Fixed-point area route optimization method and system for responsive public transportation service

Abstract

The present invention provides a method and system for optimizing a fixed-point area path of a responsive public transportation service, which relates to the field of intelligent transportation technology. The method and system include constructing a responsive public transportation service system model, determining a mathematical expression of a fixed-point area path optimization problem; designing a path optimization solver by using an improved ant colony algorithm, executing a local search operation, and locally optimizing an evaluated path to obtain an optimized path set; updating a global pheromone distribution based on the optimized path set; constructing a multi-objective optimization framework, taking the updated global pheromone distribution as prior information, and solving a Pareto optimal solution set by using an improved NSGA-II algorithm under the multi-objective optimization framework; generating a candidate path set according to short-term passenger demand, the Pareto optimal solution set and the updated global pheromone distribution; selecting an optimal path from the candidate path set by using a deep reinforcement learning algorithm, and performing vehicle allocation and scheduling.

Description

Technical Field

[0001] The present invention relates to intelligent transportation technology, and in particular to a method and system for optimizing a fixed-point area path for a responsive public transportation service. Background Art

[0002] Existing responsive bus service systems have many deficiencies in optimizing routes in fixed-point areas. Traditional fixed-route bus services lack flexibility and cannot effectively adapt to dynamically changing passenger demand, especially in areas with low passenger flow or dispersed demand, resulting in inefficiency and waste of resources. Existing route planning methods such as heuristic algorithms and mathematical programming models are inefficient in dealing with large-scale, real-time changing scenarios and are difficult to meet the needs of responsive services. These methods also have difficulty in simultaneously considering multiple conflicting optimization objectives such as service quality, operating costs, and environmental impact.

[0003] Although artificial intelligence technology provides new ideas for solving complex path optimization problems, it still faces many challenges in practical applications. How to effectively represent complex transportation networks and diverse passenger needs, how to achieve real-time optimization with limited computing resources, and how to ensure the interpretability and reliability of algorithms are all urgent issues to be solved. In addition, the existing system is also insufficient in accurately predicting changes in passenger demand in the short term and considering external factors such as road conditions and weather conditions. The integration with the existing traffic management system also needs to be improved. Summary of the invention

[0004] The embodiments of the present invention provide a method and system for optimizing a fixed-point area path of a responsive public transportation service, which can solve the problems in the prior art.

[0005] According to a first aspect of the embodiments of the present invention,

[0006] A method for optimizing the route of a fixed area to provide responsive bus services, including:

[0007] Constructing a responsive public transportation service system model, including a passenger demand module, a vehicle status module and a road network information module, and determining a mathematical expression for a fixed-point area path optimization problem based on the responsive public transportation service system model;

[0008] A path optimization solver is designed using an improved ant colony algorithm, wherein the improved ant colony algorithm includes an adaptive pheromone update mechanism and a dynamic candidate set strategy; an initial path set is constructed using the path optimization solver, and the paths in the initial path set are evaluated according to a preset evaluation index; a local search operation is performed to locally optimize the evaluated paths to obtain an optimized path set; and based on the optimized path set, the global pheromone distribution is updated;

[0009] A multi-objective optimization framework is constructed, which includes a passenger satisfaction objective function, an operating cost objective function and an environmental impact objective function; the updated global pheromone distribution is used as prior information, and the improved NSGA-II algorithm is used to solve the Pareto optimal solution set under the multi-objective optimization framework; short-term passenger demand is predicted based on historical data and real-time information; a candidate path set is generated according to short-term passenger demand, the Pareto optimal solution set and the updated global pheromone distribution; a deep reinforcement learning algorithm is used to select the optimal path from the candidate path set, and vehicles are allocated and scheduled; the global pheromone distribution is updated according to the actual operation results, and the updated information is fed back to the path optimization solver, the multi-objective optimization framework and the dynamic path planning algorithm to continuously optimize the responsive bus service system model.

[0010] In an optional embodiment,

[0011] A responsive public transportation service system model is constructed, including a passenger demand module, a vehicle status module and a road network information module. Based on the responsive public transportation service system model, the steps of determining a mathematical expression of a fixed-point area path optimization problem include:

[0012] The passenger demand module uses a spatiotemporal distribution model to describe passenger demand, and uses a time series prediction model based on a long short-term memory network to predict demand; the vehicle status module uses a multi-dimensional state vector to represent the vehicle status, and uses a Kalman filter algorithm to predict the future status of the vehicle; the road network information module uses a weighted directed graph to represent the road network, and uses a traffic flow prediction model based on a graph convolutional network to update the road network status in real time;

[0013] Generate passenger demand distribution data for the current and future periods based on the spatiotemporal distribution model of the passenger demand module and the time series prediction model of the long short-term memory network; use the passenger demand distribution data as input parameters for the fixed-point area route optimization problem;

[0014] Based on the multi-dimensional state vector and Kalman filter algorithm of the vehicle state module, the real-time state information of the vehicle is calculated and updated; the real-time state information of the vehicle is used as the constraint condition of the fixed-point area path optimization problem;

[0015] The traffic flow prediction model based on the weighted directed graph and graph convolutional network of the road network information module generates real-time road network status data; the real-time road network status data is used to calculate the driving distance and time window constraints of the fixed-point area path optimization problem;

[0016] Based on the constructed responsive bus service system model, the mathematical expression of the fixed-point area path optimization problem is determined by utilizing passenger demand distribution data, real-time vehicle status information, and real-time road network status data. The mathematical expression includes an objective function and constraints. The objective function considers travel distance, vehicle energy consumption, and passenger satisfaction. The constraints include vehicle capacity constraints, time window constraints, and power constraints.

[0017] In an optional embodiment,

[0018] A path optimization solver is designed using an improved ant colony algorithm, wherein the improved ant colony algorithm includes an adaptive pheromone update mechanism and a dynamic candidate set strategy; an initial path set is constructed using the path optimization solver, and the paths in the initial path set are evaluated according to a preset evaluation index; a local search operation is performed to locally optimize the evaluated paths to obtain an optimized path set; based on the optimized path set, the steps of updating the global pheromone distribution include:

[0019] The adaptive pheromone update mechanism uses an adaptive factor that changes dynamically with the number of iterations to adjust the pheromone increment; the dynamic candidate set strategy uses a dynamic threshold that changes with the iteration process to control the size of the candidate set;

[0020] Using a path optimization solver to construct an initial path set, and generating multiple candidate paths through state transition rules and roulette selection method; evaluating the candidate paths in the initial path set according to preset evaluation indicators, wherein the preset evaluation indicators include path length, time reliability, energy efficiency and passenger satisfaction; and selecting high-quality paths according to the evaluation results;

[0021] Perform local search operations on the selected high-quality paths, and use variable neighborhood search strategies to design a variety of neighborhood structures, including 2-opt exchange, node insertion, and subpath reversal; obtain the optimized path set by changing the neighborhood structure;

[0022] Based on the optimized path set, the global pheromone distribution is updated, the idea of ​​elite ant colony system is adopted, and the pheromone increments of the current iterative optimal path and the global optimal path are considered at the same time; the idea of ​​maximum and minimum ant colony system is introduced to limit the pheromone concentration within the maximum and minimum range of dynamic adjustment;

[0023] The initial path set construction, path evaluation, local search optimization and global pheromone distribution update steps are repeated until the preset convergence condition is met or the maximum number of iterations is reached.

[0024] In an optional embodiment,

[0025] The probability of selection in the roulette wheel selection method is defined as:

[0026] ;

[0027] Among them, each parameter represents the meaning:

[0028] p ij (t): The probability that the ant chooses to go to node j from the current node i at time t;

[0029] τ ij (t): pheromone concentration on the path connecting node i and node j at time t;

[0030] η ij (t): heuristic information from node i to node j;

[0031] α: pheromone importance factor;

[0032] β: heuristic information importance factor;

[0033] l: index variable, used to represent all possible next nodes of node i;

[0034] τ il (t): pheromone concentration on the path connecting node i and node l at time t;

[0035] η il (t): heuristic information from node i to node l;

[0036] N i (t): At time t, the candidate set of node i, that is, the set of next nodes that the ant can currently choose;

[0037] j∈N i (t): indicates that node j is an element in the candidate set of node i;

[0038] The numerator of the formula [τ ij (t)] α *[η ij (t)] β represents the total attraction of the path from node i to node j;

[0039] The denominator of the formula Σ[τ il (t)] α *[η il (t)] β is a normalization factor, which is the sum of the total attractiveness of all possible choices. By dividing by this sum, we get a probability distribution that ensures that the sum of the probabilities of all possible choices is 1;

[0040] By adjusting the values ​​of α and β, the balance between utilizing pheromones and heuristic information can be controlled.

