocation scheme to obtain an allocation instruction set; sort the allocation instruction set in real time to obtain a hierarchical allocation sequence; perform matching and optimization based on the hierarchical allocation sequence and load balancing metrics to obtain the optimal order allocation strategy; generate rider task parameters based on the optimal order allocation strategy to obtain task parameter configuration instructions. It should be further explained that, in the specific implementation process, the process of obtaining task parameter configuration instructions includes: Obtain the orders to be assigned, perform status analysis and priority marking on them to obtain the order status matrix; calculate the order-rider matching degree based on load balancing metrics and load balancing allocation scheme to obtain the matching fitness matrix; The allocation decision is generated based on the order status matrix and the matching fitness matrix to obtain the initial allocation instruction set; The timeliness of the initial allocation instruction set is evaluated based on the segmented path density distribution data to obtain instruction timeliness level data; Based on instruction timeliness level data, the allocation instruction set is sorted by real-time priority to obtain a hierarchical allocation sequence; Order matching simulation is performed based on hierarchical allocation sequence and load balancing indicators to obtain matching prediction results; then conflict detection and resolution are performed on the matching prediction results to obtain conflict resolution strategies. The optimal order allocation strategy is generated based on the matching prediction results and conflict resolution strategy; the task parameter configuration instructions are then generated based on the optimal order allocation strategy.

[0045] In this embodiment, when acquiring orders to be assigned and performing status analysis and priority marking, a multi-dimensional order feature extraction framework is used to process different types of delivery orders. First, new order information is obtained in real time through the API interface of the order management center, including basic attributes such as order ID, delivery address, product type, order time, and expected delivery time. Then, the order status is analyzed, and the orders are divided into four states: pending assignment, assigned, in transit, and completed. At the same time, the urgency of the order is analyzed, and an urgency index is calculated based on factors such as remaining delivery time, customer level, and product value. Priority marking adopts a multi-level scoring mechanism, comprehensively considering three dimensions: time sensitivity, distance factor, and value weight, and using a weighted summation formula to calculate the order priority score. Through matrix processing, the status information, priority scores, and feature attributes of all orders are organized into a structured order status matrix, providing a data foundation for subsequent matching calculations.

[0046] When calculating order-rider matching based on load balancing metrics and allocation schemes, a multi-factor matching evaluation model is adopted, comprehensively considering three core dimensions: distance matching, capability matching, and load matching. Distance matching is calculated using a modified Euclidean distance formula: D_match = exp(-α0 × distance ÷ D_max), where D_match represents the distance matching degree; α0 is the attenuation coefficient; distance is the actual delivery distance; and D_max is the maximum acceptable delivery distance. Capability matching assesses the suitability of a rider's delivery capabilities to order demands, including factors such as load capacity, delivery experience, and familiarity with product types. Load matching analyzes the current workload of riders, calculating the degree of matching between remaining rider capacity and time windows. A fuzzy comprehensive evaluation method is used to weightedly fuse the matching degrees of the three dimensions to form a comprehensive matching fitness. By batch calculating the matching fitness of all orders with all available riders, a complete matching fitness matrix is ​​constructed.

[0047] When generating allocation decisions based on the order status matrix and matching fitness matrix, a multi-objective optimization decision framework is adopted. The decision-making process first establishes a set of constraints, including rider capacity constraints, time window constraints, geographical range constraints, and service quality constraints. The rider capacity constraint ensures that the number of orders allocated to each rider does not exceed their processing capacity; the time window constraint ensures that orders are delivered within a specified time; the geographical range constraint limits the rider's service radius; and the service quality constraint maintains customer satisfaction standards. Under these constraints, a hybrid strategy combining particle swarm optimization (PSO) and heuristic search is used to generate an initial allocation scheme. PSO prioritizes the order-rider combination with the highest matching degree, while heuristic search avoids the local optimum trap of PSO through local optimization. The decision generation process adopts a batch processing mechanism, grouping the orders to be allocated according to urgency and geographical location, with each group undergoing independent initial allocation, followed by coordinated optimization at the global level. The final result is an initial allocation instruction set containing the allocation objects, allocation parameters, and execution sequence.

