Intelligent distribution capacity real-time allocation system for multi-service scenario adaptive switching

Abstract

The application relates to the field of artificial intelligence and discloses an intelligent distribution capacity real-time allocation system for adaptive switching of multiple service scenarios. The system adopts a three-layer collaborative distributed architecture of cloud prediction, edge decision and terminal execution. The system fuses multi-source heterogeneous data such as orders, capacity, road conditions and environment through a space-time gridding method, and realizes multi-dimensional prediction of order demand and road condition state by using a space-time graph convolution network and a hybrid neural network. The core innovation lies in introducing an adaptive switching mechanism based on fuzzy logic, dynamically generating scheduling weights according to real-time service scenarios, guiding a reinforcement learning strategy network to perform real-time assignment of capacity and predictive prescheduling actions. In combination with model compression technology, the system realizes millisecond-level response at the edge node, and performs automatic recovery of the assignment scheme through an abnormality monitoring algorithm. The application effectively shortens the distribution response time, reduces the capacity empty running rate, and improves the scheduling efficiency and operation stability of the system in a complex dynamic environment.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to an intelligent real-time distribution system for adaptive switching across multiple business scenarios. Background Technology

[0002] With the rapid development of on-demand delivery services, scenarios such as same-city delivery, food delivery, and fresh food e-commerce have placed higher demands on delivery timeliness and dynamic order response. However, existing vehicle routing methods mostly adopt static planning models, uniformly planning routes before delivery begins, resulting in the system's inability to respond in real time to new orders generated during the delivery process. Due to the lack of effective predictive capabilities for the spatiotemporal distribution of future orders and the evolution of road conditions, the delivery system is often in a state of reactive adjustment after the fact. New orders typically require a long average waiting time before being assigned, and delivery personnel are prone to entering sections of road that are about to become congested, resulting in a high level of empty-running capacity.

[0003] At the scheduling strategy execution level, some dynamic insertion methods rely primarily on manually designed local optimization rules or greedy strategies. These methods typically set fixed scheduling parameters and decision thresholds, making it difficult to adaptively adjust to real-time changes in business scenarios. When faced with complex scenarios such as order surges during lunchtime peak hours, capacity redundancy during off-peak hours, or reduced traffic efficiency due to severe weather, fixed-weight scheduling logic cannot achieve a dynamic balance between timeliness, cost, and customer satisfaction. Furthermore, traditional machine learning methods, due to their limited expressive power, struggle to capture the complex and non-linear temporal dependencies in delivery scenarios, resulting in insufficient accuracy in delivery time prediction and limiting the accuracy of assignment decisions.

[0004] From a system architecture perspective, existing delivery scheduling largely relies on a centralized cloud computing model, with all order processing and decision-making instructions issued by a central server. When business expansion leads to a sharp increase in order volume, the centralized architecture often faces significant computational pressure and communication latency, making it difficult to guarantee millisecond-level response times for assignment decisions. Simultaneously, centralized systems lack effective edge-side fault tolerance mechanisms; if cloud nodes experience response bottlenecks or network fluctuations, it will directly impact the overall regional scheduling efficiency. Furthermore, due to the lack of efficient model compression technology, complex scheduling models are difficult to deploy on computing nodes closer to the terminals, hindering the implementation of distributed scheduling solutions. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent real-time delivery capacity allocation system that adaptively switches between multiple business scenarios. This system solves the problems of existing delivery scheduling schemes, which suffer from slow response in dynamic environments, high empty-running rates, and inability to automatically switch scheduling focus for different business loads due to the lack of predictive capabilities, fixed assignment parameters, and large computational delays caused by centralized architecture.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an intelligent real-time delivery capacity allocation system with adaptive switching across multiple business scenarios, comprising: The cloud-based intelligent layer 100 is deployed on a central server cluster and is used to perform offline model training, global resource optimization, and long-term prediction tasks.

[0007] Edge computing layer 200 is deployed in a distributed manner in urban grid areas to perform regional real-time assignment decisions, local route optimization, and delivery anomaly monitoring.

[0008] The terminal execution layer 300 runs on mobile terminal devices and is used to report the delivery personnel's geographical coordinates and task load status in real time, and to receive assignment decision instructions.

[0009] The core principle of this invention lies in constructing an adaptive scheduling system based on spatiotemporal feature perception and fuzzy logic guidance. The system uses a multimodal data acquisition and fusion system 55 to spatially align orders, transportation capacity, road conditions, and environmental variables, and divides them into geographic grid units according to preset specifications to construct a high-dimensional feature vector containing order density, average road conditions, and point of interest distribution.

[0010] In the perception dimension, the deep learning prediction engine 57 models the spatial adjacency and functional similarity relationships between grids through a spatiotemporal graph convolutional network to predict order demand in future time periods; at the same time, it uses a hybrid architecture of recurrent neural networks and convolutional neural networks to extract the temporal features of road segments to predict the congestion status of road segments.