[0041] In an optional embodiment,

[0042] A local search operation is performed on the selected high-quality paths, and a variable neighborhood search strategy is used to design a variety of neighborhood structures, including 2-opt exchange, node insertion, and subpath reversal. The steps of obtaining the optimized path set by changing the neighborhood structure include:

[0043] The selected high-quality path is used as the current solution;

[0044] Design a variety of neighborhood structures and build a collection of neighborhood structures, including 2-opt exchange structure, node insertion structure and subpath inversion structure;

[0045] A local search operation is performed on the current solution, wherein the local search operation adopts a variable neighborhood search strategy, including:

[0046] Randomly select nodes or edges in the path in the current neighborhood structure, and generate new solutions according to the rules of the current neighborhood structure. The 2-opt exchange structure randomly selects two non-adjacent edges to disconnect and reconnect to form a new path, the node insertion structure randomly selects a node and inserts it to another position in the path, and the subpath reversal structure randomly selects a subpath and reverses its order.

[0047] Performing local optimization on the rule-generated new solution of the current neighborhood structure to obtain a local optimal solution;

[0048] Calculate the objective function value of the local optimal solution and compare it with the objective function value of the current solution; update the search state according to the comparison result. When the local optimal solution is better than the current solution, set the local optimal solution as the new current solution and reset the neighborhood structure index to the initial value; otherwise, increase the neighborhood structure index and switch to the next neighborhood structure;

[0049] Determine whether the neighborhood structure index exceeds the size of the neighborhood structure set. If not, return to the step of randomly generating a new solution in the current neighborhood structure and continue searching in the current neighborhood structure. If it exceeds, enter the next round of search.

[0050] The local search operation is repeated until a preset stopping criterion is met.

[0051] In an optional embodiment,

[0052] A multi-objective optimization framework is constructed, wherein the multi-objective optimization framework includes a passenger satisfaction objective function, an operating cost objective function, and an environmental impact objective function; the updated global pheromone distribution is used as prior information, and the steps of solving the Pareto optimal solution set using the improved NSGA-II algorithm under the multi-objective optimization framework include:

[0053] The passenger satisfaction objective function takes into account the passenger waiting time; the operating cost objective function takes into account the vehicle's driving distance and energy consumption, and the energy consumption calculation involves the vehicle type, load and road conditions; the environmental impact objective function takes into account the vehicle emission factor and energy consumption, and the emission factor is dynamically determined according to the vehicle type and the use of new energy;

[0054] Establish constraints, including vehicle capacity constraints, time window constraints, and power constraints;

[0055] Creating and maintaining a global pheromone distribution matrix, wherein the global pheromone distribution matrix represents path attractiveness and is dynamically adjusted through pheromone update rules;

[0056] The improved NSGA-II algorithm is used to solve the Pareto optimal solution set. The improved NSGA-II algorithm includes: generating an initial population based on the global pheromone distribution, and using a priority coding method to represent individuals; performing adaptive crossover and mutation operations, and the crossover and mutation probabilities are dynamically adjusted according to the individual fitness; using an improved crowding calculation method based on k-nearest neighbors to evaluate the diversity of solutions; performing simulated annealing local search on elite individuals to improve the quality of solutions; performing non-dominated sorting and environmental selection to generate a new generation of populations; updating the global pheromone distribution and strengthening the path corresponding to the high-quality solution; and using the hypervolume indicator to evaluate the quality and convergence of the Pareto frontier.

[0057] Repeat the iteration until the termination condition is met;

[0058] The density of the final Pareto solution set is controlled and the boundary is preserved, and the Pareto optimal solution set is output.

[0059] In an optional embodiment,

[0060] The steps of selecting the optimal path from the candidate path set and performing vehicle allocation and scheduling using the deep reinforcement learning algorithm include:

[0061] Collect vehicle state vectors, passenger demand matrices, traffic condition matrices, candidate path sets and current time information to construct initial state information; construct a state space based on the initial state information, and design an action space representing the path assigned to the vehicle;

[0062] Design a path feature extraction module, including a multi-layer perceptron, to extract the total distance, estimated completion time, number of passengers, passenger satisfaction estimate, and path complexity features of candidate paths;

[0063] Using the multi-layer perceptron to perform feature encoding on each candidate path to obtain a path feature encoding vector;

[0064] The attention mechanism is used to aggregate the path feature encoding vectors to generate path set encoding;

[0065] Connect the path set encoding with the initial state information to form an enhanced state representation;

[0066] Use historical route data to pre-train the route feature extraction module through self-supervised learning;

[0067] Construct a deep Q network, including an input layer, multiple hidden layers, and an output layer, where the input layer receives the enhanced state representation and the output layer corresponds to the size of the action space;

[0068] Initialize the experience replay buffer to store state transition samples;

[0069] Adopting the priority experience replay strategy, the sample priority is calculated according to the temporal difference error, and sampling is performed based on the priority;

[0070] Use a dual deep Q network structure, including an online network and a target network, where the online network is used for action selection and the target network is used to calculate the target Q value;

[0071] The loss function is defined based on mean square error with importance sampling weights;

[0072] A dynamic ε-greedy strategy is used for exploration and exploitation, and the ε value decays with the number of training steps;

[0073] The path feature extraction module and the main Q network parameters are optimized simultaneously through the back propagation algorithm;

[0074] Update the target network parameters regularly to keep them consistent with the online network parameters;

[0075] Input the current enhanced state representation into the trained deep Q network and calculate the Q value of each path in the candidate path set; select the path with the highest Q value from the candidate path set as the optimal path, and assign the optimal path to the corresponding vehicle;

[0076] Execute the selected optimal path, observe environmental changes, and obtain new status information and rewards;

[0077] The new state information, executed actions, rewards obtained, and next state information are stored in the experience replay buffer as state transition samples;

[0078] Repeat the steps from constructing the initial state information to storing the state transition samples to optimize the deep Q network.

[0079] According to a second aspect of the embodiments of the present invention,

[0080] A point-to-point regional route optimization system that provides responsive bus services, including:

[0081] The first unit is used to construct a responsive public transportation service system model, including a passenger demand module, a vehicle status module and a road network information module, and to determine a mathematical expression of a fixed-point area path optimization problem based on the responsive public transportation service system model;

[0082] The second unit is used to design a path optimization solver using an improved ant colony algorithm, wherein the improved ant colony algorithm includes an adaptive pheromone update mechanism and a dynamic candidate set strategy; construct an initial path set using the path optimization solver, and evaluate the paths in the initial path set according to a preset evaluation index; perform a local search operation, locally optimize the evaluated paths, and obtain an optimized path set; and update the global pheromone distribution based on the optimized path set;

[0083] The third unit is used to construct a multi-objective optimization framework, which includes a passenger satisfaction objective function, an operating cost objective function, and an environmental impact objective function; using the updated global pheromone distribution as prior information, and using the improved NSGA-II algorithm to solve the Pareto optimal solution set under the multi-objective optimization framework; predicting short-term passenger demand based on historical data and real-time information; generating a candidate path set based on short-term passenger demand, the Pareto optimal solution set, and the updated global pheromone distribution; using a deep reinforcement learning algorithm to select the optimal path from the candidate path set, and perform vehicle allocation and scheduling; updating the global pheromone distribution based on actual operation results, and feeding back the updated information to the path optimization solver, the multi-objective optimization framework, and the dynamic path planning algorithm to continuously optimize the responsive bus service system model.

[0084] According to a third aspect of the embodiments of the present invention,

[0085] An electronic device is provided, comprising:

[0086] processor;

[0087] a memory for storing processor-executable instructions;

[0088] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0089] A fourth aspect of the embodiments of the present invention is:

[0090] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0091] The variable neighborhood search strategy adopted in this paper has significant advantages in path optimization problems. First, the design of multiple neighborhood structures increases the diversity of the search and improves the probability of finding the global optimal solution. Second, the mechanism of dynamically switching the neighborhood structure helps to jump out of the local optimum and avoid being stuck in search stagnation. Furthermore, the adaptive adjustment mechanism can dynamically adjust the strategy according to the feedback information of the search process to improve the efficiency of the algorithm. Finally, this method is easy to implement and has relatively low computational complexity, which is suitable for large-scale practical problems. Through this method, high-quality path optimization solutions can be obtained in a relatively short time, providing strong decision-making support for traffic scheduling, logistics distribution and other fields, and ultimately achieving reduced operating costs and improved service quality.

