Online car-hailing intelligent order sending method and system based on multi-dimensional rule

Abstract

The invention relates to the technical field of online car-hailing order dispatching, in particular to an online car-hailing intelligent order dispatching method and system based on a multi-dimensional rule, and the method comprises the steps: building a standardized multi-dimensional feature set through collecting passenger orders, driver states, traffic conditions, environmental weather and regional events, and carrying out the combined modeling of regional order demands and driver online conditions based on a recurrent neural network, thereby achieving the intelligent order dispatching of the online car-hailing. Outputting expected passenger and driver thermodynamic distribution, constructing expected difference thermodynamic distribution according to the expected passenger and driver thermodynamic distribution, training a scheduling strategy through reinforcement learning, enabling an unloaded vehicle to actively migrate to a supply and demand gap area before an order is generated, and further combining an order generation condition and driver expected off-duty information to obtain an order generation result; and calculating the accumulated driving time required by the driver to return to the expected off-duty place after the driver completes the service, and selecting the work order for distribution when the time difference between the accumulated driving time and the expected off-duty place is minimum, thereby realizing accurate matching between the order distribution strategy and the driver work cycle. The empty driving rate is effectively reduced, and the quick response capability and the resource configuration efficiency of the system in a dynamic supply and demand environment are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of online car-hailing order dispatching, and particularly to an intelligent online car-hailing order dispatching method and system based on multi-dimensional rules. Background Art

[0002] Online car-hailing travel services based on intelligent terminals have become one of the transportation methods frequently used by urban residents. To achieve an efficient matching of orders and transportation capacity resources, various online car-hailing platforms have generally deployed dispatching systems to distribute real-time orders. However, most existing dispatching systems rely on a small number of static factors such as the geographical location distance between drivers and passengers and the order response duration for matching, lacking the ability to fuse and process multi-source heterogeneous data, and unable to construct a regional-level dynamic supply and demand state representation at the system level. In addition, the prediction ability for short-term supply and demand fluctuations is insufficient, making it difficult for the system to timely identify high-demand areas and guide transportation capacity resources to arrive in advance during peak hours or when emergencies occur, thereby causing local congestion and a large number of empty driving behaviors. At the same time, in existing order dispatching schemes, parameters such as the driver's working cycle, off-duty time, and off-duty location are generally not incorporated into the dispatching system, resulting in some drivers having to perform service tasks that are far away and deviate from the off-duty path after receiving an order, thereby increasing the non-operating driving time and affecting the platform's dispatching satisfaction and efficiency. Summary of the Invention

[0003] To solve the above problems, the present invention provides an intelligent online car-hailing order dispatching method and system based on multi-dimensional rules.

[0004] To achieve the above object, the technical solution adopted by the present invention is:

[0005] An intelligent online car-hailing order dispatching system based on multi-dimensional rules, comprising: a multi-dimensional information collection module, a travel demand prediction module, a pre-dispatching module, and an intelligent order dispatching module that are connected in sequence;

[0006] The multi-dimensional information collection module is used to collect passenger order data, driver real-time status data, traffic condition data, environmental weather data, and regional event data, and generate a multi-dimensional feature set;

[0007] The travel demand prediction module is used to, according to the multi-dimensional feature set, use a recurrent neural network model to perform real-time prediction on the order demand, vehicle supply status, and future order generation trend in each region, and generate a short-term supply and demand prediction result including the expected driver heat distribution and the expected passenger heat distribution;

[0008] The pre-dispatching module is used to, according to the expected driver heat distribution and the expected passenger heat distribution, construct an expected difference heat distribution, and guide empty online car-hailing vehicles to the expected demand region based on a reinforcement learning model;

[0009] The intelligent order dispatching module is used to calculate the estimated arrival time for a online car-hailing driver to complete each order and return to the expected off-duty location based on the expected demand value in the expected demand area and the expected off-duty information of the online car-hailing driver, calculate the difference from the expected off-duty time, obtain the off-duty time difference corresponding to each order, and select the order with the smallest absolute value of the off-duty time difference to dispatch the order to the online car-hailing driver.

[0010] Further, the multi-dimensional information collection module is used to perform the following steps:

[0011] Obtain passenger order data, including passenger identity information, order generation time, departure place, destination, expected passenger arrival time, and passenger historical order records;

[0012] Obtain driver real-time status data, including driver identity information, current location, vehicle status, online duration, driver's expected off-duty time and expected off-duty location;

[0013] Obtain traffic condition data, including real-time road traffic flow, average driving speed of each road section, congestion index, and traffic event information obtained through the navigation API;

[0014] Obtain environmental weather data, including temperature, humidity, precipitation, wind speed, and visibility information obtained through the weather platform API;

[0015] Obtain regional event data, including event information of public holidays and large-scale regional activities obtained through the news information API.

[0016] Further, the generation of the multi-dimensional feature set includes the following steps:

[0017] Preprocess the collected passenger order data, driver real-time status data, traffic condition data, environmental weather data, and regional event data to generate a standardized original data set;

[0018] Based on the standardized original data set, synchronize and align various types of data according to the time and geographical location dimensions to obtain structured input data;

[0019] Perform unified coding and normalization processing on the structured input data to generate a standardized multi-dimensional feature set.

[0020] Further, the training steps of the recurrent neural network model include:

[0021] Use the standardized multi-dimensional feature set as the model input to construct regional input samples arranged in time series;

[0022] Construct an expected passenger heat distribution label based on the actual number of passenger orders in each region within the historical time period, and construct an expected driver heat distribution label based on the actual number of online drivers in each region;

[0023] Use the input sample of the said area, the corresponding expected passenger heat distribution label and the expected driver heat distribution label as supervised learning samples, and input them into a recurrent neural network model for joint training;

[0024] During the training process, introduce a perturbation factor related to regional event data, and perform sample marking and perturbation correction for holidays and large-scale events;

[0025] Construct a joint loss function based on the heat distribution prediction error, calculate the loss value through the joint loss function and perform backpropagation until the model converges.

[0026] Furthermore, the operation formula of the recurrent neural network is as follows:

[0027] h t = σ(W x ·x t + W h ·h t-1 + b);

[0028] Wherein, h t is the hidden state at the current time step t; h t-1 is the hidden state at the previous time step; x t is the normalized multi-dimensional feature set; W x and W h are the input weight matrix and the state transition weight matrix respectively; b is the bias term; σ is the activation function.