[0011] In terms of decision-making switching, this invention introduces a business scenario identification mechanism based on a fuzzy logic controller. This mechanism takes real-time order density, average system waiting time, and delivery personnel load rate as inputs, and infers a dynamic weight set for the current scenario through a fuzzy rule base. The generated timeliness weight, cost weight, and satisfaction weight are fed back into the reward function of the reinforcement learning policy network in real time, enabling the assignment decision to switch to a timeliness mode during order surges and a cost optimization mode during periods of capacity redundancy.

[0012] In terms of allocation and execution, the system utilizes a predictive delivery pre-scheduling mechanism to solve for the offset path with the lowest movement cost based on the prediction results and optimization algorithms. This guides idle delivery capacity to converge towards high-demand areas before orders are actually generated. To ensure deployment performance, the system uses knowledge distillation technology to compress the cloud model into a lightweight model and deploy it to edge computing layer nodes, enabling high-frequency regional assignment reasoning.

[0013] In addition, the system analyzes delivery deviations in real time through the anomaly monitoring algorithm 372, uses unsupervised algorithms to identify delivery delays or route deviations, and triggers an automatic recovery mechanism to perform order transfers or route replanning.

[0014] This invention provides an intelligent real-time delivery capacity allocation system with adaptive switching across multiple business scenarios. It offers the following advantages: 1. This invention achieves accurate prediction of order demand and road conditions through spatiotemporal graph convolutional networks and hybrid neural networks, and executes a two-stage pre-scheduling mechanism based on the prediction results. This forward-looking capacity allocation method enables delivery personnel to converge to high-demand areas in advance, effectively shortening order response time and reducing the empty-running rate of delivery personnel.

[0015] 2. This invention introduces an adaptive switching mechanism based on fuzzy logic, which can dynamically adjust the optimization weights of timeliness, cost, and satisfaction according to indicators such as real-time order density and load rate. This technical feature ensures that the scheduling strategy can automatically switch between various business scenarios such as peak periods, off-peak periods, and severe weather, achieving dynamic balancing and optimization of multi-objective delivery tasks.

[0016] 3. This invention employs an edge-cloud collaborative distributed architecture, combined with model compression technologies such as knowledge distillation, to deploy computing tasks to regional edge nodes for execution. This hierarchical assignment model significantly reduces the response latency of system decisions, improving the throughput of high-concurrency orders while enhancing the fault tolerance and operational stability of the scheduling system in the event of node failures or network anomalies. Attached Figure Description

[0017] Figure 1 This is a structural diagram of the present invention.

[0018] Among them, 55. Multimodal data acquisition and fusion system; 57. Deep learning prediction engine; 58. Order demand prediction model; 59. Road condition prediction model; 60. Delivery time prediction model; 61. Adaptive scheduling strategy; 66. Predictive deliveryman pre-scheduling mechanism; 68. Edge-cloud collaborative scheduling architecture; 72. Business scenario adaptive switching mechanism; 100. Cloud intelligence layer; 110. Model training module; 120. Global optimization module; 130. Long-term prediction module; 200. Edge computing layer; 210. Regional scheduling module; 220. Fast decision-making module; 230. Local optimization module; 240. Anomaly monitoring module; 300. Terminal execution layer; 310. Navigation module; 320. Status reporting module; 330. Feedback interaction module; 354. Fuzzy logic controller; 360. Fuzzy rule base; 372. Anomaly detection algorithm; 376. Rule engine. Detailed Implementation

[0019] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example: Please see the appendix Figure 1 This invention provides an intelligent real-time delivery capacity allocation system with adaptive switching across multiple business scenarios. The system employs a three-layer collaborative distributed architecture of cloud prediction, edge decision-making, and terminal execution. The system includes a cloud intelligence layer 100, an edge computing layer 200, and a terminal execution layer 300.

[0021] The cloud-based intelligent layer 100 is deployed on a high-performance server cluster, integrating a model training module 110, a global optimization module 120, and a long-term prediction module 130. The model training module 110 utilizes stored historical multimodal data to perform offline training and parameter iteration on the deep learning prediction model and reinforcement learning policy network. The global optimization module 120 is responsible for the non-real-time allocation of resources and macro-level strategy guidance across the entire city. The long-term prediction module 130 performs large-scale order demand and traffic trend analysis on a minute-by-minute basis.

[0022] The edge computing layer 200 consists of multiple edge computing nodes distributed across different geographical grid areas within the city. The edge computing layer 200 includes a regional scheduling module 210, a fast decision-making module 220, a local optimization module 230, and an anomaly monitoring module 240. The regional scheduling module 210 directly responds to real-time delivery requests within its assigned region. The fast decision-making module 220 calls a compressed, lightweight model to perform millisecond-level capacity allocation calculations. The local optimization module 230 performs real-time dynamic corrections to route planning within the region. The anomaly monitoring module 240 is used to identify deviation events in the delivery process and trigger local recovery logic.

[0023] The terminal execution layer 300 runs on the mobile terminal device and mainly includes a navigation module 310, a status reporting module 320, and a feedback interaction module 330. The navigation module 310 receives scheduling instructions and displays the optimal route. The status reporting module 320 collects the delivery person's geographical coordinates and task load status in real time. The feedback interaction module 330 is used to receive order confirmation signals and user error feedback.