[0092] The deep reinforcement learning-based method adopted in this paper has many advantages. First, it can handle high-dimensional and complex state spaces and adapt to dynamically changing traffic environments. Secondly, through experience replay and priority sampling, the sample utilization efficiency and learning effect are improved. Furthermore, the introduction of the path feature extraction module enables the model to better understand the intrinsic characteristics of the path. In addition, the application of the attention mechanism enhances the model's ability to capture key information. Finally, the dual Q network and soft update mechanism improve the stability of training. This method can not only achieve efficient vehicle scheduling in complex urban traffic networks, but also has strong generalization ability, can adapt to path planning problems of different scales and types, and provides strong support for the development of intelligent transportation systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] Figure 1 A schematic diagram of a flow chart of a method for optimizing a fixed-point area path for a responsive public transportation service according to an embodiment of the present invention;

[0094] Figure 2 It is a structural schematic diagram of a fixed-point area path optimization system for responsive public transportation service according to an embodiment of the present invention. DETAILED DESCRIPTION

[0095] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0096] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0097] Figure 1 FIG. 1 is a flow chart of a method for optimizing a fixed-point area path for a responsive public transportation service according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0098] S1. Construct a responsive bus service system model, including a passenger demand module, a vehicle status module and a road network information module, and determine a mathematical expression for a fixed-point area path optimization problem based on the responsive bus service system model;

[0099] S2. Design a path optimization solver using an improved ant colony algorithm, wherein the improved ant colony algorithm includes an adaptive pheromone update mechanism and a dynamic candidate set strategy; construct an initial path set using the path optimization solver, and evaluate the paths in the initial path set according to a preset evaluation index; perform a local search operation, locally optimize the evaluated paths, and obtain an optimized path set; update the global pheromone distribution based on the optimized path set;

[0100] S3. Construct a multi-objective optimization framework, which includes a passenger satisfaction objective function, an operating cost objective function, and an environmental impact objective function; use the updated global pheromone distribution as prior information, and use the improved NSGA-II algorithm to solve the Pareto optimal solution set under the multi-objective optimization framework; predict short-term passenger demand based on historical data and real-time information; generate a candidate path set based on short-term passenger demand, the Pareto optimal solution set, and the updated global pheromone distribution; use a deep reinforcement learning algorithm to select the optimal path from the candidate path set, and perform vehicle allocation and scheduling; update the global pheromone distribution based on actual operation results, and feed the updated information back to the path optimization solver, the multi-objective optimization framework, and the dynamic path planning algorithm to continuously optimize the responsive bus service system model.

[0101] In an optional embodiment,

[0102] A responsive public transportation service system model is constructed, including a passenger demand module, a vehicle status module and a road network information module. Based on the responsive public transportation service system model, the steps of determining a mathematical expression of a fixed-point area path optimization problem include:

[0103] The passenger demand module uses a spatiotemporal distribution model to describe passenger demand, and uses a time series prediction model based on a long short-term memory network to predict demand; the vehicle status module uses a multi-dimensional state vector to represent the vehicle status, and uses a Kalman filter algorithm to predict the future status of the vehicle; the road network information module uses a weighted directed graph to represent the road network, and uses a traffic flow prediction model based on a graph convolutional network to update the road network status in real time;

[0104] Generate passenger demand distribution data for the current and future periods based on the spatiotemporal distribution model of the passenger demand module and the time series prediction model of the long short-term memory network; use the passenger demand distribution data as input parameters for the fixed-point area route optimization problem;

[0105] Based on the multi-dimensional state vector and Kalman filter algorithm of the vehicle state module, the real-time state information of the vehicle is calculated and updated; the real-time state information of the vehicle is used as the constraint condition of the fixed-point area path optimization problem;

[0106] The traffic flow prediction model based on the weighted directed graph and graph convolutional network of the road network information module generates real-time road network status data; the real-time road network status data is used to calculate the driving distance and time window constraints of the fixed-point area path optimization problem;

[0107] Based on the constructed responsive bus service system model, the mathematical expression of the fixed-point area path optimization problem is determined by utilizing passenger demand distribution data, real-time vehicle status information, and real-time road network status data. The mathematical expression includes an objective function and constraints. The objective function considers travel distance, vehicle energy consumption, and passenger satisfaction. The constraints include vehicle capacity constraints, time window constraints, and power constraints.

[0108] For example, the construction of a responsive bus service system model is the key foundation for optimizing the route in a fixed area. The model contains three core modules: passenger demand module, vehicle status module, and road network information module. These modules provide comprehensive input parameters and constraints for the route optimization problem by integrating data from different dimensions.

[0109] The passenger demand module uses a spatiotemporal distribution model to describe the dynamic characteristics of passenger demand. The model divides the urban area into multiple grids, each of which represents a specific geographical location. For each time period, the model records the passenger boarding and alighting demands of each grid. For example, during the morning rush hour of 7:00-9:00, the boarding demand in commercial area A may be 100 people / hour, while the alighting demand in residential area B may be 80 people / hour. In order to predict future demand changes, the module also integrates a time series prediction model based on the long short-term memory network (LSTM). By analyzing the long-term and short-term dependencies of historical data, the LSTM network can accurately predict the passenger demand trend in the next few hours or even days.

[0110] The vehicle status module uses a multidimensional state vector to represent the real-time status of each bus. This state vector contains multiple key parameters, such as the current position coordinates, speed, passenger capacity, remaining power, etc. For example, the state of an electric bus may be: position (116.3, 39.9), speed 20km / h, current passenger capacity 30, remaining power 70%. In order to predict the future state of the vehicle, the module applies the Kalman filter algorithm. By combining the vehicle's motion model with real-time observation data, the algorithm can effectively estimate the vehicle's position and state at future moments, and maintain high accuracy even in the presence of measurement noise and system errors.

[0111] The road network information module uses a weighted directed graph to represent the urban traffic network. In this graph, nodes represent intersections or stations, edges represent road segments, and the weight of the edge reflects the travel time or congestion level of the road. For example, a main road may be represented as an edge from node A to node B, with a weight of 10 minutes when it is unobstructed and may increase to 20 minutes when it is congested. In order to update the road network status in real time, the module integrates a traffic flow prediction model based on a graph convolutional network (GCN). GCN can capture the spatial dependencies of the road network and use historical traffic data and real-time information to predict future traffic flow and road conditions.

[0112] Based on the output of the passenger demand module, the system generates passenger demand distribution data for the current and future periods. This data includes the expected number of passengers, departure and destination information for each grid area in different time periods. This information is directly used as input parameters for the fixed-point area path optimization problem, helping the system plan the most efficient route to meet passenger demand.

[0113] Using the output of the vehicle status module, the system calculates and updates the status information of each bus in real time. This information is converted into constraints for the route optimization problem. For example, the remaining battery power of the vehicle determines the maximum distance it can travel, and the current passenger capacity affects whether it can accommodate more passengers at a specific stop.

[0114] The system can accurately calculate the travel distance and estimated travel time of different routes through the real-time road network status data generated by the road network information module. These data are used to set the time window constraints in the path optimization problem to ensure that the bus can complete the designated route within a reasonable time.

[0115] Based on the outputs of the above three modules, the system finally determines the mathematical expression of the fixed-point regional route optimization problem. The objective function takes into account three key factors: driving distance, vehicle energy consumption, and passenger satisfaction. For example, the system may weigh the choice of a slightly longer route with lower energy consumption, or a route that can serve more passengers but has a slightly longer driving distance. Constraints include vehicle capacity constraints (to ensure no overloading), time window constraints (to ensure punctuality), and power constraints (to ensure that the vehicle can complete the entire journey).