[0029] Furthermore, the loss function is as follows:

[0030]

[0031] Wherein, is the total loss value; and P i are the predicted passenger heat distribution value and the corresponding true value for the i-th area respectively; and D i are the predicted driver heat distribution value and the corresponding true value for the i-th area respectively; N is the number of areas participating in the calculation; α is the weighting coefficient of the passenger heat distribution loss term and the driver heat distribution loss term.

[0032] Furthermore, the reinforcement learning model is trained through the following steps:

[0033] Construct a state space, which includes the expected difference heat distribution output by the vehicle demand prediction module, the geographical location of the empty-running drivers within the current time slice, the empty-running state identifier, the traffic passing level of the area where they are located, and the historical dispatching state vector;

[0034] Construct an action space, where the action space includes a set of actions of dispatching an empty-running driver to any one of several candidate differential heat regions or maintaining the current geographical location unchanged;

[0035] Define a reward function, which is a weighted function. The input variables include the empty-running distance from the driver's current location to the target region, the actual order generation rate of the target region, and the number of real-time orders accessed after the driver is dispatched. The output is an immediate reward value in scalar form;

[0036] Based on the state space, action space, and reward function, construct a reinforcement learning training framework, and use a deep Q-network to train the policy network, updating the policy parameters to maximize the target policy function.

[0037] Further, calculating the estimated arrival time for the online car-hailing driver to complete each work order and return to the expected off-duty location according to the expected off-duty information of the online car-hailing driver based on the expected demand value of the expected demand region includes the following steps:

[0038] Obtain a set of work orders to be assigned based on the expected demand value of the expected demand region, extract the pick-up point, destination, and expected service time interval of each order, and construct work order path information;

[0039] Obtain the driver's current location, expected off-duty time, and expected off-duty location, and use them as matching constraint parameters;

[0040] For each work order to be assigned, calculate the cumulative driving time required for the driver to pick up the order from the current location, complete the passenger service, and drive to the expected off-duty location, and calculate the estimated arrival time for the online car-hailing driver to complete each work order and return to the expected off-duty location according to the cumulative driving time and the current timestamp.

[0041] Further, the obtaining of the set of work orders to be assigned based on the expected demand value of the expected demand region includes:

[0042] By setting the order waiting time, collect the work orders within the preset radius of the driver and in the regions where the corresponding expected demand values are greater than the preset threshold to construct a set of work orders to be assigned.

[0043] A method for intelligent order dispatching of online car-hailing based on multi-dimensional rules, applied to the system for intelligent order dispatching of online car-hailing based on multi-dimensional rules described in any one of the foregoing, includes:

[0044] Collect passenger order data, driver real-time status data, traffic condition data, environmental weather data, and regional event data to generate a multi-dimensional feature set;

[0045] According to the multi-dimensional feature set, a recurrent neural network model is used to perform real-time prediction on the order demand, vehicle supply status, and future order generation trend in each region, and generate a short-term supply and demand prediction result including the expected driver heat distribution and the expected passenger heat distribution;

[0046] According to the expected driver heat distribution and the expected passenger heat distribution, an expected difference heat distribution is constructed, and based on a reinforcement learning model, the empty-running online car-hailing is guided to the expected demand area;

[0047] Based on the expected demand value in the expected demand area, according to the expected off-duty information of the online car-hailing driver, calculate the estimated arrival time for the online car-hailing driver to complete each work order and return to the expected off-duty location, and perform a difference calculation with the expected off-duty time to obtain the off-duty time difference corresponding to each work order, and select the work order with the smallest absolute value of the off-duty time difference to dispatch the online car-hailing driver.

[0048] The beneficial effects of the present invention are as follows: The present invention generates a standardized multi-dimensional feature set by collecting passenger order data, driver real-time status data, traffic condition data, environmental weather data, and regional event data, and after preprocessing, structured alignment, and unified coding; the vehicle demand prediction module constructs time series samples based on this feature set, uses a recurrent neural network to jointly model the order demand and driver online situation in each region, outputs the expected passenger heat distribution and the expected driver heat distribution, constructs an expected difference heat distribution by comparing the expected passenger heat distribution and the expected driver heat distribution, and uses a reinforcement learning algorithm to train the scheduling strategy, so that the empty-running online car-hailing actively migrates to the area with a large supply and demand gap before the order is generated; further, in combination with the order generation situation in the expected demand area and the expected off-duty information of the driver, calculate the cumulative driving time for the driver to complete the work order and return to the expected off-duty location, and perform a difference calculation with the expected off-duty time, and dispatch by selecting the work order with the smallest absolute value of the off-duty time difference, realizing the precise matching of the dispatching strategy and the driver's individual work cycle. It constitutes a closed-loop scheduling from data collection, supply and demand prediction, empty-running pre-scheduling to personalized dispatching, significantly improving the response ability and resource allocation efficiency of the system in a dynamic supply and demand environment, realizing the reduction of the driver's non-operating driving time, and improving the platform scheduling satisfaction and efficiency. Description of the Drawings

[0049] Figure 1 is a schematic structural diagram of an online car-hailing intelligent dispatching system based on multi-dimensional rules in the present invention.

[0050] Figure 2 is a flow chart of the training steps of the recurrent neural network model in the present invention. Detailed Embodiment

[0051] Please refer to Figure 1 - Figure 2As shown in the figure, the present invention relates to an intelligent order dispatching system for online car-hailing based on multi-dimensional rules, including: a multi-dimensional information collection module, a vehicle usage demand prediction module, a pre-scheduling module, and an intelligent order dispatching module that are connected in sequence;

[0052] The multi-dimensional information collection module is used to collect passenger order data, driver real-time status data, traffic condition data, environmental weather data, and regional event data, and generate a multi-dimensional feature set;

[0053] The vehicle usage demand prediction module is used to, according to the multi-dimensional feature set, utilize a recurrent neural network model to perform real-time prediction on the order demand, vehicle supply status, and future order generation trend in each region, and generate a short-term supply and demand prediction result including the expected driver heat distribution and the expected passenger heat distribution;

[0054] The pre-scheduling module is used to, according to the expected driver heat distribution and the expected passenger heat distribution, construct an expected difference heat distribution, and guide the empty online car-hailing to the expected demand area based on a reinforcement learning model;

[0055] The intelligent order dispatching module is used to, based on the expected demand value in the expected demand area, according to the expected off-duty information of the online car-hailing driver, calculate the expected arrival time for the online car-hailing driver to complete each work order and return to the expected off-duty location, and perform a difference calculation with the expected off-duty time to obtain the off-duty time difference corresponding to each work order, and select the work order with the smallest absolute value of the off-duty time difference to dispatch the work order to the online car-hailing driver.

