Dynamic rescheduling system and method based on real-time demand

Abstract

The invention discloses a dynamic rescheduling system based on a real-time demand, which comprises a demand prediction module for generating minute-level and grid-level travel demand probability prediction results through a space-time diagram convolutional network and a long and short-term memory network according to a real-time data stream; the multi-target path planning module calculates the optimal path of the empty vehicle through an improved algorithm according to the demand prediction result, the real-time traffic data, the vehicle SOC state and the charging station distribution data; and the collaborative arbitration module carries out arbitration through a game model according to the path planning result and the global vehicle state data to obtain an arbitrated path suggestion. According to the method, the problems of high empty driving mileage, unbalanced supply and demand and electric quantity risk caused by the fact that real-time demand prediction, traffic conditions and electric quantity constraints are not comprehensively considered in an existing unmanned taxi empty driving strategy are solved, the empty driving mileage is remarkably reduced, the operation efficiency is improved, electric quantity exhaustion events are reduced, and the platform transport capacity utilization rate is improved.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, specifically to a dynamic rescheduling system and method based on real-time demand. Background Technology

[0002] In the existing driverless taxi operation model, when a vehicle is empty after completing an order, it is usually dispatched using two strategies: waiting at a nearby station or taking a fixed cruising route. Although these two strategies can maintain the basic operation of the vehicle to a certain extent, their inherent technical limitations are becoming increasingly prominent as the scale of operation expands and the requirements for real-time performance increase.

[0003] Existing technical solutions suffer from the following main problems: In demand forecasting, existing models rely heavily on historical order data, weather, holidays, and other macro-level characteristics for prediction. Their forecast granularity is typically hourly or coarser, lacking the ability to capture dynamic demand at the minute or meter level, resulting in an inability to effectively respond to sudden changes in real-time travel demand. Regarding route planning, existing solutions primarily focus on optimizing the order dispatch stage, with insufficient research on proactive route planning during the no-order, empty-running phase. They generally ignore key constraints such as real-time traffic conditions and remaining vehicle battery power, leading to poor feasibility of planned routes and even increasing the risk of vehicles running out of power. Furthermore, existing solutions often rely on individual vehicle decisions, lacking a platform-level collaborative arbitration mechanism. This can easily lead to multiple vehicles converging on the same hotspot area, creating new supply-demand imbalances. Additionally, the lack of an effective dynamic backoff mechanism to address situations where orders have not yet materialized after vehicle arrival can easily result in secondary empty runs. Summary of the Invention

[0004] The purpose of this invention is to provide a dynamic rescheduling system based on real-time demand. This invention integrates minute-level demand forecasting, real-time traffic information and power constraints, and has a dynamic rescheduling system and method with platform collaboration and fault tolerance capabilities, which significantly reduces empty mileage, improves capacity utilization and operational safety.

[0005] To achieve this objective, the present invention provides a dynamic rescheduling system based on real-time demand, comprising: The multi-objective path planning module is used to calculate the optimal path for empty vehicles based on demand forecast results, real-time traffic data, real-time monitored SOC status information and charging station distribution data, and output the path planning results. The collaborative arbitration module is used to arbitrate based on the path planning results and global vehicle status data through a game model, and obtain the arbitrated path suggestion.

[0006] By integrating demand forecasting, power monitoring, multi-objective path planning, and collaborative arbitration into a closed-loop system, a fundamental shift from passive response to proactive scheduling has been achieved. This breaks the limitations of isolated decision-making by each module in traditional scheduling, enabling empty vehicles to make comprehensive decisions based on high-precision forecasts, real-time power consumption, and traffic conditions. It solves problems such as high empty mileage, regional supply and demand imbalance, and power consumption risks, thereby systematically improving capacity utilization and safety.

[0007] Preferably, it also includes a local execution module, which is used to receive route suggestions on the vehicle side, perform local verification and execution, and output the execution status; When the vehicle terminal receives a route suggestion, it automatically triggers a local verification process. Verification fails if the following conditions are met: Condition 1: The estimated remaining battery power upon reaching the destination must be greater than the absolute safety threshold, and a set safety margin must be reserved. If the estimated remaining battery power upon reaching the destination is less than the absolute safety threshold plus the set safety margin, the local verification will fail. Condition 2: Detect whether there are any sudden situations ahead of the path that have not been updated in time by the cloud. If a sudden obstacle that prevents passage is detected, the local verification will fail. Condition 3: Check if the route contains no-entry, height-restricted, weight-restricted, or incorrect turning instructions. If the route violates traffic rules or vehicle physical constraints, the verification will fail.