[0024] The intelligent delivery capacity real-time allocation system provided by this invention performs the following macro-level workflow during operation: The S101 multimodal data acquisition and fusion system acquires multi-source data in real time, including order flow, capacity distribution, road network index, and environmental variables. The system uses a spatiotemporal gridding method to divide the target geographic space into feature units of preset sizes, and constructs a feature vector for each unit containing order density, average road condition, and POI density. Regarding GPS coordinate deviation calibration and invalid data filtering, those skilled in the art can implement it using conventional filtering algorithms and coordinate transformation formulas. The specific process is well-known in the field and will not be elaborated here.

[0025] S102, Deep Learning Prediction Engine 57 performs multi-dimensional state prediction. Order Demand Prediction Model 58 predicts the probability distribution of order demand within the next 15 to 30 minutes based on a spatiotemporal graph convolutional network. Traffic prediction model 59 uses a hybrid network to predict the congestion index of road segments within the next 10 to 20 minutes. The delivery time prediction model 60 calculates the estimated duration of a single delivery task using a deep neural network, based on order, capacity, and road network characteristics.

[0026] S103, the business scenario adaptive switching mechanism 72 is activated, adjusting the scheduling optimization weights based on real-time environmental indicators through the fuzzy logic controller 354. Controller input variables include order density. Average system wait time and capacity load factor Through fuzzy inference and centroid defuzzification, the system outputs a dynamic weight set for different scenarios. :

[0027] In the formula, Represents timeliness weight. Represents cost weight, These three factors represent satisfaction weights, and their sum satisfies a normalization constraint. The system switches scheduling strategies in real time based on this weight set, for example, automatically increasing [the load] during peak periods. The percentage.

[0028] S104, Adaptive Scheduling Strategy 61 utilizes a reinforcement learning network to perform real-time capacity allocation. The scheduling engine encodes the features of the current orders to be assigned, the status of candidate delivery personnel, and prediction information into a high-dimensional state vector. And input it into a deep Q-network (DQN). The system is based on the action space. Output optimal allocation action And combined with the dynamic weights generated in step S103 Calculate instant reward value :

[0029] In the formula, , , These are the feedback rewards for the corresponding dimensions. As a load balancing reward, To constrain violations and impose penalties, These are preset constant coefficients.

[0030] S105, the predictive delivery pre-scheduling mechanism 66 executes capacity convergence in advance based on the order demand forecast results. When the ratio of predicted demand to real-time capacity in a specific area exceeds a preset threshold, the system calculates the optimal scheduling offset instruction through a two-stage optimization model, guiding idle capacity to move in advance to the expected order surge point.

[0031] S106, the edge-cloud collaborative scheduling architecture 68 ensures efficient command execution. Long-term optimization parameters are distributed from the cloud, and edge nodes 200 utilize local inference models to achieve rapid decision-making with a response time of less than 100 milliseconds. Upon detecting delivery delays or route anomalies, the anomaly detection algorithm 372 automatically triggers emergency scheduling logic, executing order transfers or route replanning.

[0032] S107, the terminal execution layer 300 provides feedback on fulfillment data through the instruction interaction module after the task is completed. The system transmits the actual delivery time and trajectory data back to the cloud and stores it in the priority experience replay pool for online fine-tuning and continuous learning of the strategy network.

[0033] Step S101 in the overall assignment and allocation process of the aforementioned system, namely the multimodal data feature extraction and fusion process, is further refined into sub-steps S201 to S204.

[0034] S201, the system uses the multimodal data acquisition and fusion system 55 to perceive and acquire multi-source heterogeneous data. The acquired data set covers order data streams. Delivery personnel data stream Traffic data stream and environmental data stream Order data stream This includes real-time order identifiers, pickup address coordinates, delivery address coordinates, order timestamp, expected delivery time, order weight, product type, and priority. Delivery driver data stream. This includes real-time latitude and longitude coordinates obtained using GPS or BeiDou positioning components, the task status reported by the terminal (idle, delivering, resting), remaining load capacity, current battery level, and online duration. Traffic data stream. Includes road congestion index Estimated travel time and information on potential road incidents are included, with the update frequency for main roads configured to be every 30 seconds. Environmental data stream. It includes meteorological monitoring data (including features such as temperature, rainfall, and visibility), POI distribution data within a specific grid (such as the number of restaurants, supermarkets, and residential areas), and attribute information such as holidays and traffic control.

[0035] S202, the multimodal data acquisition and fusion system 55 performs urban spatial gridding and coordinate mapping. The system divides the target scheduling geographic area into areas with a size of geographic grid unit For any physical coordinates collected Mapped to the corresponding grid index using the spatial projection formula. This achieves spatial alignment of heterogeneous data at a unified grid scale. Basic preprocessing such as coordinate calibration and outlier filtering can be implemented using conventional filtering algorithms and coordinate transformation matrices, and the specific algorithms are well-known in the field and will not be elaborated upon here.