[0056] In some embodiments, first, the multi-dimensional information collection module is deployed at the platform-side data access layer to collect multi-source heterogeneous data in real time, including passenger order data, driver real-time status data, traffic condition data, environmental weather data, and regional event data. Taking the core area of a certain city as an example, the passenger order data includes attributes such as the order initiation time, the coordinates of the departure and destination, and the expected arrival time of the passenger; the driver status data includes the driver's current location, online status, current task information, expected off-duty time, and off-duty location; the traffic and weather data are pulled in real time through a third-party API interface, and combined with the geographical grid coding method, all data are uniformly mapped to spatial grid cells. The collected data is processed through cleaning, time synchronization, and location standardization, and finally constitutes a structured standardized multi-dimensional feature set, providing a consistent input for the downstream model. Subsequently, the system calls the vehicle demand prediction module to perform regional-level modeling of the supply-demand status. Based on the constructed spatio-temporal grid input features, this module uses a recurrent neural network (RNN) to jointly model the order density and driver activity within each time window. Specifically, a multi-layer RNN structure is adopted to embed the order and driver status information in the time series, constructing a bidirectional prediction path, and respectively outputting the expected passenger heat distribution and the expected driver heat distribution. During the model training process, the historical regional order number and the number of online drivers are used as supervision labels, and weather and holiday event perturbation factors are introduced as external perturbation terms, and a joint loss function is used for optimization, where the passenger heat error and the driver heat error are weighted and fused with weight coefficients. Through this model, a supply-demand status map covering the entire region can be generated in each prediction cycle (such as 5 minutes), providing a decision-making basis for pre-scheduling and order dispatching. The pre-scheduling module further uses the above heat prediction results to construct a heat difference matrix, that is, calculates the difference between the expected passenger heat distribution and the expected driver heat distribution in units of grids to form a difference heat map representing the degree of supply-demand imbalance. This module introduces a reinforcement learning algorithm based on a deep Q-network (DQN), constructs a state space with the difference heat distribution, the driver's current location, the traffic congestion level, and historical scheduling records, defines the action space as dispatching to each candidate heat region or maintaining the current location, and continuously optimizes the scheduling strategy function by designing an immediate reward function (such as an increase in the order acceptance success rate, a shortening of the empty driving distance, etc.). The system jointly trains the strategy through offline historical data and online real-time feedback, and finally enables empty driving drivers to be actively guided to future high-demand regions before orders are generated, realizing the pre-positioning of resources. At the order assignment stage, the intelligent order assignment module makes an optimized decision by combining the supply-demand prediction results and the driver's personalized off-duty information.Specifically, based on the prediction that an order will be generated in a certain high-demand area, the system retrieves the drivers who are currently within the dispatching range and in the order-receiving state. For the drivers who are about to finish work, their current location, expected off-duty time, and off-duty location are extracted, and the cumulative driving time for them to complete the service and return to the off-duty location after receiving the order is calculated, and the difference is calculated with their expected off-duty time. The system constructs a sequence of off-duty time differences for all candidate orders and preferentially assigns the order that minimizes the absolute value of this difference, ensuring the smallest deviation between the assigned route and the driver's original itinerary plan, thereby reducing non-operating mileage and improving driver satisfaction and system dispatching stability.

[0057] By constructing an intelligent order dispatching architecture with multi-module collaboration, multiple mechanisms such as multi-dimensional feature perception, time series modeling, supply-demand difference prediction, reinforcement learning scheduling optimization, and personalized path matching are introduced at the algorithm level, forming a full-process closed-loop system composed of data-driven prediction, policy-driven dispatching, and constraint-driven order dispatching, achieving significant technological breakthroughs and systematic innovations overall. Specifically, on the one hand, different from the existing systems that rely on a small amount of static information for order dispatching, by introducing a standardized multi-dimensional feature set and a recurrent neural network model, the system realizes regional-level time series prediction of the supply-demand state and constructs a dynamic difference expression of the heat distribution of passengers and drivers, fundamentally improving the platform's prediction ability for future supply-demand evolution and response pre-emption. On the other hand, the reinforcement learning mechanism is embedded in the dispatching decision-making. Not only is the difference heat map used as the core variable of the reinforcement learning state space, but the historical dispatching behavior of empty-running drivers and traffic conditions are jointly modeled to realize a self-learning and adaptive empty-running guidance strategy, thereby dynamically alleviating the supply-demand mismatch problem at the system level. In addition, in terms of personalized dispatching, the driver's expected off-duty time and off-duty location are incorporated into the order dispatching constraints. By constructing a time simulation model for the return path, the total path time consumption after the order service is completed is accurately calculated, and with the core optimization goal of "minimizing the off-duty time difference", the dynamic coordination between the platform task dispatching and the driver's work rhythm is realized. Combining the expected demand value filtering mechanism introduced in the screening logic of candidate orders further ensures the dual optimization of order matching in terms of geographical supply-demand heat and time constraints.

[0058] Furthermore, the multi-dimensional information acquisition module is used to perform the following steps:

[0059] Obtain passenger order data, including passenger identity information, order generation time, departure place, destination, passenger expected arrival time, and passenger historical order records;

[0060] Obtain driver real-time status data, including driver identity information, current location, vehicle status, online duration, driver expected off-duty time, and expected off-duty location;

[0061] Obtain traffic condition data, including real-time road traffic flow, average driving speed of each section, congestion index and traffic event information obtained through the navigation API;

[0062] Obtain environmental weather data, including temperature, humidity, precipitation, wind speed and visibility information obtained through the weather platform API;

[0063] Obtain regional event data, including event information on public holidays and large-scale regional activities obtained through the news information API.