[0008] By introducing a local execution module and setting strict verification conditions, a vehicle-side security redundancy layer is added on top of cloud-based decision-making. This grants vehicles the right to veto decisions in extreme situations, ensuring that dispatch instructions will not lead to vehicles being in dangerous or illegal situations due to information lag in the cloud. This greatly enhances the robustness and security of the system. At the same time, an auditing mechanism balances vehicle-side autonomy with platform collaboration.

[0009] Preferably, it also includes a demand forecasting module for generating demand forecasting results based on real-time data streams using a hybrid forecasting model that integrates a spatiotemporal graph convolutional network and a long short-term memory network. The specific steps for generating demand forecasting results using a hybrid prediction model that integrates spatiotemporal graph convolutional networks and long short-term memory networks include: Based on real-time traffic data and historical order data from the real-time data stream, spatial encoding is performed. Each order point is converted into a unique code for a grid of a set spatial size using the GeoHash algorithm (a geohashing algorithm used to encode two-dimensional latitude and longitude coordinates into a short string, achieving spatial discretization and hierarchical partitioning). This code is then converted into a numerical vector. Temporal encoding is performed using Relative time embedding technology (a technique used to convert time information into a continuous numerical vector). This results in a feature vector for each grid node that integrates historical orders, real-time traffic, and time features. A spatiotemporal graph is constructed, where each geographic grid is a node in the spatiotemporal graph, and the feature vector generated by each grid node is the corresponding node feature. The edges between nodes are defined based on the spatial adjacency or road network connectivity between grids. A spatiotemporal graph describing the spatial relationships of the city is constructed based on each geographic grid node, the feature vector generated by each grid node, and the edges between nodes. Based on the spatiotemporal graph, a multi-layer spatiotemporal graph convolutional network is used to aggregate the feature information of each geographic grid node with its neighboring grids to obtain spatial dependencies. At the same time, by performing temporal convolution in the time dimension, the short-term dynamic changes of each node's features are captured. Based on spatial dependencies and the short-term dynamic changes of each node's features, a spatiotemporal feature tensor is obtained through a multi-layer spatiotemporal graph convolutional network. The feature vector of each grid node is input into the Long Short-Term Memory (LSTM) network. The LSTM network automatically learns the long-term temporal dependencies of the feature vectors and performs decoding and prediction according to the time order. The output prediction value is mapped to a probability value through the sigmoid activation function in the last fully connected layer of the LSTM, thus obtaining the grid demand probability matrix and prediction confidence. The grid demand probability matrix and prediction confidence are the demand prediction results.

[0010] By capturing spatial correlations through spatiotemporal graph convolutional networks and combining them with LSTM to learn long-term temporal patterns, the hybrid prediction model can simultaneously capture the spatiotemporal dynamics of demand, achieving high-precision predictions at the minute and hundred-meter levels. This overcomes the shortcomings of traditional historical data predictions, which have coarse granularity and slow response, and provides more accurate demand prediction results for subsequent path planning, thereby improving scheduling effectiveness from the source.

[0011] Preferably, the implementation steps for calculating the optimal path for an empty vehicle using a multi-objective path planning algorithm include: When a vehicle enters an empty driving state, a route planning request is immediately triggered; The path planning request is used to find a path that maximizes the comprehensive score function: the urban road network is represented as a graph structure, the heuristic cost is calculated based on the power consumption, the multi-objective path planning algorithm starts from the starting point of the route, continuously combines adjacent nodes, calculates the total cost, and prioritizes expanding the node with the minimum total cost. When the algorithm reaches a target node that satisfies the constraints, the search terminates and the optimal path is obtained by backtracking. The constraints include the set minimum safe battery level and the set maximum no-load distance.

[0012] Preferably, the specific formula for calculating the heuristic cost is as follows: ; in, Weighted by order revenue. Weighting for power consumption penalty The heuristic cost, i.e., the estimated cost from the current node n to the target point, This represents the maximum grid probability along the path from n to the target. This represents the difference between the current battery level and the safe battery level. Total cost The specific calculation formula is as follows: ; in, The actual cost from the starting point to the current node n is given by the following formula: ; in, For road section The congestion index, Weighting of congestion penalties For the length of the road segment, For road section exist The probability of demand at any given moment. For passing through the section The estimated travel time.