[0036] S203, System Construction of High-Dimensional Spatiotemporal Feature Vectors This enables feature representation of multimodal information. Within each preset sampling time step (updated every 1 minute), the system performs feature representation on each grid cell. Calculate and generate the corresponding feature components, and the fused feature vector. Represented as:

[0037] The meanings of the symbols in the formula are defined as follows: Representing the grid Formed by fusion at the present moment 3D eigenvectors. This represents the real-time order density within the grid at the current moment. This represents the current effective delivery personnel density within the grid. This represents the predicted average congestion index for the road segment within that grid. This represents the distribution density of points of interest such as restaurants and supermarkets within the grid. This represents the average number of orders for that grid during the same historical time period (e.g., the same minute intervals over the past 7 days). This represents the normalized environmental meteorological feature vector. It represents the time encoding characteristics, using one-hot encoding to represent the current hour, weekday attribute, and whether it is a holiday. The geographic coordinate code representing the center point of the grid.

[0038] S204, the system performs standardized processing and tensor serialization storage of feature data. Data acquisition and fusion system 55 targets... The continuous numerical components in the data undergo min-max normalization, mapping their numerical range to... The interval is defined to eliminate the influence of different physical dimensions on the gradient of the subsequent deep learning prediction model training. The standardized spatiotemporal feature vectors are serialized and stored in memory in chronological order to form a four-dimensional feature tensor structure (the dimension includes grid coordinates). Grid coordinates Time step and feature dimensions This directly supports the inference input requirements of the order demand prediction model 58, road condition prediction model 59, and delivery time prediction model 60 in the subsequent step S102.

[0039] The feature vectors determined by the above sub-steps S201 to S204 This ensures that the system can fully perceive real-time dynamic changes in multiple business scenarios during subsequent processing, and provides a complete data foundation for adaptive weight switching and reinforcement learning assignment decisions.

[0040] For step S103 in the aforementioned overall workflow, namely the business scenario adaptive switching mechanism, the technical solution is specifically implemented through sub-steps S301 to S305.

[0041] S301, the system performs a real-time extraction process of business scenario feature indicators. The business scenario adaptive switching mechanism 72 extracts features from the feature vector. Key evaluation indicators are extracted and used as input variables for the fuzzy logic controller 354. The input variables include: Order density The number of orders generated per unit area per unit time reflects the intensity of market demand.

[0042] Average waiting time The average time from when an order is generated to when it is accepted by a delivery person reflects the lag in the current delivery capacity response.

[0043] Capacity load factor The ratio of the time online delivery riders spend performing tasks to their total online time reflects the saturation state of transportation capacity resources.

[0044] Weather condition coefficient The discrete values ​​obtained by mapping meteorological monitoring data represent the degree of interference of the external environment on delivery efficiency.

[0045] S302, the system performs fuzzification processing on the input indicators. The fuzzy logic controller 354 configures a corresponding membership function for each input variable, converting precise numerical indicators into fuzzy sets.

[0046] For order density Its fuzzy set is defined as { The mapping is performed using a trapezoidal membership function.

[0047] For average waiting time Its fuzzy set is defined as { The mapping is performed using a triangular membership function.

[0048] For capacity load factor Its fuzzy set is defined as { This reflects whether the transport capacity has redundancy.

[0049] S303, the system performs fuzzy inference based on a preset fuzzy rule base 360. The system uses the Mamdani inference method, triggering corresponding scheduling center adjustment rules based on the membership degree of the input variables.

[0050] Typical rule definitions include: when the system identifies order density "Extremely high" and average waiting time When the timeframe is "long," it is determined that the current scenario is at an extremely high peak, and the rule output indicates an increase in the weight of timeliness. And suppress cost weight When order density "Low" and capacity load factor When the value is "low", it is determined that the current scenario is off-peak, and the rule output points to increasing the cost weight. To optimize delivery personnel's routes. (Weather condition coefficient) When the condition is deemed "bad," it is determined to be an abnormal environment, and the rule output will increase the weight of satisfaction. This allows for a relaxation of time constraints in scheduling algorithms, focusing instead on safely completing tasks.

[0051] S304, The system performs defuzzification to generate an accurate set of scheduling weights. The system uses the centroid method to calculate the precise value of each weight component.

[0052] In the formula, This represents the deblurred weight output value. For the first The membership value of the activated rule. This corresponds to the output fuzzy set center value. After calculation, the system normalizes and corrects the weights to ensure timely weighting. Cost weight and satisfaction weight The algebraic sum of these is always equal to 1. The generated weight set The reward function is fed back in real time to the reinforcement learning scheduling strategy. This can change the optimization preferences of the decision network in the current scenario.

[0053] S305, the system performs smoothing and dynamic updates for weight switching. To avoid frequent changes in scheduling strategies due to instantaneous data fluctuations, the system uses an exponential moving average method to smooth the weights over time.

[0054] In the formula, The current execution weight after smoothing. The weights were applied to the previous period. The target weights output in step S304 The preset smoothing coefficient is used. The system performs a periodic weight update every 5 minutes, or automatically triggers an instant update when the order density change rate exceeds the preset threshold, ensuring that the focus of capacity allocation can adaptively switch as business scenarios evolve.

[0055] Steps S301 to S305 above, through quantitative identification and logical reasoning of physical environment characteristics, realize the transformation of system scheduling logic from "post-event passive adjustment" to "scenario-aware adaptive switching". In this way, the present invention can provide the most suitable capacity allocation scheme for various complex business scenarios such as lunch peak, nighttime off-peak, or sudden severe weather.