[0064] It should be noted that in terms of passenger order data collection, the data stream interface on the platform side is responsible for receiving in real time the order request information submitted by the passenger client. Taking a certain area in the center of Beijing as an example, the platform will record the passenger identity number, the order generation timestamp, the origin and destination coordinates (both in latitude and longitude format), the passenger's estimated arrival time (set by the passenger or estimated by the platform), and the order behavior records within a certain time window in the past, such as statistical parameters of the average car-hailing time and cancellation rate within 30 days. These data are not only used to describe the individual behavior of passengers, but also can be used to construct the regional order density characteristics in subsequent predictions. In the collection of real-time driver status data, the system continuously obtains the location information of currently online drivers, vehicle status (such as whether carrying passengers, whether empty, vehicle operation status, etc.), driver online duration (cumulative running time since going online), expected off-duty time and expected off-duty location through real-time synchronization with the driver-side application. Among them, the driver's expected off-duty time and off-duty location can be obtained through the preset information fields in the system by the driver, or automatically calculated by the platform through the driver's past off-duty habit data. This part of the data is an important input for realizing subsequent dispatching optimization based on the individual work cycle. The acquisition of traffic condition data mainly relies on the third-party navigation API service interface. The system obtains information such as the real-time traffic flow, the average vehicle speed of each road section, the congestion index (such as TPI), and sudden traffic events (such as road closures, construction, traffic accidents, etc.) by setting the sampling period and regional scope. Taking a 2-kilometer by 2-kilometer geographical grid as a unit, the system collects traffic condition information every 60 seconds and maps it to the dispatching area grid to construct a road map model reflecting the current accessibility and travel cost, which is used to support driver route estimation and travel time calculation. The acquisition of environmental weather data relies on the weather platform API (such as the data interface of the China Meteorological Administration or third-party services such as HeFeng Weather) to obtain meteorological observation information covering the dispatching range. The collection dimensions include but are not limited to indicators such as the current temperature, humidity, precipitation, wind speed, and visibility. For example, when the system detects that heavy rainfall will occur in a certain area within the next hour, the platform can consider in advance the impact of the weather on the pick-up time and the travel tendency of passengers, and make adaptive adjustments to the heat prediction and dispatching logic. The meteorological data can be obtained at the hourly level and synchronized to the standardized feature set. In terms of regional event data collection, the system extracts structured regional public event information, including legal holidays, major festival activities, traffic control announcements, large-scale performances and competitions, etc. by accessing news platform APIs and government announcement interfaces. These events usually have the characteristics of suddenness and high aggregation, and have a significant impact on the short-term order demand in a specific area. The system will perform event annotation and time window coding processing on these event information. For example, the "music festival activity" within a certain time period in a certain area is marked as a high-passenger-flow event, and a perturbation term weight will be introduced in the supply-demand prediction model for the corresponding area for correction to improve the robustness of the model to abnormal fluctuation scenarios.

[0065] Further, the generating of the multi-dimensional feature set includes the following steps:

[0066] Preprocess the collected passenger order data, driver real-time status data, traffic condition data, environmental weather data, and regional event data to generate a standardized original data set;

[0067] Based on the standardized original data set, synchronously align various types of data according to the time and geographical location dimensions to obtain structured input data;

[0068] Perform unified encoding and normalization processing on the structured input data to generate a standardized multi-dimensional feature set.

[0069] In some embodiments, the collected passenger order data, driver real-time status data, traffic condition data, environmental weather data, and regional event data are first uniformly preprocessed. Taking the morning rush hour period in a certain city in April 2025 as an example, the system receives a large amount of asynchronously collected data within a unit time (such as every 5 minutes) window. The passenger order data includes information such as order ID, generation time, departure location coordinates, destination location coordinates, and estimated passenger boarding time; the driver status data includes the driver's unique identifier, longitude and latitude coordinates, passenger-carrying status, online duration, expected off-duty time, and destination; the traffic data includes average vehicle speed, congestion index, and real-time traffic events in terms of road segments; the weather data includes regional temperature, wind speed, precipitation intensity, visibility, etc.; the event data encodes whether there are large-scale activities or emergencies within a specific spatio-temporal range through an event recognition model. All data is first de-duplicated and missing value filled at the source, and text fields or symbolic data fields are processed using unified mapping (such as label encoding or hash encoding) to form a standardized original data set with a unified field structure. Subsequently, the system synchronously aligns the standardized original data set based on the dual dimensions of "time window + geographic grid". In terms of the time dimension, the system performs sample slicing with a fixed sliding window (such as 5 minutes) to provide a basis for alignment of different types of data in the same time period; in terms of the spatial dimension, the urban space is divided into regular geographic grids (such as based on S2 or GeoHash spatial encoding), and all order and status data are mapped into the corresponding grid cells to ensure spatial consistency of the data input. For example, in the area numbered Grid_245, the order volume is 12, the number of active drivers is 8, the average congestion index is 0.67, the weather type is rainfall, and the regional event is marked as "there is a large exhibition activity" during the period from 08:00 to 08:05. The system generates a structured input data matrix at this step, where each sample is in the form of "time slice + area code + multi-dimensional index vector". Finally, the system performs unified encoding and normalization processing on the above structured input data to improve the numerical perception stability and gradient propagation performance of the model for input features. Continuous variables such as online duration, average vehicle speed, temperature, order interval, etc. are processed using the Min-Max normalization or Z-score standardization method to scale them to the [0,1] or zero-mean unit variance distribution; discrete variables such as weather category, event type, driver service level, etc. are processed using One-Hot encoding or Embedding mapping to give them a learnable low-dimensional space representation form. A cross-feature construction mechanism is also introduced during the encoding process. For example, "high temperature + holiday" and "heavy rain + traffic congestion" can form composite input variables to enhance the model's representation ability for boundary scenarios.After the above processing, the generated standardized multi-dimensional feature set is organized in the form of a tensor, with dimensions of T×N×D, where T is the number of time slices, N is the number of spatial grids, and D is the number of feature dimensions for each spatio-temporal unit, meeting the condition of being directly input into a recurrent neural network for regional-level sequence prediction.

[0070] Further, the training steps of the recurrent neural network model include:

[0071] Using the standardized multi-dimensional feature set as the model input, construct regional input samples arranged in a time series;

[0072] Construct an expected passenger heat distribution label based on the actual number of passenger orders in each region within the historical time period, and construct an expected driver heat distribution label based on the actual number of online drivers in each region;

[0073] Use the regional input samples and the corresponding expected passenger heat distribution labels and expected driver heat distribution labels as supervised learning samples, and input them into the recurrent neural network model for joint training;

[0074] Introduce a perturbation factor related to regional event data during training, and perform sample marking and perturbation correction for holidays and large-scale events;

[0075] Construct a joint loss function based on the heat distribution prediction error, calculate the loss value through the joint loss function, and perform backpropagation until the model converges.