[0048] The timeliness of the initial allocation instruction set is evaluated based on segmented path density distribution data. This evaluation considers three key factors: path congestion level, rider mobility efficiency, and order processing time. Path congestion level is calculated based on segmented path density distribution data. By mapping riders' expected paths to a path density map, the impact of congestion on the path is quantified, resulting in a congestion impact index. This index is determined based on the real-time congestion level and length of the congested road segment. Rider mobility efficiency is predicted based on rider type data and historical behavior patterns, as different rider types exhibit varying movement speeds under different road conditions. Order processing time includes three parts: pickup waiting time, delivery time, and delivery time, predicted through historical data statistics and real-time store status. These factors are combined to obtain the expected completion time and timeliness level for each allocation instruction. The timeliness level is divided into four grades: Excellent (expected early completion), Good (on time), Average (slight delay), and Poor (severe delay), forming the instruction timeliness level data.

[0049] When prioritizing the allocation instruction set in real time based on instruction timeliness level data, the sorting process first performs a coarse classification by timeliness level, with instructions with lower timeliness levels having the highest priority and requiring priority processing or reallocation. Then, within the same timeliness level, a more refined sort is performed based on order priority, customer level, delivery distance, and rider load. The priority calculation formula is: Priority_score = α1 × Timeliness Weight + β1 × Order Weight + γ1 × Customer Weight + δ1 × Distance Weight + ε1 × Load Weight, where α1, β1, γ1, δ1, and ε1 are the weight coefficients for each component, dynamically adjusted according to business strategies. The sorting process also considers real-time changing factors, such as new order additions, rider status changes, and traffic condition updates, employing an incremental sorting strategy that only re-sorts the affected components, improving sorting efficiency. Finally, a hierarchical allocation sequence is formed, arranged from highest to lowest priority, providing the execution order for subsequent matching simulations.

[0050] Step S4: Adjust order push for each rider according to the task parameter configuration instructions to obtain real-time rider task configuration data; monitor the execution status of the real-time rider task configuration data to obtain task execution status data; evaluate the load efficiency based on the task execution status data to obtain load performance evaluation results; perform feedback optimization based on the load performance evaluation results to obtain optimization parameters; apply the optimization parameters to the load balancing distribution scheme.

[0051] It should be further explained that, in the specific implementation process, the timeliness of the initial allocation instruction set is evaluated based on the segmented path density distribution data to obtain instruction timeliness level data, including: The rider's expected delivery route is decomposed into road segments, and the route is mapped to segmented road density distribution data to obtain the congestion impact index; Based on rider type data and historical behavior patterns, combined with path characteristics, rider movement efficiency is predicted to obtain an estimated rider movement speed. Based on historical data and real-time status of merchants, the order pickup waiting time is calculated, and the delivery time is estimated by combining the route length and rider movement speed to obtain the order processing time prediction value; By combining the congestion impact index, the estimated rider movement speed, and the predicted order processing time, the expected completion time of the assigned instructions is calculated to obtain timeliness prediction data. The timeliness prediction data is compared with the expected delivery time of the order, and the timeliness level of the instruction is classified according to the time difference to obtain the instruction timeliness level data.

[0052] In this embodiment, when decomposing the rider's expected delivery route into segments, the entire delivery route is first divided into multiple segment units according to natural breakpoints such as intersections, turning points, and functional area boundaries. Then, a unique identifier is assigned to each segment, and basic information such as its start and end point coordinates, length, road type, and traffic characteristics are recorded. Next, the segment is projected onto the spatiotemporal coordinate system of the segmented path density distribution data to achieve precise alignment between the path and density data. Spatial interpolation technology is used during the mapping process to handle cases where the density sampling points and the segment locations do not perfectly match, ensuring the continuity and accuracy of the mapping results. For each segment, the density value within its corresponding time period is extracted as a congestion indicator, and a congestion impact index is calculated. The congestion impact index comprehensively considers the segment density value, segment length, and segment importance (main roads have a greater impact from congestion), quantifying the potential impact of path congestion on delivery time. The calculation of the congestion impact index pays special attention to areas of sudden density changes and periods of large fluctuations, which are often high-risk points for delivery delays. When predicting rider mobility efficiency based on rider type data and historical behavior patterns, the process first extracts speed characteristics from historical rider trajectory data, including average speed, peak-hour speed, speed on various roads, and acceleration / deceleration characteristics. Then, riders are categorized and labeled into multiple types, such as high-speed, stable, and cautious riders. Next, the differences in speed performance among different rider types under various road conditions and weather conditions are analyzed. Finally, an adaptive speed prediction model is established to predict the rider's actual speed during the delivery process based on current rider type characteristics, route characteristics, and environmental conditions. The prediction model pays special attention to speed adjustments under changing road conditions and abnormal weather conditions to improve the robustness of the prediction. The prediction results form an estimate of the rider's speed, accurate to the road segment level, providing a basis for subsequent delivery time calculations.