[0013] By quantitatively assessing order revenue, congestion costs, and battery risk, the algorithm can weigh the conflicts of multiple objectives in real time and find the comprehensive optimal solution. Integrating battery power as a state dimension into graph search can accurately predict range, prevent vehicles from failing midway due to improper battery planning, and ensure the feasibility of scheduling.

[0014] Preferably, the specific formula for calculating the scoring function is as follows: ; in, The nodes are represented in a set format. Each node not only represents a geographical location, but also shows the remaining battery power at the corresponding location through (intersection location, SOC range).

[0015] The planning objectives were further refined by defining scoring functions and nodes, emphasizing the simultaneous optimization of geographical location and battery status in path evaluation. This ensures that the algorithm not only focuses on short-term order acquisition but also takes into account the long-term operational capacity of the vehicle, preventing over-consumption of battery power due to excessive pursuit of orders, thereby improving the continuous operational efficiency and safety of the vehicle.

[0016] Preferably, it also includes dynamic recalculation of path planning, used to recalculate the optimal path for empty vehicles through a multi-objective path planning algorithm when the new path gain reaches a set condition, reaches a set time, or the traffic event level reaches a set condition.

[0017] By introducing a dynamic recalculation mechanism, the system can adjust its planning in real time according to environmental changes, avoiding vehicles sticking to outdated routes, and quickly responding to sudden changes in road conditions or shifts in demand. This maintains the timeliness and adaptability of the scheduling strategy and reduces efficiency losses caused by information update delays.

[0018] Preferably, the implementation steps for arbitration using a game theory model, based on path planning results and global vehicle state data, include: Input the path planning results, i.e., the grid of intended routes submitted by all empty vehicles, construct a game model, define each vehicle as a player, and determine which grid each vehicle chooses to go to. Calculate the payoff function for path planning, the formula of which is: ; Let g be the probability of demand. This parameter represents the number of vehicles currently planning to travel to grid g. It's used to prevent swarming; the more vehicles there are, the lower the benefit. Let i be the distance from vehicle i to grid g; The optimal response iterative algorithm is used to calculate the scheduling scheme for empty vehicles when the Nash equilibrium is stable. After one round of iteration, that is, after all vehicles have reselected, the platform checks whether any vehicles have changed their strategies. If the strategies of all vehicles have not changed in this round, it means that the system has reached the Nash equilibrium and the algorithm has converged. If any vehicles have changed their strategies, the next round of iteration begins, and the optimal path for empty vehicles is recalculated using the multi-objective path planning algorithm until convergence is successful or the set number of rounds is reached. If convergence is not achieved after the set number of rounds is reached, the platform will initiate forced arbitration.

[0019] The game arbitration model dynamically adjusts vehicle distribution through a payoff function, transforming individual vehicle decisions into platform-level collaborative optimization. This avoids new supply-demand imbalances caused by blind vehicle aggregation, improves the efficiency of overall capacity allocation, and ensures the enforceability and stability of arbitration results through an iterative convergence mechanism.

[0020] Preferably, a dynamic rescheduling method based on real-time demand includes: Based on demand forecasting results, real-time traffic data, real-time monitored SOC status information, and charging station distribution data, the optimal path for empty vehicles is calculated using a multi-objective path planning algorithm, and the path planning results are output. Based on the route planning results and global vehicle status data, arbitration is conducted through a game theory model to obtain the arbitration-based route suggestion.

[0021] The beneficial effects of this invention are as follows: This invention proposes a dynamic rescheduling system based on real-time demand. The demand prediction module generates travel demand probabilities at the minute and hundred-meter levels, solving the problem of excessively coarse granularity in prediction time and space in existing technologies, and achieving the goals of reducing empty mileage and increasing the average daily effective operating time per vehicle. The multi-objective path planning module integrates order probability, real-time traffic data, remaining battery power, and charging station distribution, solving the problem of ignoring traffic mutations and battery constraints, and achieving the goals of shortening vehicle arrival time in high-demand areas and reducing the average waiting time for users to hail a ride. The collaborative arbitration module performs global vehicle scheduling based on a game theory model, solving the clustering problem caused by the lack of platform-level collaboration in individual vehicle decision-making, and achieving the goals of improving the overall platform capacity utilization rate and reducing the operating cost per unit kilometer. The local execution module performs on-board verification and fault tolerance processing, solving the problems of poor path executability and lack of dynamic backoff mechanism, and achieving the goals of reducing low-battery shutdown events and increasing the charging return route success rate. This invention effectively solves the shortcomings of high energy consumption, low efficiency and insufficient reliability in the existing empty-running dispatch of driverless taxis by real-time data fusion, multi-objective optimization and distributed collaboration, and significantly improves the capacity utilization rate and user experience. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the structure of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of 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 a part of the embodiments of the present invention, not all of them. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to represent selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0024] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: Example 1 A dynamic rescheduling system based on real-time demand, such as Figure 1 As shown, it includes: The multi-objective path planning module is used to calculate the optimal path for empty vehicles based on demand forecast results, real-time traffic data, real-time monitored SOC status information and charging station distribution data, and output the path planning results. The collaborative arbitration module is used to arbitrate based on the path planning results and global vehicle status data through a game model, and obtain the arbitrated path suggestion.