[0076] It should be noted that, first of all, the system uses a standardized multi-dimensional feature set as the model input, and combines the defined spatio-temporal grid (such as a 5-minute time slice and a 2-kilometer × 2-kilometer spatial grid) to construct a spatio-temporal sequence input tensor, where T is the number of time slices, N is the number of regional grids, and D is the number of multi-dimensional features contained in each unit. These input features include order density, the number of online drivers, road traffic efficiency, weather category, visibility, holiday flag, regional activity code, etc. After normalization and encoding processing, they form input samples in a unified spatio-temporal format. Taking a certain day from 07:00 to 09:00 as an example, in a sequence of T = 24 time slices, the model can obtain the continuous evolution path of the urban operation state and capture the trends and rhythms of supply and demand changes. For the design of the supervision signal, this embodiment does not adopt the traditional single-label modeling, but proposes a joint supervision mechanism based on double labels. That is, a passenger heat distribution label is constructed respectively based on the actual number of passenger orders in each region within the corresponding historical time window, and a driver heat distribution label is constructed based on the actual number of online drivers in the region. The two labels are respectively used as the GroundTruth of the prediction target, and the consistency of the spatial dimension and the time series is retained. Such a design can enable the model to learn the spatial distribution evolution from two perspectives simultaneously during the training process, which helps to enhance its ability to distinguish the "supply-demand asymmetry" phenomenon and provides a more stable basis for constructing the difference heat distribution. In terms of the model structure, this embodiment uses a multi-layer RNN structure to construct a time series modeling path, and parallel prediction branches are set at the output end to independently model the passenger heat distribution and the driver heat distribution. During the training process, regional event data is introduced as a perturbation factor. By constructing special samples of scenarios such as holidays, large-scale events, and traffic restriction measures, and marking them as event perturbation samples, perturbation weights or specific correction factors are introduced for these samples. For example, during holidays, order demands usually shift forward and are concentrated, while during large-scale events, the peak demand position may shift to the surrounding areas of the event venue. Therefore, local amplification weights and spatial translation annotations are introduced in the training samples during such periods to improve the model's ability to identify abnormal spatio-temporal changes. In order to optimize the prediction accuracy of the two output channels simultaneously, this embodiment constructs a joint loss function based on the heat distribution prediction error. This loss function is composed of a weighted combination of the passenger heat distribution error term (such as mean square error) and the driver heat distribution error term.Different from the prior art which only conducts heat modeling based on passenger orders or only focuses on the global balance of supply and demand ratio, the training mechanism of this embodiment has the following remarkable effects: First, by jointly predicting the heat distribution on both the supply and demand sides, the model is equipped with the native support ability for the differential heat distribution, providing a structured prediction basis for subsequent empty driving scheduling strategies; Second, the introduction of the regional event perturbation mechanism significantly enhances the model's perception ability of sudden abnormal scenarios, improving the stability and accuracy of predictions during special holidays and events; Third, the design of the joint loss function makes the model training objective closer to the actual scheduling optimization requirements, effectively avoiding the problem of misleading resource allocation strategies caused by the prediction bias of single-objective models.

[0077] Further, the operation formula of the recurrent neural network is as follows:

[0078] h t =σ(W x ·x t +W h ·h t-1 +b);

[0079] Where h t is the hidden state at the current time step t; h t-1 is the hidden state at the previous time step; x t is the standardized multi-dimensional feature set; W x and W h are the input weight matrix and the state transition weight matrix respectively; b is the bias term; σ is the activation function.

[0080] Specifically, the model takes the corresponding normalized multi-dimensional feature set at each time slice as the current input. By jointly acting with the hidden state of the previous time slice, it generates the hidden state vector of the current time slice, thereby achieving a deep expression of the regional supply and demand state at the current moment. This hidden state not only contains the input feature information at the current moment but also integrates the context information accumulated in historical time slices, enabling the model to have the capabilities of temporal memory and trend perception. In the network structure, the input features mainly include information such as passenger order density, the number of online drivers, traffic congestion index, weather conditions, visibility, holiday identification, and regional event status, all of which have been normalized and standardized in the previous module. At each time step, the model maps the current input features to a high-dimensional space through a set weight matrix, and performs weighted superposition with the state vector of the previous time step. At the same time, a bias term is introduced to enhance the non-linear modeling ability. After being processed by the activation function, the hidden state vector of the current time step is output, which is used to express the position and role of the current regional supply and demand state in the whole-time series evolution process. This recursive calculation process unfolds step by step over the entire time series, finally forming a regional state evolution path containing continuous time slices. The predicted value of each region at each time point is generated jointly driven by its historical state and the current input, thereby achieving continuous modeling and fitting prediction of the short-term supply and demand fluctuation trend. The advantage of this structure is that it not only considers the static features of the current input but also retains the state transfer between time series, enabling the model to still have strong response and dynamic adaptation capabilities under non-stationary conditions such as holiday mutations, event perturbations, or sudden weather.

[0081] Furthermore, the loss function is as follows:

[0082]

[0083] Where, is the total loss value; and P i are respectively the predicted passenger heat distribution value and the corresponding true value of the i-th region; and D i are respectively the predicted driver heat distribution value and the corresponding true value of the i-th region; N is the number of regions participating in the calculation; α is the weighting coefficient of the passenger heat distribution loss term and the driver heat distribution loss term.

[0084] It should be noted that within each time slice, the system outputs two prediction values for each area after urban division, namely the predicted value of the order demand density on the passenger side and the predicted value of the transport capacity supply density on the driver side. At the same time, the system extracts the actual number of passenger orders and the number of online drivers under this time slice from historical real data as supervision labels. During the training process, the model calculates the prediction error of the passenger heat distribution and the prediction error of the driver heat distribution respectively, and superimposes the two to form a total loss value. In order to adjust the optimization weights of the two target variables in model training, the system introduces two configurable weighting parameters, corresponding to the error terms on the passenger side and the driver side respectively, so that the model has adaptability in different application scenarios. This joint loss function continuously iteratively updates the parameters of the neural network through the backpropagation mechanism, enabling the model to learn the evolution laws of the two targets simultaneously. Spatially, it can capture the regional differences in supply and demand distribution; temporally, it can track the trend of the evolution of the supply and demand relationship over time, thus providing a high-precision prediction basis for subsequent differential heat distribution calculation and resource scheduling strategy generation. Different from traditional models that usually only model the passenger order volume or the driver activity number separately, the joint loss mechanism adopted in this embodiment breaks the separation state between supply and demand prediction tasks and improves the model's overall understanding ability of the regional supply and demand structure. Especially in the face of complex spatio-temporal perturbation scenarios such as holidays, large-scale events, and extreme weather, joint modeling can more effectively identify supply and demand imbalance areas and improve prediction robustness.