[0053] The model calculates order pickup waiting time based on historical merchant data and real-time status using a multi-factor comprehensive prediction model. First, it extracts regularities in merchant preparation time from historical data, such as average preparation time and its fluctuation range under different time periods, different dish types, and different order volumes. Then, it obtains the merchant's current real-time status information, including current backlog of orders, kitchen busyness, and food preparation speed. Next, it incorporates special circumstances, such as new product launches, promotional activities, or equipment malfunctions that may affect food preparation speed. Finally, it calculates the expected pickup waiting time through weighted fusion. Simultaneously, based on path length and predicted rider speed, it calculates the time required for delivery; and based on delivery location type (e.g., residential areas, office buildings, schools) and historical delivery records, it estimates the final delivery time. These three time components are combined to form the predicted order processing time, comprehensively covering the entire process from pickup to delivery.

[0054] The expected completion time of delivery instructions is calculated by combining the congestion impact index, rider speed estimates, and order processing time predictions. First, the entire delivery process is divided into three main stages: pickup, delivery, and final delivery. Then, the base time for each stage is calculated: the pickup stage is based on predicted pickup waiting time, the delivery stage on path length and rider speed, and the final delivery stage on delivery location characteristics and delivery complexity. Next, a congestion risk correction factor is introduced to adjust the delivery stage time upwards based on the congestion impact index. Finally, the times of each stage are combined, and a reasonable time buffer is added to form the final expected completion time. The calculation process pays special attention to high-risk time points and critical path nodes to ensure that the prediction results reflect both normal conditions and potential anomalies. The calculation results form timely delivery prediction data, including the expected completion time and possible ranges for the completion time (considering error margins).

[0055] When comparing timeliness prediction data with the expected delivery time of orders, the time difference between the expected completion time and the expected delivery time is first calculated. A positive value indicates a possible delay, while a negative value indicates a possible early delivery. Then, based on the magnitude and direction of the time difference, instructions are classified into different timeliness levels: "Excellent" indicates service quality exceeding expectations when expected to complete more than 10 minutes ahead of schedule; "Good" indicates service quality meeting expectations when expected to complete on time or less than 10 minutes ahead of schedule; "Average" indicates service quality slightly decreasing but still acceptable when expected to complete 5-15 minutes behind schedule; and "Poor" indicates service quality significantly decreasing and requiring intervention. The timeliness level classification process considers order type differences. For high-value orders or special service orders, stricter classification standards are used to ensure that critical orders receive higher quality service guarantees. The classification results form instruction timeliness level data, directly affecting subsequent priority ranking and resource allocation, and are a key decision-making basis for intelligent order allocation.

[0056] It should be further explained that, in the specific implementation process, the process of obtaining the order push adjustments for each rider based on the task parameter configuration instructions includes: The task parameter configuration instructions are parsed and processed to obtain the task parameter set; the parameter update is pushed to each rider according to the task parameter set to obtain the rider parameter update data; the validity of the push is verified on the rider parameter update data to obtain the verification result. Based on the verification results, rider task performance is monitored to obtain task performance data; real-time status data of task performance data is collected to obtain task execution status data. Efficiency indicators are calculated based on task execution status data to obtain efficiency evaluation data; delivery quality is evaluated based on task execution status data to obtain delivery quality indicators. A comprehensive load performance evaluation is conducted based on efficiency assessment data and delivery quality indicators to obtain the load performance evaluation results. The load performance evaluation results are analyzed for potential improvement to obtain data on improvement directions; a parameter tuning strategy is generated based on the improvement direction data to obtain tuning parameters; and the tuning parameters are fed back to the load balancing distribution scheme.

[0057] In this embodiment, when parsing and processing task parameter configuration instructions, a layered parsing framework is used to handle different types of configuration instructions. For standardized JSON or XML format configuration instructions, a syntax parser is first used to extract the instruction structure and parameter values. Then, a parameter mapping table is used to convert general parameters into specific parameters for rider terminals, ultimately forming a complete set of task parameters, including delivery rules, priority settings, route configuration, and quality control. The parameters are then distributed to each rider terminal via mobile application API, push notifications, or SMS interfaces to complete the parameter configuration update. During the update process, the application status of each parameter and rider feedback information are recorded to form rider parameter update data.