[0116] The responsive bus service system model based on multi-module integration adopted in this paper has significant advantages. First, it can comprehensively consider various factors that affect the quality of bus services, including passenger demand, vehicle status and road network conditions, so as to formulate a more optimized operation strategy. Secondly, through the application of real-time data updates and prediction models, the system can quickly respond to changes in the traffic environment, dynamically adjust path planning, and improve the flexibility and efficiency of bus services. Finally, the multi-objective optimization method enables the system to find the best balance between operating costs, service quality and environmental impact, and achieve sustainable development of bus services. This method can not only improve passenger satisfaction, but also reduce the operating costs of bus companies, while reducing the environmental burden of urban transportation, providing an innovative solution for smart city traffic management.

[0117] In an optional embodiment,

[0118] A path optimization solver is designed using an improved ant colony algorithm, wherein the improved ant colony algorithm includes an adaptive pheromone update mechanism and a dynamic candidate set strategy; an initial path set is constructed using the path optimization solver, and the paths in the initial path set are evaluated according to a preset evaluation index; a local search operation is performed to locally optimize the evaluated paths to obtain an optimized path set; based on the optimized path set, the steps of updating the global pheromone distribution include:

[0119] The adaptive pheromone update mechanism uses an adaptive factor that changes dynamically with the number of iterations to adjust the pheromone increment; the dynamic candidate set strategy uses a dynamic threshold that changes with the iteration process to control the size of the candidate set;

[0120] Using a path optimization solver to construct an initial path set, and generating multiple candidate paths through state transition rules and roulette selection method; evaluating the candidate paths in the initial path set according to preset evaluation indicators, wherein the preset evaluation indicators include path length, time reliability, energy efficiency and passenger satisfaction; and selecting high-quality paths according to the evaluation results;

[0121] Perform local search operations on the selected high-quality paths, and use variable neighborhood search strategies to design a variety of neighborhood structures, including 2-opt exchange, node insertion, and subpath reversal; obtain the optimized path set by changing the neighborhood structure;

[0122] Based on the optimized path set, the global pheromone distribution is updated, the idea of ​​elite ant colony system is adopted, and the pheromone increments of the current iterative optimal path and the global optimal path are considered at the same time; the idea of ​​maximum and minimum ant colony system is introduced to limit the pheromone concentration within the maximum and minimum range of dynamic adjustment;

[0123] The initial path set construction, path evaluation, local search optimization and global pheromone distribution update steps are repeated until the preset convergence condition is met or the maximum number of iterations is reached.

[0124] For example, the improved ant colony algorithm is an effective method for solving complex path optimization problems. The algorithm significantly improves the convergence speed and solution quality of the algorithm by introducing an adaptive pheromone update mechanism and a dynamic candidate set strategy.

[0125] The adaptive pheromone update mechanism is one of the core innovations of the algorithm. This mechanism uses an adaptive factor that changes dynamically with the number of iterations to adjust the pheromone increment. Specifically, in the early stages of the algorithm, the adaptive factor is large, promoting global exploration; as the number of iterations increases, the adaptive factor gradually decreases, strengthening local search. For example, the initial adaptive factor may be set to 0.9, and then reduced by 0.1 every 100 iterations until it reaches a minimum value of 0.1. This mechanism can achieve a dynamic balance between exploration and utilization during the operation of the algorithm.

[0126] The dynamic candidate set strategy is another important improvement. This strategy uses a dynamic threshold that changes with the iteration process to control the size of the candidate set. In the early stage of the algorithm, the threshold is low and the candidate set is large to promote diversity; as the iterations proceed, the threshold is gradually increased, the candidate set is reduced, and the focus is on high-quality solutions. For example, the initial threshold can be set to 0.3, and it is increased by 0.05 every 50 iterations until it reaches a maximum value of 0.8. This strategy effectively balances the exploration of the solution space and the use of high-quality solutions.

[0127] The construction of the path optimization solver is a key step in the implementation of the algorithm. The solver first generates multiple candidate paths through state transition rules and roulette selection. The state transition rule takes into account pheromone strength and heuristic information such as distance, time and other factors. The roulette selection method selects the path based on the comprehensive weight of these factors. For example, for a decision point, if there are three optional paths with comprehensive weights of 0.5, 0.3 and 0.2 respectively, then the probability of selecting these three paths is 50%, 30% and 20% respectively.

[0128] After the initial set of paths is constructed, the system evaluates the paths according to the preset evaluation indicators. The evaluation indicators include path length, time reliability, energy efficiency and passenger satisfaction. Path length directly affects the operating cost and can be calculated by accumulating the distance of each section. Time reliability reflects the punctuality of the path and the average delay time can be calculated through historical data. Energy efficiency considers the energy consumption of vehicles in different sections and can be estimated based on road conditions and vehicle characteristics. Passenger satisfaction can be comprehensively evaluated by factors such as ride time and number of transfers. The system sorts the paths according to the weighted sum of these indicators and selects high-quality paths.

[0129] For the selected high-quality paths, the system performs local search operations for further optimization. A variable neighborhood search strategy is adopted and a variety of neighborhood structures are designed. The 2-opt exchange operation selects two non-adjacent nodes in the path and exchanges the subpaths between them. The node insertion operation moves a node from its current position to another position in the path. The subpath reversal operation selects a subpath in the path and reverses its order. The system tries these operations in turn and accepts the new solution if a better solution is generated, otherwise the original solution remains unchanged.

[0130] The optimized path set is used to update the global pheromone distribution. The update process adopts the idea of ​​the elite ant colony system, and considers the pheromone increments of the current iterative optimal path and the global optimal path. For example, if the length of the current iterative optimal path is 100km and the length of the global optimal path is 95km, weights of 0.8 and 1.0 can be given respectively to increase the pheromone on the corresponding path. At the same time, the idea of ​​the maximum and minimum ant colony system is introduced to limit the pheromone concentration within the dynamically adjusted maximum and minimum range to prevent the algorithm from converging to the local optimal solution too early. For example, the pheromone concentration can be limited to the range of [0.1, 2.0], and this range can be gradually narrowed as the iteration proceeds.

[0131] The system repeats the above steps, including initial path set construction, path evaluation, local search optimization, and global pheromone distribution update, until the preset convergence condition is met or the maximum number of iterations is reached. The convergence condition can be set to no significant improvement in the optimal solution after multiple consecutive iterations, for example, the improvement in the optimal solution after 50 consecutive iterations is less than 0.1%. The maximum number of iterations can be set according to the problem scale and computing resources, and can usually be set to 1000-5000 times.

[0132] The improved ant colony algorithm adopted in this paper has significant advantages in solving complex fixed-point regional path optimization problems. First, the adaptive pheromone update mechanism and dynamic candidate set strategy improve the convergence speed of the algorithm and the quality of the solution, and can find high-quality solutions in a relatively short time. Secondly, the diversified local search strategy enhances the ability of the algorithm to jump out of the local optimum and increases the probability of finding the global optimal solution. Thirdly, the path evaluation method considering multiple evaluation indicators makes the optimization results more comprehensive and can simultaneously meet multiple goals such as operational efficiency, service quality and environmental friendliness. Finally, the flexibility of the algorithm enables it to adapt to path optimization problems of different scales and characteristics, providing a powerful decision support tool for intelligent transportation systems. In this way, the efficiency and quality of bus services can be significantly improved, operating costs can be reduced, environmental pollution can be reduced, and ultimately a more convenient and comfortable travel experience can be provided to passengers.

[0133] In an optional embodiment,

[0134] The probability of selection in the roulette wheel selection method is defined as:

[0135] ;

[0136] Among them, each parameter represents the meaning:

[0137] p ij (t): The probability that the ant chooses to go to node j from the current node i at time t;

[0138] τ ij (t): pheromone concentration on the path connecting node i and node j at time t;

[0139] η ij (t): heuristic information from node i to node j;

[0140] α: pheromone importance factor;

[0141] β: heuristic information importance factor;

[0142] l: index variable, used to represent all possible next nodes of node i;

[0143] τ il (t): pheromone concentration on the path connecting node i and node l at time t;

[0144] η il (t): heuristic information from node i to node l;

[0145] N i (t): At time t, the candidate set of node i, that is, the set of next nodes that the ant can currently choose;

[0146] j∈N i (t): indicates that node j is an element in the candidate set of node i;

[0147] The numerator of the formula [τ ij (t)] α *[η ij (t)] β represents the total attraction of the path from node i to node j;

[0148] The denominator of the formula Σ[τ il (t)] α *[η il (t)] β is a normalization factor, which is the sum of the total attractiveness of all possible choices. By dividing by this sum, we get a probability distribution that ensures that the sum of the probabilities of all possible choices is 1;

[0149] By adjusting the values ​​of α and β, the balance between utilizing pheromones and heuristic information can be controlled.