[0056] For the basic computational logic of fuzzy set theory and membership function, those skilled in the art can use well-known controller parameter tuning methods to implement it. The mathematical implementation process is a well-known technology in this field and will not be elaborated here.

[0057] Step S102 in the aforementioned overall assignment and allocation process, namely the process of the deep learning prediction engine 57 performing multi-dimensional state prediction, is further refined into sub-steps S401 to S403.

[0058] S401, the system performs forward-looking prediction of grid-level order density through order demand prediction model 58. This model uses a spatiotemporal graph convolutional network (ST-GCN) combined with a multi-head attention mechanism to model the spatiotemporal correlation characteristics of the urban grid.

[0059] During the graph structure construction process, the system defines the graph. , where nodes Represents a geographic grid cell, edge It includes spatially adjacent edges and functionally similar edges. Spatially adjacent edges are defined as edges whose Euclidean distance is less than a preset threshold (e.g., ...). Connections between grids; functionally similar edges are defined as those where the cosine similarity of the POI distribution is greater than a preset threshold (e.g., ...). (Connections between grids)

[0060] In the spatial feature aggregation stage, the first Layer node feature updates follow the formula below:

[0061] in, Representing the The node feature matrix of the layer contains Each node Each time step and Dimensional features; Represents the normalized adjacency matrix; Represents the learnable weight matrix; This represents the activation function.

[0062] The system introduces a multi-head attention mechanism to dynamically adjust the weights of different edge types. :

[0063] in, and Represents the feature vector of a node. Represents the attention parameter vector. This represents a vector concatenation operation.

[0064] During the time-series modeling phase, the system utilizes gated recurrent units (GRUs) to capture the time dependencies of order demands:

[0065]

[0066]

[0067]

[0068] in, To update the door, To reset the door, Candidate state This is the final hidden state. The output layer predicts the future. Probability distribution of order quantity in each grid .

[0069] S402, the system predicts the traffic conditions of road segments using traffic prediction model 59. This model employs a hybrid architecture of one-dimensional convolutional network (1D-CNN) and long short-term memory network (LSTM).

[0070] The system constructs a multi-dimensional temporal feature sequence for road segments, which is then input into a 1D-CNN layer to extract local temporal patterns. The feature vector output from the convolution operation is subsequently fed into an LSTM layer for long-term dependency modeling, with the state update logic as follows:

[0071]

[0072]

[0073]

[0074]

[0075] in, , , These represent the forget gate, input gate, and output gate, respectively. Represents cellular memory state. This represents a hidden state.

[0076] The model is trained using the Huber loss function, which is robust to outliers.

[0077] in, For the true value, For predicted values, This is a preset threshold parameter.

[0078] S403, the system calculates the estimated order fulfillment time using delivery time prediction model 60. This model is based on a 45-dimensional input feature vector and uses a deep neural network (DNN) for nonlinear regression.

[0079] The input feature vectors cover order dimensions (such as road network distance and priority), delivery personnel dimensions (such as current load and historical on-time rate), road condition dimensions (such as path congestion index and number of traffic lights), and environmental spatiotemporal dimensions (such as time encoding and grid order density).

[0080] The DNN architecture consists of an input layer, multiple hidden layers, and a linear output layer. Each hidden layer performs the following computation:

[0081] To prevent overfitting, each hidden layer is configured with a preset ratio (e.g., ...). to The model employs a Dropout mechanism. It is trained using the Mean Squared Error (MSE) loss function and outputs the estimated delivery time in minutes. .

[0082] For the selection of gradient optimization algorithms and regularization parameters in deep learning models, those skilled in the art can use conventional methods such as the Adam optimizer and L2 regularization. The specific training details are well-known technologies in this field and will not be elaborated here.

[0083] The prediction data generated in the above sub-steps S401 to S403 are directly input into the reinforcement learning policy network in the subsequent step S104, providing a forward-looking state representation for the generation of real-time assigned actions.

[0084] For steps S104 and S105 in the aforementioned overall assignment and allocation process, the specific technical solutions are implemented through sub-steps S501 to S504.

[0085] S501, the system uses Markov decision process modeling to model the real-time capacity allocation process. State space. This includes the global system status, the current status of orders awaiting assignment, and the status of candidate delivery personnel. The global system status includes the total number of online delivery personnel, the size of orders awaiting assignment, and the predicted demand distribution for each grid. Order status includes order coordinates, remaining available time, load weight, and priority. Delivery personnel status includes their real-time coordinates, current load, remaining capacity, and estimated travel time to the pickup point. Action Space This includes discrete actions and continuous actions. Discrete actions include assigning orders to specific delivery personnel, storing orders in a delayed allocation pool for processing in the next cycle, or making order splitting decisions. Continuous actions include adjusting the trade-off coefficients between timeliness and cost in route planning.