[0058] When validating the configuration of rider parameter update data, the following steps are performed: First, parameter reception is verified by checking the configuration confirmation receipt on the rider's terminal to confirm whether the parameters are correctly received and applied. Next, functional verification is performed by sending test orders or simulating delivery tasks to check if the rider's delivery function is normal. Finally, performance verification is performed to check whether the delivery metrics after parameter adjustments meet expectations. Verification metrics include parameter reception rate, functional compliance rate, and performance compliance rate. The parameter reception rate is calculated as the ratio of the number of riders who confirmed receiving the data to the total number of riders. The functional compliance rate is calculated as the ratio of the number of riders with normal functionality to the number of riders who confirmed receiving the data. The performance compliance rate is calculated as the ratio of the number of riders with satisfactory performance to the number of riders with normal functionality. The combined results of these three aspects form a complete configuration verification result.

[0059] When monitoring rider task performance based on configuration verification results, a layered monitoring framework is adopted, covering three levels: individual performance, regional collaboration, and the overall system. Individual performance monitoring mainly focuses on the work indicators of individual riders, such as delivery volume, average delivery time, customer satisfaction, and online time. Regional collaboration monitoring focuses on the cooperation effect among riders within a region, such as the balance of order allocation, the integrity of regional coverage, and the coordination of emergency response. Overall system monitoring focuses on performance indicators related to the overall delivery network, such as overall delivery efficiency, network load distribution, and resource utilization. Monitoring data collection adopts a hierarchical sampling strategy, with high-frequency sampling for key indicators, medium-frequency sampling for general indicators, and low-frequency sampling for auxiliary indicators. At the same time, an anomaly trigger mechanism is set up to increase the sampling frequency and activate alarms when indicators exceed the preset range.

[0060] Real-time status analysis is performed on the collected task performance data. The sliding window statistical method is used to calculate the statistical characteristics of various performance indicators, including mean, median, standard deviation, and trend, to form a task performance feature vector. At the same time, a time-series anomaly detection algorithm is used to analyze performance change patterns and identify performance anomaly events and trend changes. Through these analyses, comprehensive task execution status data is obtained.

[0061] When calculating efficiency indicators based on task execution status data, differentiated calculation models are used for efficiency in different dimensions; delivery efficiency is represented by the number of effective delivery orders completed per unit time; route efficiency is represented by the ratio of actual delivery route to optimal route; time efficiency is represented by the ratio of expected delivery time to actual delivery time; these indicators are combined and weighted to obtain the overall efficiency evaluation index.

[0062] When assessing delivery quality based on task execution status data, a multi-dimensional customer experience indicator system is employed, encompassing three dimensions: timeliness, service, and satisfaction. Timeliness is assessed by evaluating delivery accuracy, calculated as the ratio of on-time delivered orders to the total number of orders. Service quality is assessed by evaluating the quality of the delivery process, calculated as the ratio of the weighted sum of perfect, good, and average deliveries to the total number of orders. Satisfaction is assessed by evaluating customer feedback, calculated as the ratio of the average customer rating to the highest rating. These three dimensions are then combined using a weighted summation formula to calculate the overall delivery quality indicator.

[0063] When conducting a comprehensive load performance evaluation based on efficiency assessment data and delivery quality indicators, a multi-objective balancing strategy is adopted to seek the optimal balance between efficiency and quality; the performance evaluation function is: Performance_score = α² × Efficiency_total + β² × Quality_index - γ² × Cost_factor; where α², β², and γ² are balancing weights, Efficiency_total represents the overall efficiency evaluation index, Quality_index represents the delivery quality index, and Cost_factor is the cost impact factor. This evaluation method allows the system to comprehensively consider efficiency and quality, avoiding the negative impacts that may result from optimizing a single index, and obtaining a comprehensive load performance evaluation result.

[0064] When analyzing the potential for improvement of load performance evaluation results, two methods are used: target gap analysis and resource balance analysis. Target gap analysis calculates the gap between the current performance and the target performance. When the gap exceeds a preset threshold, the optimization process is triggered. Resource balance analysis detects the load balance of riders in each area and calculates the load balance degree. When the balance degree is lower than the threshold, it identifies an imbalance in load distribution that needs to be adjusted. Based on these analysis results, the system identifies the adjustment direction of specific parameters and forms optimization direction data.