[0150] For example, the roulette wheel selection method is a key decision-making mechanism in the ant colony algorithm, which is used to determine the behavior of ants in the path selection process. This method calculates the probability of ants choosing each possible path by comprehensively considering the pheromone concentration and heuristic information, thereby realizing intelligent path selection.

[0151] The core of the roulette wheel selection method is the calculation of selection probability. For each time point t, when the ant is at node i, the probability of choosing to go to node j depends on several factors. The first is the pheromone concentration on the path connecting node i and node j, denoted as τ ij (t). The pheromone concentration reflects the historical ant selection preference. The higher the concentration, the more times the path is selected, and the better the quality may be. The second is the heuristic information from node i to node j, denoted by η ij (t). Heuristic information is usually related to the actual physical characteristics of the path, such as the inverse of the distance. The shorter the path, the greater the heuristic information value.

[0152] In order to balance the influence of pheromone and heuristic information, two important parameters are introduced: pheromone importance factor α and heuristic information importance factor β. These two parameters are used to adjust the relative importance of pheromone concentration and heuristic information in the decision-making process. By adjusting the values ​​of α and β, the balance between the algorithm's use of historical experience (pheromone) and current local information (heuristic information) can be controlled.

[0153] The calculation process of the selection probability can be described as follows: First, for each possible next node j of node i, its attractiveness is calculated. The attractiveness is represented by the product of the α power of the pheromone concentration and the β power of the heuristic information. Then, the attractiveness of the node is divided by the sum of the attractiveness of all possible selected nodes to obtain the probability of selecting the node.

[0154] Specifically, suppose the current node i has three optional next nodes: A, B, and C. The pheromone concentration of node A is 0.5, and the heuristic information is 2; the pheromone concentration of node B is 0.3, and the heuristic information is 3; the pheromone concentration of node C is 0.4, and the heuristic information is 1.5. If α=1, β=2 are set, the probability of selecting these three nodes can be calculated:

[0155] First calculate the attraction of each node:

[0156] The attraction of node A: 0.5^1*2^2=2;

[0157] The attractiveness of node B: 0.3^1*3^2=2.7;

[0158] The attraction of node C: 0.4^1*1.5^2=0.9;

[0159] The total attractiveness is: 2+2.7+0.9=5.6;

[0160] Then calculate the selection probability:

[0161] The probability of selecting node A: 2 / 5.6≈35.7%;

[0162] The probability of selecting node B: 2.7 / 5.6≈48.2%;

[0163] The probability of selecting node C: 0.9 / 5.6≈16.1%;

[0164] This calculation process ensures that the sum of the probabilities of all possible choices is 1, forming a valid probability distribution.

[0165] In practical applications, the values ​​of α and β have an important impact on the performance of the algorithm. A larger α value will make the algorithm more inclined to choose a path with a high pheromone concentration, enhancing the use of historical experience; a larger β value will make the algorithm more inclined to choose a path that currently looks better, enhancing the use of local information. Usually, the value range of α is between 0 and 5, and the value range of β is between 1 and 5. It is necessary to find the parameter combination that best suits the specific problem through experiments.

[0166] The implementation of the roulette wheel selection method also needs to consider the candidate set N i(t). The candidate set contains all the next nodes that the ant can currently choose. When constructing the candidate set, it is necessary to consider the constraints of the problem, such as the connectivity of the path, the capacity limit of the vehicle, the time window constraint, etc. Only nodes that meet these constraints can be included in the candidate set.

[0167] The roulette wheel selection method provides a flexible and effective decision-making mechanism for the ant colony algorithm. It can strike a balance between exploring new paths and using known good paths, improving the algorithm's ability to find the global optimal solution. By adjusting parameters, it can adapt to different types of optimization problems, so that the algorithm can achieve a good balance between convergence speed and solution quality. This method not only improves the performance of the algorithm, but also enhances its adaptability to complex path optimization problems, providing a powerful tool for solving practical problems such as traffic scheduling and logistics distribution.

[0168] In an optional embodiment,

[0169] A local search operation is performed on the selected high-quality paths, and a variable neighborhood search strategy is used to design a variety of neighborhood structures, including 2-opt exchange, node insertion, and subpath reversal. The steps of obtaining the optimized path set by changing the neighborhood structure include:

[0170] The selected high-quality path is used as the current solution;

[0171] Design a variety of neighborhood structures and build a collection of neighborhood structures, including 2-opt exchange structure, node insertion structure and subpath inversion structure;

[0172] A local search operation is performed on the current solution, wherein the local search operation adopts a variable neighborhood search strategy, including:

[0173] Randomly select nodes or edges in the path in the current neighborhood structure, and generate new solutions according to the rules of the current neighborhood structure. The 2-opt exchange structure randomly selects two non-adjacent edges to disconnect and reconnect to form a new path, the node insertion structure randomly selects a node and inserts it to another position in the path, and the subpath reversal structure randomly selects a subpath and reverses its order.

[0174] Performing local optimization on the rule-generated new solution of the current neighborhood structure to obtain a local optimal solution;

[0175] Calculate the objective function value of the local optimal solution and compare it with the objective function value of the current solution; update the search state according to the comparison result. When the local optimal solution is better than the current solution, set the local optimal solution as the new current solution and reset the neighborhood structure index to the initial value; otherwise, increase the neighborhood structure index and switch to the next neighborhood structure;

[0176] Determine whether the neighborhood structure index exceeds the size of the neighborhood structure set. If not, return to the step of randomly generating a new solution in the current neighborhood structure and continue searching in the current neighborhood structure. If it exceeds, enter the next round of search.

[0177] The local search operation is repeated until a preset stopping criterion is met.

[0178] For example, local search operation is a key step in optimizing high-quality paths, and the variable neighborhood search strategy can effectively improve the search efficiency and solution quality. This strategy designs multiple neighborhood structures and dynamically switches them during the search process to increase the probability of jumping out of the local optimum.

[0179] First, the selected high-quality path is set as the current solution. For example, suppose the current solution is a path containing 10 nodes: ABCDEFGHIJ, with a total length of 100 kilometers.

[0180] Next, we design a variety of neighborhood structures and construct a neighborhood structure set. This scheme includes three neighborhood structures: 2-opt exchange structure, node insertion structure, and subpath reversal structure. The 2-opt exchange structure randomly selects two non-adjacent edges, disconnects them, and reconnects them to form a new path. The node insertion structure randomly selects a node and inserts it to another position in the path. The subpath reversal structure randomly selects a subpath and reverses its order.

[0181] The core of the variable neighborhood search strategy is to perform a local search operation on the current solution. First, randomly select nodes or edges in the path in the current neighborhood structure, and generate a new solution according to the rules of the current neighborhood structure. Taking the 2-opt exchange structure as an example, the edges AB and EF may be randomly selected, disconnected and reconnected to form a new path AEDCBFGHIJ. For the node insertion structure, the node D may be selected and inserted between G and H to form a new path ABCEFGDHIJ. For the subpath reversal structure, the subpath CDE may be selected and its order may be reversed to form a new path ABEDCFGHIJ.

[0182] Perform local optimization on the generated new solution to obtain the local optimal solution. Local optimization can use a simple greedy strategy, such as trying to exchange adjacent nodes and accepting the exchange if it can improve the path length. For example, for the path AEDCBFGHIJ after 2-opt exchange, it is possible to obtain a further optimized path AECDBFGHIJ by exchanging C and D.

[0183] Calculate the objective function value of the local optimal solution and compare it with the objective function value of the current solution. The objective function can be the total length of the path or a weighted sum of multiple factors. Assume that the length of the new optimized path is 98 kilometers, which is less than the 100 kilometers of the current solution.

[0184] Update the search state based on the comparison result. When the local optimal solution is better than the current solution, set the local optimal solution as the new current solution and reset the neighborhood structure index to the initial value. In this example, the new path AECDBFGHIJ becomes the current solution and the neighborhood structure starts again from 2-opt exchange. If the local optimal solution fails to improve the current solution, increase the neighborhood structure index and switch to the next neighborhood structure, such as switching from 2-opt exchange to node insertion structure.