[0085] Furthermore, the reinforcement learning model is trained through the following steps:

[0086] Construct a state space, which includes the expected differential heat distribution output by the vehicle demand prediction module, the geographical location of empty-running drivers within the current time slice, the empty-running state flag, the traffic passing level of the area where they are located, and the historical scheduling state vector;

[0087] Construct an action space, which includes the action set of dispatching an empty-running driver to any one of several candidate differential heat regions or maintaining the current geographical location unchanged;

[0088] Define a reward function, which is a weighted function. The input variables include the empty-running distance from the driver's current location to the target area, the actual order generation rate of the target area, and the real-time order access times after the driver is dispatched. The output is an immediate reward value in scalar form;

[0089] Based on the state space, action space, and reward function, construct a reinforcement learning training framework, and use a deep Q-network to train the policy network, updating the policy parameters to maximize the target policy function.

[0090] It should be noted that, first of all, the system constructs the state space in reinforcement learning. This state space comprehensively considers the expected difference heat distribution output by the vehicle usage demand prediction module, that is, the dynamic gap between the passenger order demand and the driver capacity supply in each region within the future time window; at the same time, it incorporates the geographical location information of all empty-running drivers in the current time slice, the identifier of whether they are in the empty-load state, and the traffic passing grade of the region where they are located (such as the congestion index or average passing speed from the navigation API); in addition, the historical scheduling trajectory vector of each driver in the past several time slices is also introduced to capture the past scheduling distribution and response behavior of the system to this driver, so as to form a high-dimensional state feature input. Secondly, the system defines the action space, that is, the set of scheduling actions that the system can choose for each empty-running driver. The action space includes guiding the driver to several candidate regions with relatively high expected supply-demand differences, or choosing to maintain the current position unchanged. These candidate regions are screened from the regions with the top-ranked supply-demand differences in the difference heat map. For example, if a driver is currently located in the southwestern region of the urban area, and the system finds through the difference heat map that there are dense orders but a lack of drivers in the northeastern region in the next 15 minutes, then this region is taken as one of the action options. The system can vectorize each action through action encoding so that the reinforcement learning network can identify and evaluate it. In terms of the design of the reward function, in this embodiment, a weighted combination form is adopted to quantify the comprehensive impact of the scheduling behavior on the overall efficiency of the system as an immediate reward value. Specifically, the reward function considers three core variables: First, the empty-running distance from the driver's current position to the target region, and the shorter the distance, the smaller the penalty; Second, the order generation rate in the target region, which represents the immediate potential revenue capacity of this region; Third, the actual number of orders accessed by the driver in the target region after scheduling, which is used to quantify the scheduling effect. The system sets the corresponding weight coefficients according to business needs, and normalizes the three variables and then weights and synthesizes them into a scalar form of the reward value. This reward value will be used as the direct feedback index for updating the policy network. Finally, based on the above state space, action space and reward function, the system constructs a reinforcement learning training framework. During the training process, a deep Q-network is adopted, combined with the experience replay and target network update mechanisms. In each round of interaction, the optimal action is selected from the state according to the current policy, and after execution, the reward value feedback by the environment is received, and the policy network parameters are updated to maximize the expected cumulative reward. To enhance the stability and convergence speed of the model, mechanisms such as the double Q-learning structure and the fixed Q-target network can be used during training to alleviate the problem of overestimation of Q-values. On the one hand, traditional dispatching systems mostly adopt scheduling algorithms based on the principle of the nearest current position, ignoring the supply-demand trend changes at the system level; on the other hand, some systems are usually based on static rules or empirical parameters and cannot adaptively adjust the policy according to real-time system feedback.In this embodiment, by introducing state modeling driven by the expected difference thermal distribution as the core, reinforcement learning policy generation, and feedback-driven iterative optimization, it can not only dynamically perceive the future supply and demand situation of the system, but also autonomously adjust the policy according to the driver's historical behavior and geographical distribution, realizing the intelligence of the empty driving scheduling policy in a complex traffic system.

[0091] Further, based on the expected demand value of the expected demand area, according to the expected off-duty information of the online car-hailing driver, calculating the estimated arrival time for the online car-hailing driver to complete each order and return to the expected off-duty location includes the following steps:

[0092] Obtain the set of orders to be assigned based on the expected demand value of the expected demand area, extract the passenger boarding point, destination, and expected service time interval of each order, and construct the order path information;

[0093] Obtain the driver's current location, expected off-duty time, and expected off-duty location, and use them as matching constraint parameters;

[0094] For each order to be assigned, calculate the cumulative driving time required for the driver to pick up the order from the current location, complete the passenger service, and drive to the expected off-duty location, and calculate the estimated arrival time for the online car-hailing driver to complete each order and return to the expected off-duty location according to the cumulative driving time and the current timestamp.

[0095] In some specific embodiments, first, in combination with the expected passenger heat distribution output by the vehicle usage demand prediction module, high-demand areas are identified, and a candidate order assignment set is constructed based on the expected demand values of each area. In this set, each order work order includes a pick-up point, a destination, and the corresponding service time interval. For example, the pick-up time window is 08:30–08:40, and the expected service duration is 15 minutes, etc. The system establishes a path planning sub-module based on the historical average commuting speed and traffic index, models the expected service path for each work order, and generates corresponding path information, including structured path features such as the driving time, driving distance, and sections passed from the pick-up point to the destination. At the same time, the system extracts the real-time location information of the current driver and obtains the expected off-duty time and off-duty location set by the driver as path constraint parameters. For example, a driver is currently located in urban area A, with a planned off-duty time of 10:00 and a planned off-duty location at a parking lot in urban area B. These parameters will be used for subsequent path time series simulation and adaptation judgment for each work order. For each candidate work order, the system executes the following path time calculation process: First, calculate the driving time for the driver to travel from the current location to the order pick-up point; second, simulate the time required for the service path from the pick-up point to the destination; finally, starting from the destination, calculate the return time for the driver to travel to the preset off-duty location along the shortest passable path. The cumulative sum of the above three stages constitutes the total driving time for the driver to execute this work order. The system adds this cumulative time to the current timestamp to obtain the estimated arrival time for the driver to complete the work order and return to the off-duty location, and calculates the difference from the expected off-duty time set by the driver. The system further records the time difference corresponding to each work order in absolute value form to construct a "work order - time difference" mapping vector. For example, the off-duty time difference for work order A is +5 minutes, for work order B is -3 minutes, and for work order C is +30 minutes. The system sorts all candidate work orders in ascending order of the absolute value of this difference and selects the work order with the smallest absolute value of the time difference as the optimal matching scheme to complete the precise order assignment decision for the current driver. If there are multiple work orders with the same difference, factors such as order service score and path congestion index are introduced for optimization in the secondary sorting.