[0065] When generating parameter tuning strategies based on optimization direction data, a combination of incremental tuning and machine learning is employed. Incremental tuning uses a gradient descent strategy to make small adjustments to the parameters and observe the effects. Machine learning uses historical tuning data to train a predictive model and predicts the optimal parameter settings through feature matching. Combining these two methods, the system generates the final tuning parameters, including load weights, priority adjustments, and path optimizations, and feeds them back to the load balancing allocation scheme, completing the final stage of closed-loop optimization.

[0066] This invention constructs a route optimization model by acquiring real-time rider location data and traffic dynamic data within the delivery area. This model accurately predicts the spatiotemporal distribution characteristics of the delivery network load, providing a scientific basis for order allocation and improving the foresight and accuracy of allocation strategies. By dividing the delivery area into rider clusters, differentiated analysis can be performed on the load characteristics of different areas, resulting in a more refined load distribution heatmap. This provides spatial accuracy for order allocation and improves the efficiency of rider resource utilization. By prioritizing the allocation instruction set, limited rider resources can be rationally allocated in the event of order competition, resolving allocation conflicts and ensuring the delivery service quality in key areas. Furthermore, the task parameter configuration instructions generated based on the optimal order allocation strategy can accurately guide riders in each area to adjust their task configurations, improving delivery accuracy and execution efficiency.

[0067] Example 2 Please see Figure 2 As shown, parts not described in detail in this embodiment are described in Embodiment 1. A smart order allocation system considering rider load balancing is provided, including: The load distribution module is used to acquire real-time location data of riders, historical order generation data, and traffic dynamic data within the delivery area; construct a rider route optimization model based on the real-time location data and traffic dynamic data to obtain route optimization prediction results; construct an order demand prediction model based on historical order generation data and route optimization prediction results to obtain peak demand prediction data; and generate a preliminary distribution strategy based on the peak demand prediction data to obtain a load balancing distribution scheme. The indicator acquisition module is used to divide delivery areas into rider clusters to obtain rider cluster data; perform load spatiotemporal analysis based on rider cluster data and route optimization prediction results to obtain a load distribution heatmap; and calculate the load balance index from the load distribution heatmap. The order allocation module generates allocation instructions for orders to be allocated based on load balancing metrics and load balancing allocation schemes, resulting in an allocation instruction set; it performs real-time priority sorting on the allocation instruction set to obtain a hierarchical allocation sequence; it performs matching and optimization based on the hierarchical allocation sequence and load balancing metrics to obtain the optimal order allocation strategy; and it generates rider task parameters based on the optimal order allocation strategy to obtain task parameter configuration instructions. The feedback optimization module is used to adjust order pushes for each rider according to the task parameter configuration instructions, and obtain real-time rider task configuration data; monitor the execution status of the real-time rider task configuration data to obtain task execution status data; evaluate the load efficiency based on the task execution status data to obtain load performance evaluation results; perform feedback optimization based on the load performance evaluation results to obtain optimization parameters; and apply the optimization parameters to the load balancing distribution scheme.

[0068] The modules are connected via wired and / or wireless means to enable data transmission between them.

[0069] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0070] All formulas in this manual are dimensionless and calculated numerically. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0071] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A smart order allocation method considering rider load balancing, characterized in that, include: Step S1: Obtain real-time location data of riders, historical order generation data, and traffic dynamic data within the delivery area; A rider route optimization model is constructed based on rider real-time location data and traffic dynamic data to obtain route optimization prediction results; an order demand prediction model is constructed based on order generation history data and route optimization prediction results to obtain peak demand prediction data. A preliminary allocation strategy is generated based on peak demand forecast data, resulting in a load balancing allocation scheme. Step S2: Divide the delivery area into rider clusters to obtain rider cluster data; perform load spatiotemporal analysis based on rider cluster data and route optimization prediction results to obtain a load distribution heatmap; calculate the load balance index based on the load distribution heatmap. Step S3: Generate allocation instructions for the orders to be allocated based on load balancing metrics and load balancing allocation scheme to obtain an allocation instruction set; sort the allocation instruction set in real time to obtain a hierarchical allocation sequence; perform matching and optimization based on the hierarchical allocation sequence and load balancing metrics to obtain the optimal order allocation strategy; Rider task parameters are generated based on the optimal order allocation strategy, and task parameter configuration instructions are obtained. Step S4: Adjust the order push for each rider according to the task parameter configuration instructions to obtain the real-time configuration data of the rider's task; monitor the execution status of the real-time configuration data of the rider's task to obtain the task execution status data; The load efficiency is evaluated based on the task execution status data to obtain the load performance evaluation results; the load performance evaluation results are used for feedback tuning to obtain tuning parameters; and the tuning parameters are applied to the load balancing distribution scheme.