[0185] Determine whether the neighborhood structure index exceeds the size of the neighborhood structure set. If not, return to the step of randomly generating a new solution and continue searching in the current neighborhood structure. If it exceeds, enter the next round of search and start again from the first neighborhood structure.

[0186] The local search operation is repeated until a preset stopping criterion is met. The stopping criterion may be reaching a maximum number of iterations, such as 1000 iterations, or there is no improvement after multiple consecutive iterations, such as the path length improvement does not exceed 0.1% after 50 consecutive iterations.

[0187] An important feature of the variable neighborhood search strategy is its adaptability. For example, the probability of using each neighborhood structure can be dynamically adjusted according to the frequency with which it produces improvements. Assuming that in 100 iterations, 2-opt exchange, node insertion, and subpath reversal produce 20, 15, and 10 improvements, respectively, then in the next round of search, the probability of using these three structures can be adjusted to 44.4%, 33.3%, and 22.2%.

[0188] In addition, in order to increase the diversity of the search, a perturbation mechanism can be introduced. For example, every certain number of iterations (such as every 50 times), a large random perturbation is performed on the current optimal solution, such as randomly exchanging the positions of multiple nodes in the path. This helps the algorithm escape the local optimal trap and explore a wider solution space.

[0189] The variable neighborhood search strategy adopted in this paper has significant advantages in path optimization problems. First, the design of multiple neighborhood structures increases the diversity of the search and improves the probability of finding the global optimal solution. Second, the mechanism of dynamically switching the neighborhood structure helps to jump out of the local optimum and avoid being stuck in search stagnation. Furthermore, the adaptive adjustment mechanism can dynamically adjust the strategy according to the feedback information of the search process to improve the efficiency of the algorithm. Finally, this method is easy to implement and has relatively low computational complexity, which is suitable for large-scale practical problems. Through this method, high-quality path optimization solutions can be obtained in a relatively short time, providing strong decision-making support for traffic scheduling, logistics distribution and other fields, and ultimately achieving reduced operating costs and improved service quality.

[0190] In an optional embodiment,

[0191] A multi-objective optimization framework is constructed, wherein the multi-objective optimization framework includes a passenger satisfaction objective function, an operating cost objective function, and an environmental impact objective function; the updated global pheromone distribution is used as prior information, and the steps of solving the Pareto optimal solution set using the improved NSGA-II algorithm under the multi-objective optimization framework include:

[0192] The passenger satisfaction objective function takes into account the passenger waiting time; the operating cost objective function takes into account the vehicle's driving distance and energy consumption, and the energy consumption calculation involves the vehicle type, load and road conditions; the environmental impact objective function takes into account the vehicle emission factor and energy consumption, and the emission factor is dynamically determined according to the vehicle type and the use of new energy;

[0193] Establish constraints, including vehicle capacity constraints, time window constraints, and power constraints;

[0194] Creating and maintaining a global pheromone distribution matrix, wherein the global pheromone distribution matrix represents path attractiveness and is dynamically adjusted through pheromone update rules;

[0195] The improved NSGA-II algorithm is used to solve the Pareto optimal solution set. The improved NSGA-II algorithm includes: generating an initial population based on the global pheromone distribution, and using a priority coding method to represent individuals; performing adaptive crossover and mutation operations, and the crossover and mutation probabilities are dynamically adjusted according to the individual fitness; using an improved crowding calculation method based on k-nearest neighbors to evaluate the diversity of solutions; performing simulated annealing local search on elite individuals to improve the quality of solutions; performing non-dominated sorting and environmental selection to generate a new generation of populations; updating the global pheromone distribution and strengthening the path corresponding to the high-quality solution; and using the hypervolume indicator to evaluate the quality and convergence of the Pareto frontier.

[0196] Repeat the iteration until the termination condition is met;

[0197] The density of the final Pareto solution set is controlled and the boundary is preserved, and the Pareto optimal solution set is output.

[0198] For example, the construction of a multi-objective optimization framework is the key to solving complex vehicle path planning problems. The framework comprehensively considers the three objectives of passenger satisfaction, operating costs and environmental impact, and solves the Pareto optimal solution set through the improved NSGA-II algorithm to achieve a balance between multiple objectives.

[0199] First, three objective functions are constructed. The passenger satisfaction objective function mainly considers the passenger waiting time. For example, the waiting time of each passenger can be weighted and summed, and the weight can be set according to the passenger type (such as ordinary passengers, VIP passengers). Assuming there are 10 passengers and the average waiting time is 15 minutes, the satisfaction score may be 150. The operating cost objective function considers the vehicle's driving distance and energy consumption. The energy consumption calculation needs to consider the vehicle type, load and road conditions. For example, an electric car with a full load may consume 20 kWh of energy when traveling 100 kilometers on urban roads. The environmental impact objective function considers the vehicle emission factor and energy consumption. The emission factor is dynamically determined according to the vehicle type and the use of new energy. For example, a hybrid vehicle may emit 15 kg of carbon dioxide per 100 kilometers under the same conditions.

[0200] Next, set up constraints. Vehicle capacity constraints ensure that each vehicle does not carry more than its maximum capacity, such as a medium-sized bus can carry a maximum of 50 passengers. Time window constraints ensure that the service is completed within a specified time range, such as the service time of a station must be between 9:00 and 11:00 am. Power constraints ensure that electric vehicles have enough power during the journey, such as requiring the remaining power to be no less than 20% when arriving at the charging station.

[0201] Creating and maintaining a global pheromone distribution matrix is ​​an important part of the algorithm. This matrix represents the path attractiveness and can be represented by a two-dimensional array. Initially, the initial value can be set according to the inverse of the path distance, and then dynamically adjusted through the pheromone update rule. For example, if path ABC performs well in multiple iterations, increase the pheromone concentration on this path.

[0202] The improved NSGA-II algorithm is the core of solving the Pareto optimal solution set. First, the initial population is generated based on the global pheromone distribution. The roulette wheel selection method can be used to probabilistically select paths according to the pheromone concentration to construct the initial solution. The priority encoding method is used to represent individuals, such as [3,1,4,2,5] to visit 5 nodes in this order.

[0203] During the evolution process, adaptive crossover and mutation operations are performed. The crossover operation can use sequential crossover (OX) or partial mapping crossover (PMX), and the mutation operation can use exchange mutation or insertion mutation. The crossover and mutation probabilities are dynamically adjusted according to the individual fitness. For example, individuals with lower fitness may have a higher mutation probability to increase diversity.

[0204] When evaluating the diversity of solutions, an improved crowding calculation method based on k-nearest neighbors is used. Traditional crowding calculation may not work well in high-dimensional target space, while the k-nearest neighbor method can better measure the distribution of solutions. For example, k=5 can be selected to calculate the average distance between each solution and its five nearest neighbors as the crowding measure.

[0205] Perform simulated annealing local search on elite individuals to improve the quality of the solution. In each generation of evolution, a certain proportion (such as the top 10%) of elite individuals are selected for local search. The simulated annealing algorithm allows slightly worse solutions to be accepted with a certain probability, which helps to escape the local optimum. For example, the temperature starts at 100 and decreases by 1% each iteration, and the search stops when the temperature drops to 1.

[0206] Non-dominated sorting and environmental selection are key steps in generating a new generation of population. Non-dominated sorting divides the population into multiple levels, and the decomposition of the first level is the current Pareto frontier. Environmental selection selects the next generation of individuals based on the non-dominated level and crowding degree to ensure the quality and diversity of the population.

[0207] When updating the global pheromone distribution, strengthen the paths corresponding to high-quality solutions. You can set a pheromone enhancement coefficient, such as 1.2, and multiply the pheromone concentration of the path corresponding to the solution on the Pareto front by this coefficient. At the same time, to prevent excessive accumulation of pheromones, you can set a global evaporation coefficient, such as 0.95, and multiply all pheromone concentrations by this coefficient after each update.

[0208] Use the hypervolume metric to evaluate the quality and convergence of the Pareto front. The hypervolume represents the volume of the multidimensional space between the Pareto front and a reference point. As the iterations proceed, the hypervolume metric should gradually increase and stabilize. For example, a threshold can be set such that the algorithm is considered to have converged when the hypervolume growth rate is less than 0.1% for 50 consecutive generations.

[0209] Finally, the Pareto solution set is density controlled and boundary preserved. Density control can be achieved through clustering methods, such as k-means clustering, to select representative solutions for each cluster. Boundary preservation ensures that extreme solutions in the target space are retained to show the trade-offs between different targets.