[0096] The prior art mainly selects the last order based on whether the driver is ready to go home. This embodiment realizes a personalized scheduling strategy based on a dynamic path prediction and time difference matching mechanism. On the one hand, this mechanism uses the orders in the expected heat area as order assignment candidates to ensure that the scheduling priority falls on high-demand areas; on the other hand, the driver's "work cycle" information is used as a scheduling constraint to participate in path calculation and time simulation, so that the order assignment result can fit the driver's off-duty path and time limit, significantly reducing the empty driving mileage and non-operating duration during the return journey, and enhancing the flexibility and interpretability of the platform's scheduling strategy.

[0097] Further, the obtaining of the work order set to be assigned based on the expected demand value of the expected demand area includes:

[0098] By setting the order waiting time, collect the work orders within the preset radius of the driver and where the corresponding expected demand value in the area is greater than the preset threshold to construct a set of work orders to be assigned.

[0099] In some embodiments, specifically, when the platform performs personalized order dispatching, it needs to screen out a part from the unassigned orders in the current system as a candidate set. If orders are extracted within a fixed radius only based on the driver's current location without screening, the following problems will occur: In low-demand areas, although there are several orders near the driver, these orders are scattered and the subsequent supply-demand density is insufficient, which is likely to cause the driver to be unable to smoothly enter the next high-demand area after completing the current work order, thereby increasing the empty driving ratio; at the same time, if the number of work orders to be assigned is too small, the algorithm cannot provide sufficient choices in terms of optimization objectives (such as the difference in off-duty time), ultimately resulting in limited matching effects. Therefore, the system first sets a preset threshold based on the demand prediction values of each area in the expected demand heat map, and screens out the high-demand areas where the expected demand value exceeds this threshold. Only when the pick-up point of the order is within these areas does the order qualify to be included in the set of work orders to be assigned. This design ensures the optional space for drivers approaching the end of their shifts. To improve the screening accuracy, the system sets a preset demand threshold for each area. Only when the expected demand value of a certain area exceeds this threshold does the order within this area have the qualification to enter the queue to be assigned. On this basis, the system then sets the service reachable radius according to the real-time positioning data of the drivers currently in the state of waiting for orders to be dispatched. This radius can be dynamically adjusted in combination with urban density, traffic conditions and the driver's historical response ability. For example, it can be set to 2 kilometers in the urban area and expanded to 5 kilometers in the suburbs. The system retrieves work orders that meet the following two conditions within the reach of the driver: (1) The expected demand value in the area where the pick-up point of its passenger is located is higher than the set threshold; (2) The order creation time of this work order is still within the allowed order waiting time range. This waiting time threshold is the maximum waiting period defined by the platform to ensure the passenger service experience. For example, it is usually set from 2 minutes to 5 minutes to ensure that the assigned orders still have the timeliness for order dispatching. All work orders that meet the above conditions will be collected to form a set of work orders to be assigned and used as the input object set in the subsequent "minimum off-duty time difference" algorithm. By synergistically controlling parameters such as the order waiting time, the driver service radius and the area demand threshold, a highly dynamic and real-time adaptable screening mechanism for work orders to be assigned is established, forming a dispatching process that is closely linked and logically closed-loop with the subsequent return path matching and personalized scheduling links, effectively improving the response efficiency and resource allocation rationality of the dispatching system in actual operation.

[0100] A method for intelligent order dispatching of online car-hailing based on multi-dimensional rules, applied to the system for intelligent order dispatching of online car-hailing based on multi-dimensional rules described in any one of the foregoing, includes:

[0101] Collect passenger order data, driver real-time status data, traffic condition data, environmental weather data, and regional event data to generate a multi-dimensional feature set;

[0102] According to the multi-dimensional feature set, use a recurrent neural network model to make real-time predictions on the order demand, vehicle supply status, and future order generation trend in each region, and generate a short-term supply and demand prediction result including the expected driver heat distribution and the expected passenger heat distribution;

[0103] According to the expected driver heat distribution and the expected passenger heat distribution, construct an expected difference heat distribution, and based on a reinforcement learning model, guide empty-running online car-hailing vehicles to the expected demand area;

[0104] Based on the expected demand value in the expected demand area, according to the expected off-duty information of online car-hailing drivers, calculate the estimated arrival time for online car-hailing drivers to complete each work order and return to the expected off-duty location, and perform a difference calculation with the expected off-duty time to obtain the off-duty time difference corresponding to each work order, and select the work order with the smallest absolute value of the off-duty time difference to dispatch work orders to online car-hailing drivers.

[0105] The above embodiments are only descriptions of the preferred embodiments of the present invention, and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary engineering and technical personnel in the art to the technical solutions of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. An intelligent order dispatching system for online car-hailing based on multi-dimensional rules, characterized in that, Including: A multi-dimensional information collection module, a car-using demand prediction module, a pre-scheduling module, and an intelligent order dispatching module that are connected in sequence; The multi-dimensional information collection module is used to collect passenger order data, driver real-time status data, traffic condition data, environmental weather data, and regional event data, and generate a multi-dimensional feature set; The car-using demand prediction module is used to, according to the multi-dimensional feature set, use a recurrent neural network model to perform real-time prediction on the order demand, vehicle supply status, and future order generation trend in each region, and generate a short-term supply and demand prediction result including the expected driver heat distribution and the expected passenger heat distribution; The pre-scheduling module is used to, according to the expected driver heat distribution and the expected passenger heat distribution, construct an expected difference heat distribution, and guide the empty-running online car-hailing vehicle to the expected demand area based on a reinforcement learning model; The intelligent order dispatching module is used to, based on the expected demand value in the expected demand area, calculate the estimated arrival time for the online car-hailing driver to complete each work order and return to the expected off-duty location according to the expected off-duty information of the online car-hailing driver, perform a difference calculation with the expected off-duty time, obtain the off-duty time difference corresponding to each work order, and select the work order with the smallest absolute value of the off-duty time difference to dispatch the work order to the online car-hailing driver.