2. The intelligent order allocation method considering rider load balancing according to claim 1, characterized in that, The process of obtaining a load balancing allocation scheme includes: By collecting riders' historical trajectory data and performing spatiotemporal pattern analysis, a rider behavior feature database is obtained; and by collecting real-time road condition data and weather impact data, environmental interference factor data is obtained. Obtain delivery network area division data; construct the path network based on the delivery network area division data to obtain path network topology data; collect order generation time series data through the order management center to obtain historical order data; A rider route optimization model is constructed based on rider behavior feature database and environmental interference factor data to obtain route behavior pattern data; dynamic feature extraction is performed on the route behavior pattern data to obtain dynamic route features. Spatial path prediction is performed based on path behavior pattern data and path dynamic characteristics to obtain a path coverage probability map; the path coverage probability map is then divided into time periods to obtain segmented path density distribution data. The segmented path density distribution data is mapped to the path network topology data to obtain order load prediction data; based on historical order data and order load prediction data, an order demand prediction model is constructed to obtain demand fluctuation prediction data. Peak detection and anomaly analysis are performed on demand fluctuation forecast data to obtain a demand surge index; Based on the path network topology data, we analyze the distribution of rider resources to obtain a rider resource capability map; we then evaluate the available capacity of the rider resource capability map to obtain a rider availability matrix. A load balancing distribution scheme is generated based on the demand surge index and rider availability matrix.

3. The intelligent order allocation method considering rider load balancing according to claim 2, characterized in that, The process of obtaining load balancing metrics includes: The delivery area is divided into geographical grids, and density cluster analysis is performed based on the rider's real-time location data to obtain rider cluster data; Based on rider cluster data and route optimization prediction results, spatiotemporal load vector decomposition is performed to extract the task backlog trend within the cluster. Perform heatmap visualization rendering of task backlog trends within the cluster to generate a load distribution heatmap. Entropy values ​​and variance balance measures are performed on the load distribution heatmap to obtain load balance indices, including inter-cluster load deviation rate and overall spatiotemporal balance score.

4. The intelligent order allocation method considering rider load balancing according to claim 3, characterized in that, The process of obtaining task parameter configuration instructions includes: Obtain the orders to be assigned, perform status analysis and priority marking on them to obtain the order status matrix; calculate the order-rider matching degree based on load balancing metrics and load balancing allocation scheme to obtain the matching fitness matrix; The allocation decision is generated based on the order status matrix and the matching fitness matrix to obtain the initial allocation instruction set; The timeliness of the initial allocation instruction set is evaluated based on the segmented path density distribution data to obtain instruction timeliness level data; Based on instruction timeliness level data, the allocation instruction set is sorted by real-time priority to obtain a hierarchical allocation sequence; Order matching simulation is performed based on hierarchical allocation sequence and load balancing indicators to obtain matching prediction results; then conflict detection and resolution are performed on the matching prediction results to obtain conflict resolution strategies. The optimal order allocation strategy is generated based on the matching prediction results and conflict resolution strategy; the task parameter configuration instructions are then generated based on the optimal order allocation strategy.

5. The intelligent order allocation method considering rider load balancing according to claim 4, characterized in that, The process of adjusting order pushes for each rider based on task parameter configuration instructions includes: The task parameter configuration instructions are parsed and processed to obtain the task parameter set; the parameter update is pushed to each rider according to the task parameter set to obtain the rider parameter update data; the validity of the push is verified on the rider parameter update data to obtain the verification result. Based on the verification results, rider task performance is monitored to obtain task performance data; real-time status data of task performance data is collected to obtain task execution status data. Efficiency indicators are calculated based on task execution status data to obtain efficiency evaluation data; delivery quality is evaluated based on task execution status data to obtain delivery quality indicators. A comprehensive load performance evaluation is conducted based on efficiency assessment data and delivery quality indicators to obtain the load performance evaluation results. The load performance evaluation results are analyzed for potential improvement to obtain data on improvement directions; a parameter tuning strategy is generated based on the improvement direction data to obtain tuning parameters; and the tuning parameters are fed back to the load balancing distribution scheme.