[0210] The multi-objective optimization method adopted in this paper has many advantages. First, it can simultaneously consider multiple conflicting objectives, such as passenger satisfaction, operating costs, and environmental impact, and provide decision makers with a series of trade-offs. Secondly, the improved NSGA-II algorithm enhances the search capability through adaptive operations and local search, improving the quality and diversity of solutions. Furthermore, the introduction of global pheromone distribution utilizes historical search information and accelerates the convergence of the algorithm. Finally, the diversity preservation mechanism and density control ensure the extensiveness and representativeness of the final solution set. This method can not only provide optimization solutions for urban public transportation systems, but can also be extended to logistics distribution, emergency rescue and other fields, providing strong support for complex decision-making problems.

[0211] In an optional embodiment,

[0212] The steps of using deep reinforcement learning algorithm to select the optimal path from the candidate path set and perform vehicle allocation and scheduling include:

[0213] Collect vehicle state vectors, passenger demand matrices, traffic condition matrices, candidate path sets and current time information to construct initial state information; construct a state space based on the initial state information, and design an action space representing the path assigned to the vehicle;

[0214] Design a path feature extraction module, including a multi-layer perceptron, to extract the total distance, estimated completion time, number of passengers, passenger satisfaction estimate, and path complexity features of candidate paths;

[0215] Using the multi-layer perceptron to perform feature encoding on each candidate path to obtain a path feature encoding vector;

[0216] The attention mechanism is used to aggregate the path feature encoding vectors to generate path set encoding;

[0217] Connect the path set encoding with the initial state information to form an enhanced state representation;

[0218] Use historical route data to pre-train the route feature extraction module through self-supervised learning;

[0219] Construct a deep Q network, including an input layer, multiple hidden layers, and an output layer, where the input layer receives the enhanced state representation and the output layer corresponds to the size of the action space;

[0220] Initialize the experience replay buffer to store state transition samples;

[0221] Adopting the priority experience replay strategy, the sample priority is calculated according to the temporal difference error, and sampling is performed based on the priority;

[0222] Use a dual deep Q network structure, including an online network and a target network, where the online network is used for action selection and the target network is used to calculate the target Q value;

[0223] The loss function is defined based on mean square error with importance sampling weights;

[0224] A dynamic ε-greedy strategy is used for exploration and exploitation, and the ε value decays with the number of training steps;

[0225] The path feature extraction module and the main Q network parameters are optimized simultaneously through the back propagation algorithm;

[0226] Update the target network parameters regularly to keep them consistent with the online network parameters;

[0227] Input the current enhanced state representation into the trained deep Q network and calculate the Q value of each path in the candidate path set; select the path with the highest Q value from the candidate path set as the optimal path, and assign the optimal path to the corresponding vehicle;

[0228] Execute the selected optimal path, observe environmental changes, and obtain new status information and rewards;

[0229] The new state information, executed actions, rewards obtained, and next state information are stored in the experience replay buffer as state transition samples;

[0230] Repeat the steps from constructing the initial state information to storing the state transition samples to optimize the deep Q network.

[0231] For example, deep reinforcement learning algorithms have shown great potential in solving complex vehicle path selection and scheduling problems. This method continuously optimizes decision-making strategies by learning the dynamic characteristics of the environment, and ultimately achieves efficient path planning and vehicle allocation.

[0232] First of all, constructing the initial state information is the basis of the algorithm. The vehicle state vector contains information such as the current position, remaining power, and passenger capacity of each vehicle. For example, an electric car may be at coordinates (100, 200), with 80% remaining power and currently carrying 3 passengers. The passenger demand matrix reflects the passenger flow between locations, such as 10 passengers demanding to go from A to B. The traffic condition matrix represents the degree of road congestion, which can be represented by a value between 0 and 1, where 0 means unobstructed and 1 means completely blocked. The candidate path set contains pre-generated feasible paths, each of which consists of a series of nodes. The current time information is used to consider time-related factors, such as peak traffic conditions.

[0233] A state space is constructed based on the initial state information, including all possible system states. The action space is designed to assign paths to all possible combinations of vehicles. For example, if there are 5 vehicles and 10 candidate paths, the action space size is 50, representing all possible vehicle-path assignments.

[0234] The path feature extraction module is one of the core components of the algorithm. It adopts a multi-layer perceptron structure and can be designed as an input layer, two hidden layers, and an output layer. The input layer receives the original features of the path, such as the total distance (in kilometers), the estimated completion time (in minutes), the number of passengers, the estimated passenger satisfaction (1-5 points), and the path complexity (such as the number of turns). The hidden layer uses the ReLU activation function, and the output layer uses the linear activation function. For example, the original features of a path may be [20,45,5,4.2,8], which means that the total length is 20 kilometers, it is expected to be completed in 45 minutes, it carries 5 passengers, the estimated satisfaction is 4.2 points, and it contains 8 turns.

[0235] The multi-layer perceptron encodes the features of each candidate path and converts the original features into a vector representation of a fixed dimension. Assuming the output dimension is 64, each path will be encoded into a 64-dimensional vector.

[0236] The introduction of the attention mechanism allows the model to focus on more important path features. The self-attention mechanism can be used to assign a weight to each path, and then perform weighted summation to obtain the overall representation of the path set. In this way, the model can dynamically adjust the attention to different path features according to the current state.

[0237] The path set encoding is connected with the initial state information to form an enhanced state representation. This representation method combines path characteristics and global state, providing a rich information basis for subsequent decision-making.

[0238] Pre-training the route feature extraction module using historical route data can improve the initial performance of the model. A self-supervised learning task can be designed, such as predicting the completion time of a route or passenger satisfaction, and training is performed using mean squared error as the loss function.

[0239] The construction of a deep Q network is the core of the algorithm. The input layer receives the enhanced state representation, and 3-4 hidden layers can be designed, each layer uses a different number of neurons (such as 512, 256, 128), and uses the ReLU activation function. The number of neurons in the output layer is the same as the size of the action space, and a linear activation function is used.

[0240] The size of the experience replay buffer needs to be set according to the problem scale. For example, the last 100,000 state transition samples can be stored. Each sample contains the current state, the action performed, the reward obtained, and the next state.

[0241] The priority experience replay strategy assigns a priority to each sample based on the temporal difference error. The temporal difference error reflects the gap between the predicted value and the actual value. The sample with a larger error is given a higher priority. When sampling, high-priority samples have a greater probability of being selected, but a certain probability is still retained to select low-priority samples to maintain diversity.

[0242] The dual deep Q network structure includes an online network and a target network, both of which have the same structure but independent parameters. The online network is used to select actions, and the target network is used to calculate the target Q value. This separation helps reduce the overestimation problem.

[0243] The loss function uses mean square error with importance sampling weights. Importance sampling weights are used to correct the bias introduced by priority sampling and ensure the stability of the learning process.

[0244] The dynamic ε-greedy strategy sets a higher ε value (such as 0.9) at the beginning of training, and gradually decays it to a lower value (such as 0.1) with the number of training steps. This encourages exploration in the early stage and makes more use of the learned knowledge in the later stage.

[0245] The path feature extraction module and the main Q network parameters are optimized simultaneously through the back propagation algorithm. The Adam optimizer can be used, the learning rate is set to 0.001, and the parameters are updated after each batch of samples are processed.

[0246] Update the target network parameters regularly, such as copying the parameters of the online network to the target network after every 1,000 training steps. This soft update mechanism helps improve the stability of training.

[0247] In practical applications, the current enhanced state representation is input into the trained deep Q network to obtain the Q value of each candidate path. The path with the highest Q value is selected as the optimal path and assigned to the corresponding vehicle. For example, if the Q value of path A is 85, which is higher than other paths, path A is selected and assigned to the vehicle of the current decision.

[0248] After executing the selected optimal path, observe the changes in the environment and obtain new status information and rewards. The reward can be designed as a weighted sum of multiple factors, such as the number of completed tasks, passenger satisfaction, energy consumption, etc.

[0249] The new state transition samples are stored in the experience replay buffer for subsequent learning. By repeating this process, the deep Q network will gradually optimize and learn better decision-making strategies.