[0086] S502, The system constructs a multi-dimensional reward function that is dynamically weighted. This guides the learning preferences of the policy network in different scenarios. Instant rewards. The calculation formula is as follows:

[0087] in, Rewards for timely feedback As a distance cost reward, Rewards for customer satisfaction As a load balancing reward, To constrain the penalty for violations. Weighting coefficients. , , The adaptive weight set generated by step S304 above Provided in real time. Especially during peak periods. When the peak hours are large, the strategy network tends to select the delivery person who can deliver the fastest; when it is in an off-peak scenario and At higher values, the network tends to choose the assignment scheme with the lowest path increment cost.

[0088] In S503, the system utilizes a Deep Q-Network (DQN) to perform online inference for assigning actions. The policy network employs a Dueling architecture, splitting the feature-extracted stream into a value stream. With dominant flow The formula for calculating the final output Q value is:

[0089] In the formula, To share parameters among convolutional layers, For value stream parameters, These are the parameters for the dominant flow. To improve training efficiency and stability, the system introduces a Priority Experience Replay (PER) mechanism, which calculates the sampling probability of samples based on the Temporal Difference Error (TD-error). :

[0090] in, , For the sample TD error, To ensure a minimal constant for non-zero sampling probability.

[0091] S504, the system executes a predictive delivery pre-scheduling mechanism. This mechanism, based on the predicted order distribution output in step S401, achieves pre-emptive intervention in delivery capacity through a two-stage optimization model. The first stage performs coarse scheduling based on the prediction results, using the Hungarian algorithm to solve for the globally optimal delivery capacity offset path. Its objective function is:

[0092] Satisfy constraints as well as ,in For deliverymen Move to target grid The cost, For binary decision variables, For grid The first stage involves forecasting the minimum required capacity. The second stage involves fine-grained assignment as orders are actually generated.

[0093] For routine logic such as delivery route planning and coordinate calculation, those skilled in the art can use open-source heuristic search algorithms or APIs provided by map service providers to implement it. The specific process is well-known in the field and will not be elaborated here.

[0094] Through the aforementioned sub-steps S501 to S504, the system achieves a closed-loop transformation from predicted data to assigned actions. Guided by dynamic weights, the reinforcement learning policy network can smoothly transition and allocate scheduling logic in real time across various business scenarios, such as order surges, capacity shortages, or regular off-peak periods.

[0095] Step S106 in the aforementioned overall assignment and allocation process, namely the process by which the edge-cloud collaborative scheduling architecture ensures instruction execution, is further refined into sub-steps S601 to S604.

[0096] The S601 system performs task decomposition and collaborative deployment between the cloud and edge sides. The cloud intelligence layer 100 is configured with a high-performance GPU server cluster to perform computationally intensive tasks such as full training of deep learning models, global resource optimization, and long-term cross-regional prediction. The edge computing layer 200 consists of edge servers deployed in various regions, performing region-level real-time assignment decisions, local path optimization, and high-frequency state data processing. The cloud and edge sides collaborate through a hierarchical decision-making mechanism, with the cloud distributing macro-level optimization parameters and prediction results, and the edge side completing the final assignment action output within a preset latency threshold.

[0097] S602, the system performs model compression processing for the edge environment. The model compression module 220 uses knowledge distillation technology to transform large-scale deep learning models in the cloud into lightweight models suitable for edge execution. During distillation, the cloud model is used as the teacher model, and the lightweight edge model is used as the student model, through a composite loss function. Guiding students in learning the parameters of the model:

[0098] In the formula, This is the balance coefficient; For real labels Compared with student model predictions Cross-entropy loss between; Predicted output for teacher model Compared with student model prediction output The Kolb-Leibler divergence between the two neural networks is used to transfer the logistic probability distribution features of the teacher model. Furthermore, the system performs model quantization, converting the 32-bit floating-point weights of the neural network into 8-bit integers to reduce memory usage and improve inference speed. Simultaneously, the system employs model pruning techniques to remove neuron connections with contributions below a preset threshold, achieving sparsity in the network structure.

[0099] In S603, the system executes asynchronous communication and data synchronization logic for edge-cloud collaboration. The cloud-based intelligent layer 100 distributes updated prediction models, global optimization parameters, and minute-level prediction data to the edge computing layer 200 via the gRPC protocol, and uses Protocol Buffers format for data serialization to compress transmission volume. The edge computing layer 200 collects delivery personnel coordinates and order status reported by the terminal execution layer 300 via the MQTT protocol, and transmits decision logs and performance indicators back to the cloud in real time. For cross-regional delivery tasks, adjacent edge nodes communicate directly through a distributed consensus protocol to collaboratively complete cross-regional capacity allocation.

[0100] S604 implements a fault tolerance and fault recovery guarantee mechanism. When a connection interruption or hardware failure is detected at an edge node, the system automatically switches the scheduling authority of the affected area to the backup node or allows the cloud intelligent layer 100 to take over until the faulty node recovers. In extreme scenarios where network communication is limited, the edge node maintains independent decision-making capabilities by utilizing locally cached lightweight models and recent periodic prediction data, and performs eventual consistency synchronization of data after the connection is restored.

[0101] For resource scheduling at the underlying hardware driver and operating system levels, those skilled in the art can use conventional containerized deployment and load balancing techniques. The specific implementation process is a well-known technology in this field and will not be elaborated here.