[0025] In some preferred embodiments, it also includes a local execution module for receiving route suggestions on the vehicle side, performing local verification and execution, and outputting the execution status; When the vehicle terminal receives a route suggestion, it automatically triggers a local verification process. Verification fails if the following conditions are met: Condition 1: The estimated remaining battery power upon reaching the destination must be greater than the absolute safety threshold, and a set safety margin must be reserved. If the estimated remaining battery power upon reaching the destination is less than the absolute safety threshold plus the set safety margin, the local verification will fail. Condition 2: Detect whether there are any sudden situations ahead of the path that have not been updated in time by the cloud. If a sudden obstacle that prevents passage is detected, the local verification will fail. Condition 3: Check if the route contains no-entry, height-restricted, weight-restricted, or incorrect turning instructions. If the route violates traffic rules or vehicle physical constraints, the verification will fail.

[0026] In some preferred embodiments, it also includes a demand forecasting module for generating demand forecasting results based on real-time data streams using a hybrid forecasting model that integrates a spatiotemporal graph convolutional network and a long short-term memory network. The specific steps for generating demand forecasting results using a hybrid prediction model that integrates spatiotemporal graph convolutional networks and long short-term memory networks include: Based on real-time traffic data and historical order data in the real-time data stream, spatial encoding is performed. Each order point is converted into a unique code for a grid of a set spatial size using the GeoHash algorithm, and this code is then converted into a numerical vector. Temporal encoding is performed, and time information is converted into a vector using Relative time embedding technology. Each grid node generates a feature vector that integrates historical orders, real-time traffic, and time features. A spatiotemporal graph is constructed, where each geographic grid is a node in the spatiotemporal graph, and the feature vector generated by each grid node is the corresponding node feature. The edges between nodes are defined based on the spatial adjacency or road network connectivity between grids. A spatiotemporal graph describing the spatial relationships of the city is constructed based on each geographic grid node, the feature vector generated by each grid node, and the edges between nodes. Based on the spatiotemporal graph, a multi-layer spatiotemporal graph convolutional network is used to aggregate the feature information of each geographic grid node with its neighboring grids to obtain spatial dependencies. At the same time, by performing temporal convolution in the time dimension, the short-term dynamic changes of each node's features are captured. Based on spatial dependencies and the short-term dynamic changes of each node's features, a spatiotemporal feature tensor is obtained through a multi-layer spatiotemporal graph convolutional network. The feature vector of each grid node is input into the Long Short-Term Memory (LSTM) network. The LSTM automatically learns the long-term temporal dependencies of the feature vectors and performs decoding and prediction according to the time order. The output prediction value is mapped to a probability value through the sigmoid activation function in the final fully connected layer of the LSTM, thus obtaining the grid demand probability matrix and prediction confidence. The grid demand probability matrix and prediction confidence are the demand prediction results.

[0027] In some preferred embodiments, the steps of calculating the optimal path for an empty vehicle using a multi-objective path planning algorithm include: When a vehicle enters an empty driving state, a route planning request is immediately triggered; The path planning request is used to find a path that maximizes the comprehensive score function: the urban road network is represented as a graph structure, the heuristic cost is calculated based on the power consumption, the multi-objective path planning algorithm starts from the starting point of the route, continuously combines adjacent nodes, calculates the total cost, and prioritizes expanding the node with the minimum total cost. When the algorithm reaches a target node that satisfies the constraints, the search terminates and the optimal path is obtained by backtracking. The constraints include the set minimum safe battery level and the set maximum no-load distance.