6. The intelligent order allocation method considering rider load balancing according to claim 2, characterized in that, The rider route optimization model, constructed based on the rider behavior feature database and environmental interference factor data, yields route behavior pattern data, including: Rider type classification analysis was performed on the rider behavior feature database to obtain rider type data; and interference sensitivity was extracted based on environmental interference factor data to obtain the path interference index. A reinforcement learning path optimization network is constructed based on rider type data to obtain a basic path optimization model. The path interference index is then used to adaptively adjust the parameters of the basic path optimization model to obtain a rider path optimization model, and path behavior pattern data is generated based on it.

7. The intelligent order allocation method considering rider load balancing according to claim 2, characterized in that, The path network topology data is represented by a weighted directed graph structure. The network nodes in the graph structure include delivery stations, merchants, customer addresses, and transportation hubs. The connections between nodes include delivery routes and logical associations. The connection weights between nodes are composed of several indicators, including distance, estimated delivery time, and road grade indicators.

8. The intelligent order allocation method considering rider load balancing according to claim 5, characterized in that, The process of obtaining load performance evaluation results includes: The efficiency evaluation data is normalized using multi-dimensional indicators to obtain normalized efficiency indicators. User feedback is mapped to delivery quality indicators to obtain satisfaction assessment data; A load performance evaluation framework is constructed, and normalized efficiency indicators and satisfaction evaluation data are dynamically weighted and integrated to obtain a comprehensive performance index. Load performance evaluation results are then generated based on the comprehensive performance index.

9. The intelligent order allocation method considering rider load balancing according to claim 4, characterized in that, The timeliness of the initial allocation instruction set is evaluated based on the segmented path density distribution data to obtain instruction timeliness level data, including: The rider's expected delivery route is decomposed into road segments, and the route is mapped to segmented road density distribution data to obtain the congestion impact index; Based on rider type data and historical behavior patterns, combined with path characteristics, rider movement efficiency is predicted to obtain an estimated rider movement speed. Based on historical data and real-time status of merchants, the order pickup waiting time is calculated, and the delivery time is estimated by combining the route length and rider movement speed to obtain the order processing time prediction value; By combining the congestion impact index, the estimated rider movement speed, and the predicted order processing time, the expected completion time of the assigned instructions is calculated to obtain timeliness prediction data. The timeliness prediction data is compared with the expected delivery time of the order, and the timeliness level of the instruction is classified according to the time difference to obtain the instruction timeliness level data.

10. A smart order allocation system considering rider load balancing, based on the smart order allocation method considering rider load balancing described in claims 1 to 9, characterized in that, include: Load distribution module; Used to obtain real-time location data of riders, historical order generation data, and traffic dynamic data within the delivery area; A rider route optimization model is constructed based on rider real-time location data and traffic dynamic data to obtain route optimization prediction results; an order demand prediction model is constructed based on order generation history data and route optimization prediction results to obtain peak demand prediction data. A preliminary allocation strategy is generated based on peak demand forecast data, resulting in a load balancing allocation scheme. The indicator acquisition module is used to divide delivery areas into rider clusters and obtain rider cluster data; based on the rider cluster data and route optimization prediction results, load spatiotemporal analysis is performed to obtain a load distribution heatmap. The load distribution heatmap is used to calculate the load balance index. Order allocation module; Based on load balancing metrics and load balancing allocation schemes, allocation instructions are generated for the orders to be allocated, resulting in an allocation instruction set; the allocation instruction set is sorted by priority in real time to obtain a hierarchical allocation sequence; and the hierarchical allocation sequence and load balancing metrics are matched and optimized to obtain the optimal order allocation strategy. Rider task parameters are generated based on the optimal order allocation strategy, and task parameter configuration instructions are obtained. Feedback optimization module; This is used to adjust order pushes for each rider based on task parameter configuration instructions, and obtain real-time rider task configuration data; and to monitor the execution status of the real-time rider task configuration data, and obtain task execution status data. The load efficiency is evaluated based on the task execution status data to obtain the load performance evaluation results; the load performance evaluation results are used for feedback tuning to obtain tuning parameters; and the tuning parameters are applied to the load balancing distribution scheme.

Images