[0102] The above sub-steps S601 to S604, through the distributed deployment of computing power and the simplification and optimization of the model structure, ensure that the system can maintain a response performance of less than 100 milliseconds when facing high-concurrency order requests, and have operational stability in complex network environments.

[0103] The anomaly monitoring algorithm 372 involved in step S106 of the aforementioned overall workflow is further refined into sub-steps S701 to S704.

[0104] S701, the system performs real-time extraction and monitoring of delivery deviation characteristics. The anomaly detection module 230 continuously retrieves real-time trajectory data and task execution status reported by the terminal, extracting multi-dimensional feature indicators including delivery time deviation, path deviation distance, delivery person's dwell time, and order cancellation frequency. Delivery time deviation is defined as the difference between the actual time consumed and the estimated completion time output by the delivery time prediction model 60. Path deviation distance is defined as the vertical Euclidean distance between the delivery person's current GPS coordinates and the path planned by the task navigation module 320.

[0105] S702 utilizes the Isolation Forest algorithm for unsupervised anomaly identification. The system extracts multi-dimensional feature indicators to construct feature vectors, which are then input into a pre-trained Isolation Forest model. This model consists of multiple isolated trees, and the degree of anomaly is determined by calculating the path length of the test sample within the tree structure. Anomaly score is then calculated. The calculation formula is as follows:

[0106] In the formula, Representative sample Path length in the isolation tree Representative sample in The expected path length in the tree set constructed from the samples. Represents a binary search tree in a sample size of The average path length over time is used to standardize the results. When the output outlier score... When the threshold is exceeded (e.g., 0.6), the system determines that the current delivery task has entered an abnormal state.

[0107] S703, the system calls rule engine 376 to perform logical judgment for a specific anomaly type. The rule engine classifies the anomaly samples identified by the isolated forest based on a preset expert rule base. The judgment rules include, but are not limited to: Delivery delay determination: If the current time exceeds the estimated delivery time by a preset threshold (e.g., 10 minutes) and the delivery person is more than the preset radius (e.g., 2km) from the destination, a delivery delay exception will be triggered.

[0108] Determination of unreachable delivery capacity: If the geographical location data of the delivery terminal is not updated within a preset time (e.g., 5 minutes), and the delivery capacity has uncompleted orders, an unreachable delivery capacity condition is triggered.

[0109] Regional backlog determination: If the number of pending orders in a specific grid exceeds the threshold and the average waiting time exceeds the preset upper limit, a regional backlog anomaly will be triggered.

[0110] S704, the system implements an automatic recovery and path correction mechanism for the allocation scheme. Based on different anomaly classification results, the rapid decision module 220 executes the corresponding response plan: In response to delivery delay anomalies, the system triggers order transfer logic, calculates available delivery capacity within a certain range around the current location of the order, and re-executes the assignment action through a reinforcement learning policy network, transferring the order task from the original delivery person to new delivery capacity that can arrive on time, and simultaneously sends a route change instruction to the mobile terminal.

[0111] In response to abnormal capacity disconnection, the system immediately blocks the order-receiving authority of the capacity after failing to establish a communication connection, and releases all pending orders held by the capacity back to the assignment pool for secondary allocation.

[0112] In response to the efficiency decline caused by regional congestion, the system activates predictive path avoidance logic. Using the congestion evolution trend output by the traffic prediction model 59, it recalculates alternative routes to avoid high-risk congested road sections, guiding regional transport capacity to detour through road sections with higher traffic efficiency.

[0113] The determination of the construction depth of the isolation trees and the sampling ratio of subsamples in the isolated forest model can be achieved by conventional parameter optimization methods such as cross-validation, which are well known in the field and will not be elaborated here.

[0114] The monitoring and recovery mechanism determined in steps S701 to S704 above enables the real-time capacity allocation system provided by the present invention to quickly correct the allocation scheme when faced with sudden scenarios such as traffic accidents and local failures caused by extreme weather.

[0115] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art 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 appended claims and their equivalents.

Claims

1. A real-time intelligent delivery capacity allocation system with adaptive switching across multiple business scenarios, characterized in that, The system adopts a three-layer collaborative distributed architecture of cloud prediction, edge decision-making, and terminal execution, including: The cloud-based intelligent layer (100) is used to perform offline model training, global resource optimization, and long-term prediction tasks. The edge computing layer (200) is distributed across the urban grid area to perform regional real-time assignment decisions, local route optimization, and delivery anomaly monitoring. The terminal execution layer (300) is deployed on mobile terminal devices to report capacity status in real time and receive delivery instructions; During the operation of the system, a spatiotemporal feature vector is constructed through a multimodal data acquisition and fusion system (55), a deep learning prediction engine (57) is used to predict the future business status, and a weight set is dynamically generated by the business scenario adaptive switching mechanism (72), and the capacity allocation action is executed in real time by the adaptive scheduling strategy (61).