[0028] Regarding heuristic costs, in some preferred embodiments, they are used to estimate the potential gains or costs of reaching the target area from an intermediate location on the path. In optional embodiments, the specific calculation formula is as follows: ; in, Weighted by order revenue. Weighting of congestion penalties Weighting for power consumption penalty The heuristic cost, i.e., the estimated cost from the current node n to the target point, This represents the maximum grid probability along the path from n to the target. The difference between the current battery level and the safe battery level. , Let n be the geometric distance from the current node n to the target point goal. For vehicle traffic speed, , For safe power consumption, Current battery level; Total cost Used to comprehensively evaluate the advantages and disadvantages of a path from the starting point to the destination, and to measure the actual cost. Compared with the estimated heuristic cost The sum of , and its specific calculation formula is: ; in, The actual cost from the starting point to the current node n is given by the following formula: ; Among them, for path From the starting point 0 to the current node Each section of the road The costs are accumulated. For road section The congestion index, For the length of the road segment, For road section exist The probability of demand at any given moment. For passing through the section The estimated travel time, based on the route The remaining electricity at the destination is estimated using an energy consumption integral model, and the specific calculation formula is as follows: ; in, For road section Predicted average speed The average slope of the road section. The power factor for air conditioning is 0.8 kWh / 10 km in summer, 1.1 kWh / 10 km in winter, and 0.3 kWh / 10 km in transitional seasons, as specified in the optional embodiment. The battery-wheel efficiency is 0.92 in an optional embodiment, where k1, k2, and k3 are the speed energy consumption coefficient, gradient energy consumption coefficient, and air conditioning energy consumption coefficient, respectively.

[0029] In optional embodiments, the parameter units and some values ​​are shown in Table 1 below: Table 1 Parameter Illustration In an optional embodiment, the specific calculation formula for the scoring function is as follows: ; in, The nodes are represented in a set format. Each node not only represents a geographical location, but also shows the remaining power at the corresponding location through (intersection location, SOC range). α+β+γ =1. The platform can be configured remotely.

[0030] Regarding the demand probability and the maximum grid probability, in some preferred embodiments, the demand probability is obtained through the demand prediction module, which generates the demand prediction result based on a hybrid prediction model that integrates a spatiotemporal graph convolutional network and a long short-term memory network, and maps the output prediction value to a probability value; the maximum grid probability is based on the grid demand probability matrix, which is dynamically tracked in the path search, and the larger of the current node demand probability value and the next node demand probability value of the driving path is selected as the maximum grid probability.

[0031] In some preferred embodiments, it also includes dynamic recalculation of path planning, used to recalculate the optimal path for empty vehicles using a multi-objective path planning algorithm when the new path gain reaches a set condition, reaches a set time, or the traffic event level reaches a set condition.

[0032] In an optional embodiment, when the new path gain is ≥5%, the time reaches 60 seconds, or the traffic incident level is ≥3, the optimal path for the empty vehicle is recalculated using a multi-objective path planning algorithm, wherein the urban road traffic incident classification standard GB / T9108-2012 is adopted. Level 1 is mild congestion (v / c 0.7–0.85); Level 2 is congestion (v / c 0.85–1.0); Level 3 is severe congestion / accident with ≥1 lane occupied (v / c > 1.0 or accident); Level 4 is disruption / control. Therefore, when the traffic incident level is ≥3, i.e., severe congestion or accident, the route must be recalculated immediately.

[0033] In some preferred embodiments, the step of arbitrating based on path planning results and global vehicle state data using a game theory model includes: Input the path planning results, i.e., the grid of intended routes submitted by all empty vehicles, construct a game model, define each vehicle as a player, and determine which grid each vehicle chooses to go to. Calculate the payoff function for path planning, the formula of which is: ; Let g be the probability of demand. This parameter represents the number of vehicles currently planning to travel to grid g. It's used to prevent swarming; the more vehicles there are, the lower the benefit. Let i be the distance from vehicle i to grid g; The optimal response iterative algorithm is used to calculate the scheduling scheme for empty vehicles in the Nash equilibrium stable state. After one round of iteration, that is, after all vehicles have reselected, the platform checks whether any vehicles have changed their strategies. If the strategies of all vehicles have not changed in this round, it means that the system has reached Nash equilibrium and the algorithm has converged. If any vehicles have changed their strategies, the next round of iteration begins, and the optimal path for empty vehicles is recalculated using the multi-objective path planning algorithm until convergence is successful or a set number of rounds is reached. If convergence is not achieved after the set number of rounds, the platform will initiate forced arbitration. In the optional implementation scheme, the platform performs arbitration every 60 seconds, and the set number of rounds is 10.