2. The intelligent order dispatching system for online car-hailing based on multi-dimensional rules according to claim 1, wherein, The multi-dimensional information collection module is used to perform the following steps: Obtain passenger order data, including passenger identity information, order generation time, departure place, destination, passenger estimated arrival time, and passenger historical order records; Obtain driver real-time status data, including driver identity information, current location, vehicle status, online duration, driver expected off-duty time, and expected off-duty location; Obtain traffic condition data, including road real-time traffic flow, average driving speed of each road section, congestion index, and traffic event information obtained through a navigation API; Obtain environmental weather data, including temperature, humidity, precipitation, wind speed, and visibility information obtained through a weather platform API; Obtain regional event data, including event information on public holidays and large-scale regional activities obtained through a news information API.

3. A car-hailing intelligent order dispatching system based on multi-dimensional rules according to claim 1, characterized in that, The generation of the multi-dimensional feature set includes the following steps: Preprocess the collected passenger order data, driver real-time status data, traffic condition data, environmental weather data, and regional event data to generate a standardized original data set; Based on the standardized original data set, synchronously align various types of data according to the time and geographical location dimensions to obtain structured input data; Perform unified encoding and normalization processing on the structured input data to generate a standardized multi-dimensional feature set.

4. The intelligent order dispatching system for online car-hailing based on multi-dimensional rules according to claim 1, wherein The training steps of the recurrent neural network model include: Use the standardized multi-dimensional feature set as the model input to construct regional input samples arranged in a time series; Construct an expected passenger heat distribution label with the actual number of passenger orders in each region within a historical time period, and construct an expected driver heat distribution label with the actual number of online drivers in each region; Use the regional input samples and the corresponding expected passenger heat distribution labels and expected driver heat distribution labels as supervised learning samples, and input them into the recurrent neural network model for joint training; During the training process, introduce perturbation factors related to regional event data, and perform sample marking and perturbation correction for holidays and large-scale events; Construct a joint loss function based on the prediction error of the heat distribution, calculate the loss value through the joint loss function and perform backpropagation until the model converges.

5. The intelligent order dispatching system for online car-hailing based on multi-dimensional rules according to claim 4, characterized in that, The operation formula of the recurrent neural network is as follows: h t = σ(W x · x t + W h · h t-1 + b); where h t is the hidden state at the current time step t; h t-1 is the hidden state at the previous time step; x t is the normalized multi-dimensional feature set; W x and W h are the input weight matrix and the state transition weight matrix respectively; b is the bias term; σ is the activation function.

6. The intelligent order dispatching system for online car-hailing based on multi-dimensional rules according to claim 5, characterized in that, The loss function is as follows: Among them, is the total loss value; and P i are respectively the predicted passenger heat distribution value and the corresponding true value in the i-th area; and D i are respectively the predicted driver heat distribution value and the corresponding true value in the i-th area; N is the number of areas participating in the calculation; α is the weighting coefficient of the passenger heat distribution loss term and the driver heat distribution loss term.

7. A car-hailing intelligent order dispatching system based on multi-dimensional rules according to claim 1, characterized in that, The reinforcement learning model is trained through the following steps: Construct a state space, which includes the expected difference heat distribution output by the vehicle demand prediction module, the geographical location of the empty-running drivers within the current time slice, the empty-running status flag, the traffic passing grade of the area where they are located, and the historical scheduling status vector; Construct an action space, which includes an action set of dispatching the empty-running drivers to any one of several candidate difference heat regions or maintaining the current geographical location unchanged; Define a reward function, which is a weighted function. The input variables include the empty-running distance from the driver's current location to the target area, the actual order generation rate of the target area, and the real-time order access times after the driver is dispatched. The output is an immediate reward value in scalar form; Based on the state space, action space and reward function, construct a reinforcement learning training framework, and use the deep Q network to train the policy network, and update the policy parameters to maximize the target policy function.

8. A ride-hailing intelligent order dispatching system based on multi-dimensional rules according to claim 1, characterized in that, The calculation of the estimated arrival time for the online car-hailing driver to complete each work order and return to the expected off-duty location based on the expected demand value in the expected demand area and according to the expected off-duty information of the online car-hailing driver includes the following steps: Obtain the set of work orders to be assigned based on the expected demand value in the expected demand area, extract the passenger pick-up point, destination and expected service time interval of each order, and construct the work order path information; Obtain the driver's current location, expected off-duty time and expected off-duty location, and use them as matching constraint parameters; For each work order to be assigned, calculate the cumulative driving time required for the driver to pick up the order from the current location, complete the passenger service, and drive to the expected off-duty location. According to the cumulative driving time and the current timestamp, calculate the estimated arrival time for the online car-hailing driver to complete each work order and return to the expected off-duty location.

9. The intelligent order dispatching system for online car-hailing based on multi-dimensional rules according to claim 8, wherein The obtaining of the set of work orders to be assigned based on the expected demand value in the expected demand area includes: By setting the order waiting time, collect the work orders within the preset radius of the driver and where the corresponding expected demand value in the area is greater than the preset threshold to construct the set of work orders to be assigned.

10. A method for intelligent order assignment of online car-hailing based on multi-dimensional rules, which is applied to an online car-hailing intelligent order assignment system according to any one of claims 1-9, and is characterized in that, Include: Collect passenger order data, driver real-time status data, traffic condition data, environmental weather data and regional event data to generate a multi-dimensional feature set; According to the multi-dimensional feature set, use the recurrent neural network model to perform real-time prediction on the order demand, vehicle supply status and future order generation trend in each area, and generate a short-term supply and demand prediction result including the expected driver heat distribution and the expected passenger heat distribution; According to the expected driver heat distribution and the expected passenger heat distribution, construct an expected difference heat distribution, and guide the empty online car-hailing vehicles to the expected demand area based on the reinforcement learning model; Based on the expected demand value in the expected demand area, according to the expected off-duty information of the online car-hailing driver, calculate the expected arrival time for the online car-hailing driver to complete each order and return to the expected off-duty location, and calculate the difference from the expected off-duty time to obtain the off-duty time difference corresponding to each order. Select the order with the smallest absolute value of the off-duty time difference to dispatch the order to the online car-hailing driver.

Images