2. The intelligent delivery capacity real-time allocation system with adaptive switching across multiple business scenarios as described in claim 1, characterized in that, The process of data perception and feature representation performed by the multimodal data acquisition and fusion system (55) includes: The geographic space is divided into geographic grid cells of preset size, and order data streams, delivery personnel data streams, road condition data streams, and environmental data streams are mapped to the corresponding grid indices; Construct a high-dimensional spatiotemporal feature vector that includes order density, delivery personnel density, average road condition index, point of interest distribution density, historical average order volume, environmental and meteorological characteristics, time coding characteristics, and geographic coordinate coding; The feature components are normalized and serialized and stored as feature tensor structures that support the input for prediction inference.

3. The intelligent delivery capacity real-time allocation system with adaptive switching across multiple business scenarios as described in claim 1, characterized in that, The adaptive switching mechanism for the business scenario (72) includes a fuzzy logic controller (354) for adjusting the scheduling focus based on real-time environmental indicators. The specific process includes: Order density, average waiting time, and capacity load rate are extracted as input variables and mapped to the corresponding fuzzy sets through membership functions. Fuzzy inference is performed based on a preset fuzzy rule base, and defuzzification is performed using the centroid method, outputting a dynamic weight set consisting of timeliness weight, cost weight, and satisfaction weight. The dynamic weight set is smoothed in time using the exponential moving average method, and the smoothed weights are fed back to the instant reward function of the adaptive scheduling strategy (61) in real time.

4. The intelligent delivery capacity real-time allocation system with adaptive switching across multiple business scenarios as described in claim 1, characterized in that, The deep learning prediction engine (57) includes an order demand prediction model (58), which uses a spatiotemporal graph convolutional network combined with a multi-head attention mechanism to predict grid-level order density. The specific process includes: Construct a graph structure that includes spatial adjacency relationships and functional similarity relationships; Spatial graph convolutional layers are used to aggregate node features, and multi-head attention mechanisms are used to dynamically adjust the connection weights for different relation types. By using a gated loop unit to capture the time dependency of order demand, the probability distribution of the number of orders in each grid area within a future preset time period is output.

5. The intelligent delivery capacity real-time allocation system with adaptive switching across multiple business scenarios according to claim 1, characterized in that, The deep learning prediction engine (57) also includes a traffic prediction model (59), which adopts a hybrid architecture of one-dimensional convolutional network and long short-term memory network to extract local features and long-term temporal dependencies of road segment traffic status, and uses a loss function that is robust to outliers for model training.

6. The intelligent delivery capacity real-time allocation system with adaptive switching across multiple business scenarios according to claim 1, characterized in that, The adaptive scheduling strategy (61) employs a reinforcement learning policy network, and its assignment logic includes: Establish a state space that includes the global system state, the status of orders to be assigned, and the status of candidate delivery personnel; The assignment action reasoning is performed using a deep Q-network, and its immediate reward function is adjusted in real time by the timeliness weight, cost weight, and satisfaction weight in the weight set. A network architecture that separates the advantage stream and the value stream is adopted, and a priority experience replay mechanism is introduced to extract experience samples based on the sampling probability determined by the time difference error.

7. The intelligent delivery capacity real-time allocation system with adaptive switching across multiple business scenarios according to claim 1, characterized in that, The system also implements a predictive delivery pre-scheduling mechanism (66), which includes: Based on the order distribution prediction results output by the deep learning prediction engine (57), a two-stage optimization model is used to achieve the pre-convergence of transportation capacity. The first stage uses an optimization algorithm to find the capacity offset path with the lowest movement cost, guiding idle capacity to move to high-demand grids. The second phase involves executing detailed assignment actions after the order actually arrives.

8. The intelligent delivery capacity real-time allocation system with adaptive switching across multiple business scenarios according to claim 1, characterized in that, The process of performing model deployment optimization in the edge computing layer (200) includes: By using knowledge distillation technology, a large model in the cloud is used as a teacher model to guide the parameter learning of a lightweight student model on the edge. Its loss function combines classification loss and distribution divergence loss. The model is subjected to weight quantization and pruning. By reducing the bit width precision of the weights and removing redundant connections, the memory usage of edge nodes is reduced and the inference speed is improved.

9. The intelligent delivery capacity real-time allocation system with adaptive switching across multiple business scenarios according to claim 1, characterized in that, The edge computing layer (200) performs correction actions through an anomaly detection algorithm (372), and its monitoring process includes: Extract feature indicators including delivery time deviation and route deviation distance, and input them into the isolated forest model to calculate the anomaly score; Based on the path length of the sample in the isolation tree and the results of standardization, when the anomaly score exceeds the preset threshold, the current delivery task is determined to be in an abnormal state.

10. The intelligent delivery capacity real-time allocation system with adaptive switching across multiple business scenarios according to claim 9, characterized in that, The system's automatic recovery process for abnormal states includes: The rule engine categorizes abnormal samples into delivery delay anomalies, capacity disconnection anomalies, or regional backlog anomalies. In case of delivery delays, the order transfer logic is automatically triggered, and the task is reassigned from the original delivery person to nearby candidate delivery personnel; In cases of abnormal loss of contact with transport capacity, the pending orders held by that capacity will be released and re-entered into the allocation process; In response to abnormal regional congestion, predictive path avoidance logic is activated to guide regional transport capacity to detour through more efficient road sections.

Images