[0034] In some preferred implementations, during forced arbitration, if the Best-Response iterations fail to converge after reaching a preset maximum number of rounds (default 10 rounds), the platform immediately initiates the "forced arbitration" process. Instead of completely random allocation, a decision is made in one go based on the following deterministic priorities: Calculate the regret value for each vehicle across all candidate grids: Regret value = highest profit in the current round – profit from the current strategy of this car, sorted in descending order of regret value, and allocated accordingly: If the regret values ​​are the same, the one with the higher remaining battery power will be compared first; if the battery power is still the same, the one with the lower vehicle ID will be compared (vehicle ID is globally unique to avoid the same order).

[0035] Once the grid capacity is full, the grid is closed, and subsequent vehicles can only choose the second-best option.

[0036] By using mandatory arbitration, allocation can be completed within 0.1 seconds. In the optional embodiment, only 0.8% of the scenarios enter mandatory arbitration, ensuring that vehicles with high regret rates get priority to the "most desired" grid, with an overall benefit loss of <1%.

[0037] Example 2 A dynamic rescheduling method based on real-time demand includes: Based on demand forecasting results, real-time traffic data, real-time monitored SOC status information, and charging station distribution data, the optimal path for empty vehicles is calculated using a multi-objective path planning algorithm, and the path planning results are output. Based on the route planning results and global vehicle status data, arbitration is conducted through a game theory model to obtain the arbitration-based route suggestion.

[0038] Example 3 A computer program product includes a computer program, characterized in that, when the computer program is executed by a processor, it implements the steps of the method described in Embodiment 2.

[0039] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

Claims

1. A dynamic rescheduling system based on real-time demand, characterized in that, It includes: The multi-objective path planning module is used to calculate the optimal path for empty vehicles based on demand forecast results, real-time traffic data, real-time monitored SOC status information and charging station distribution data, and output the path planning results. The collaborative arbitration module is used to arbitrate based on the path planning results and global vehicle status data through a game model, and obtain the arbitrated path suggestion.

2. The dynamic rescheduling system based on real-time demand according to claim 1, characterized in that: It also includes a local execution module, which is used to receive route suggestions on the vehicle side, perform local verification and execution, and output the execution status; When the vehicle terminal receives a route suggestion, it automatically triggers a local verification process. Verification fails if the following conditions are met: Condition 1: The estimated remaining battery power upon reaching the destination must be greater than the absolute safety threshold, and a set safety margin must be reserved. If the estimated remaining battery power upon reaching the destination is less than the absolute safety threshold plus the set safety margin, the local verification will fail. Condition 2: Detect whether there are any sudden situations ahead of the path that have not been updated in time by the cloud. If a sudden obstacle that prevents passage is detected, the local verification will fail. Condition 3: Check if the route contains no-entry, height-restricted, weight-restricted, or incorrect turning instructions. If the route violates traffic rules or vehicle physical constraints, the verification will fail.

3. The dynamic rescheduling system based on real-time demand according to claim 1, characterized in that: It also includes a demand forecasting module, which generates demand forecasting results based on real-time data streams using a hybrid forecasting model that integrates spatiotemporal graph convolutional networks and long short-term memory networks. The specific steps for generating demand forecasting results using a hybrid prediction model that integrates spatiotemporal graph convolutional networks and long short-term memory networks include: Based on real-time traffic data and historical order data in the real-time data stream, spatial encoding is performed. Each order point is converted into a unique code for a grid of a set spatial size using the GeoHash algorithm, and this code is then converted into a numerical vector. Temporal encoding is performed, and time information is converted into a vector using Relative time embedding technology. As a result, each grid node generates a feature vector that integrates historical orders, real-time traffic, and time features. A spatiotemporal graph is constructed, where each geographic grid is a node in the spatiotemporal graph, and the feature vector generated by each grid node is the corresponding node feature. The edges between nodes are defined based on the spatial adjacency or road network connectivity between grids. A spatiotemporal graph describing the spatial relationships of the city is constructed based on each geographic grid node, the feature vector generated by each grid node, and the edges between nodes. Based on the spatiotemporal graph, a multi-layer spatiotemporal graph convolutional network is used to aggregate the feature information of each geographic grid node with its neighboring grids to obtain spatial dependencies. At the same time, by performing temporal convolution in the time dimension, the short-term dynamic changes of each node's features are captured. Based on spatial dependencies and the short-term dynamic changes of each node's features, a spatiotemporal feature tensor is obtained through a multi-layer spatiotemporal graph convolutional network. The feature vector of each grid node is input into the Long Short-Term Memory (LSTM) network. The LSTM automatically learns the long-term temporal dependencies of the feature vectors and performs decoding and prediction according to the time order. The output prediction value is mapped to a probability value through the sigmoid activation function in the final fully connected layer of the LSTM, thus obtaining the grid demand probability matrix and prediction confidence. The grid demand probability matrix and prediction confidence are the demand prediction results.

4. The dynamic rescheduling system based on real-time demand according to claim 1, characterized in that: The steps for calculating the optimal path for an empty vehicle using a multi-objective path planning algorithm include: When a vehicle enters an empty driving state, a route planning request is immediately triggered; The path planning request is used to find a path that maximizes the comprehensive score function: the urban road network is represented as a graph structure, the heuristic cost is calculated based on the power consumption, the multi-objective path planning algorithm starts from the starting point of the route, continuously combines adjacent nodes, calculates the total cost, and prioritizes expanding the node with the minimum total cost. When the algorithm reaches a target node that satisfies the constraints, the search terminates and the optimal path is obtained by backtracking. The constraints include the set minimum safe battery level and the set maximum no-load distance.

5. A dynamic rescheduling system based on real-time demand according to claim 4, characterized in that: The specific formula for calculating the heuristic cost is as follows: ； in, Weighted by order revenue. Weighting for power consumption penalty The heuristic cost, i.e., the estimated cost from the current node n to the target point, This represents the maximum grid probability along the path from n to the target. This represents the difference between the current battery level and the safe battery level. Total cost The specific calculation formula is as follows: ； in, The actual cost from the starting point to the current node n is given by the following formula: ； in, For road section The congestion index, Weighting of congestion penalties For the length of the road segment, For road section exist The probability of demand at any given moment. For passing through the section The estimated travel time.

6. A dynamic rescheduling system based on real-time demand according to claim 4, characterized in that: The specific formula for calculating the scoring function is as follows: ； in, The nodes are represented in a set format. Each node not only represents a geographical location, but also shows the remaining battery power at the corresponding location through (intersection location, SOC range).

7. The dynamic rescheduling system based on real-time demand according to any one of claims 4-6, characterized in that: It also includes dynamic recalculation of path planning, which is used to recalculate the optimal path for empty vehicles using a multi-objective path planning algorithm when the new path gain reaches a set condition, reaches a set time, or the traffic event level reaches a set condition.

8. A dynamic rescheduling system based on real-time demand according to claim 1, characterized in that: Based on the path planning results and global vehicle status data, the arbitration process using a game theory model includes the following steps: Input the path planning results, i.e., the grid of intended routes submitted by all empty vehicles, construct a game model, define each vehicle as a player, and determine which grid each vehicle chooses to go to. Calculate the payoff function for path planning, the formula of which is: ； Let g be the probability of demand. This parameter represents the number of vehicles currently planning to travel to grid g. It's used to prevent swarming; the more vehicles there are, the lower the benefit. Let i be the distance from vehicle i to grid g; The optimal response iterative algorithm is used to calculate the scheduling scheme for empty vehicles when the Nash equilibrium is stable. After one round of iteration, that is, after all vehicles have reselected, the platform checks whether any vehicles have changed their strategies. If the strategies of all vehicles have not changed in this round, it means that the system has reached the Nash equilibrium and the algorithm has converged. If any vehicles have changed their strategies, the next round of iteration begins, and the optimal path for empty vehicles is recalculated using the multi-objective path planning algorithm until convergence is successful or the set number of rounds is reached. If convergence is not achieved after the set number of rounds is reached, the platform will initiate forced arbitration.

9. A dynamic rescheduling method based on real-time demand, characterized in that, It includes: Based on demand forecasting results, real-time traffic data, real-time monitored SOC status information, and charging station distribution data, the optimal path for empty vehicles is calculated using a multi-objective path planning algorithm, and the path planning results are output. Based on the route planning results and global vehicle status data, arbitration is conducted through a game theory model to obtain the arbitration-based route suggestion.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method of claim 9.

Images