[0250] The deep reinforcement learning-based method adopted in this paper has many advantages. First, it can handle high-dimensional and complex state spaces and adapt to dynamically changing traffic environments. Secondly, through experience replay and priority sampling, the sample utilization efficiency and learning effect are improved. Furthermore, the introduction of the path feature extraction module enables the model to better understand the intrinsic characteristics of the path. In addition, the application of the attention mechanism enhances the model's ability to capture key information. Finally, the dual Q network and soft update mechanism improve the stability of training. This method can not only achieve efficient vehicle scheduling in complex urban traffic networks, but also has strong generalization ability, can adapt to path planning problems of different scales and types, and provides strong support for the development of intelligent transportation systems.

[0251] Figure 2 FIG. 1 is a schematic diagram of the structure of a fixed-point area path optimization system for a responsive public transportation service according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0252] The first unit is used to construct a responsive public transportation service system model, including a passenger demand module, a vehicle status module and a road network information module, and to determine a mathematical expression of a fixed-point area path optimization problem based on the responsive public transportation service system model;

[0253] The second unit is used to design a path optimization solver using an improved ant colony algorithm, wherein the improved ant colony algorithm includes an adaptive pheromone update mechanism and a dynamic candidate set strategy; construct an initial path set using the path optimization solver, and evaluate the paths in the initial path set according to a preset evaluation index; perform a local search operation, locally optimize the evaluated paths, and obtain an optimized path set; and update the global pheromone distribution based on the optimized path set;

[0254] The third unit is used to construct a multi-objective optimization framework, which includes a passenger satisfaction objective function, an operating cost objective function, and an environmental impact objective function; using the updated global pheromone distribution as prior information, and using the improved NSGA-II algorithm to solve the Pareto optimal solution set under the multi-objective optimization framework; predicting short-term passenger demand based on historical data and real-time information; generating a candidate path set based on short-term passenger demand, the Pareto optimal solution set, and the updated global pheromone distribution; using a deep reinforcement learning algorithm to select the optimal path from the candidate path set, and perform vehicle allocation and scheduling; updating the global pheromone distribution based on actual operation results, and feeding back the updated information to the path optimization solver, the multi-objective optimization framework, and the dynamic path planning algorithm to continuously optimize the responsive bus service system model.

[0255] According to a third aspect of the embodiments of the present invention,

[0256] An electronic device is provided, comprising:

[0257] processor;

[0258] a memory for storing processor-executable instructions;

[0259] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0260] A fourth aspect of the embodiments of the present invention is:

[0261] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0262] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0263] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for optimizing the fixed-point area path of a responsive bus service, characterized in that: include: Constructing a responsive public transportation service system model, including a passenger demand module, a vehicle status module and a road network information module, and determining a mathematical expression for a fixed-point area path optimization problem based on the responsive public transportation service system model; Designing a path optimization solver using an improved ant colony algorithm, wherein the improved ant colony algorithm includes an adaptive pheromone update mechanism and a dynamic candidate set strategy; Use the path optimization solver to build an initial path set, and evaluate the paths in the initial path set according to the preset evaluation index; perform local search operations, locally optimize the evaluated paths, and obtain an optimized path set; based on the optimized path set, update the global pheromone distribution; A multi-objective optimization framework is constructed, wherein the multi-objective optimization framework includes a passenger satisfaction objective function, an operating cost objective function, and an environmental impact objective function; the updated global pheromone distribution is used as prior information, and the improved NSGA-II algorithm is used to solve the Pareto optimal solution set under the multi-objective optimization framework; short-term passenger demand is predicted based on historical data and real-time information; a candidate path set is generated based on short-term passenger demand, the Pareto optimal solution set, and the updated global pheromone distribution; a deep reinforcement learning algorithm is used to select the optimal path from the candidate path set, and vehicle allocation and scheduling are performed; According to the actual operation results, the global pheromone distribution is updated, and the updated information is fed back to the path optimization solver, multi-objective optimization framework and dynamic path planning algorithm to continuously optimize the responsive bus service system model; A responsive public transportation service system model is constructed, including a passenger demand module, a vehicle status module and a road network information module. Based on the responsive public transportation service system model, the steps of determining a mathematical expression of a fixed-point area path optimization problem include: The passenger demand module uses a spatiotemporal distribution model to describe passenger demand, and uses a time series prediction model based on a long short-term memory network to predict demand; the vehicle status module uses a multi-dimensional state vector to represent the vehicle status, and uses a Kalman filter algorithm to predict the future status of the vehicle; the road network information module uses a weighted directed graph to represent the road network, and uses a traffic flow prediction model based on a graph convolutional network to update the road network status in real time; Generate passenger demand distribution data for the current and future periods based on the spatiotemporal distribution model of the passenger demand module and the time series prediction model of the long short-term memory network; use the passenger demand distribution data as input parameters for the fixed-point area route optimization problem; Based on the multi-dimensional state vector and Kalman filter algorithm of the vehicle state module, the real-time state information of the vehicle is calculated and updated; the real-time state information of the vehicle is used as the constraint condition of the fixed-point area path optimization problem; The traffic flow prediction model based on the weighted directed graph and graph convolutional network of the road network information module generates real-time road network status data; the real-time road network status data is used to calculate the driving distance and time window constraints of the fixed-point area path optimization problem; Based on the constructed responsive public transportation service system model, the mathematical expression of the fixed-point area path optimization problem is determined by using passenger demand distribution data, real-time vehicle status information and real-time road network status data. The mathematical expression includes an objective function and constraints. The objective function considers travel distance, vehicle energy consumption and passenger satisfaction. The constraints include vehicle capacity constraints, time window constraints and power constraints. The steps of using deep reinforcement learning algorithm to select the optimal path from the candidate path set and perform vehicle allocation and scheduling include: Collecting vehicle state vectors, passenger demand matrices, traffic condition matrices, candidate path sets and current time information to construct initial state information; constructing a state space based on the initial state information, and designing an action space representing the path assigned to the vehicle; Design a path feature extraction module, including a multi-layer perceptron, to extract the total distance, estimated completion time, number of passengers, passenger satisfaction estimate, and path complexity features of candidate paths; Using the multi-layer perceptron to perform feature encoding on each candidate path to obtain a path feature encoding vector; The attention mechanism is used to aggregate the path feature encoding vectors to generate path set encoding; Connect the path set encoding with the initial state information to form an enhanced state representation; Use historical route data to pre-train the route feature extraction module through self-supervised learning; Construct a deep Q network, including an input layer, multiple hidden layers, and an output layer, where the input layer receives the enhanced state representation and the output layer corresponds to the size of the action space; Initialize the experience replay buffer to store state transition samples; Adopting the priority experience replay strategy, the sample priority is calculated according to the temporal difference error, and sampling is performed based on the priority; Use a dual deep Q network structure, including an online network and a target network, where the online network is used for action selection and the target network is used to calculate the target Q value; The loss function is defined based on mean square error with importance sampling weights; A dynamic ε-greedy strategy is used for exploration and exploitation, and the ε value decays with the number of training steps; The path feature extraction module and the main Q network parameters are optimized simultaneously through the back propagation algorithm; Update the target network parameters regularly to keep them consistent with the online network parameters; Input the current enhanced state representation into the trained deep Q network and calculate the Q value of each path in the candidate path set; select the path with the highest Q value from the candidate path set as the optimal path, and assign the optimal path to the corresponding vehicle; Execute the selected optimal path, observe environmental changes, and obtain new status information and rewards; The new state information, executed actions, rewards obtained, and next state information are stored in the experience replay buffer as state transition samples; Repeat the steps from constructing the initial state information to storing the state transition samples to optimize the deep Q network.

2. The method according to claim 1, characterized in that Designing a path optimization solver using an improved ant colony algorithm, wherein the improved ant colony algorithm includes an adaptive pheromone update mechanism and a dynamic candidate set strategy; Use the path optimization solver to build an initial path set, and evaluate the paths in the initial path set according to the preset evaluation index; perform local search operations, locally optimize the evaluated paths, and obtain the optimized path set; Based on the optimized path set, the steps of updating the global pheromone distribution include: The adaptive pheromone update mechanism uses an adaptive factor that changes dynamically with the number of iterations to adjust the pheromone increment; the dynamic candidate set strategy uses a dynamic threshold that changes with the iteration process to control the size of the candidate set; Using a path optimization solver to construct an initial path set, and generating multiple candidate paths through state transition rules and roulette selection method; evaluating the candidate paths in the initial path set according to preset evaluation indicators, wherein the preset evaluation indicators include path length, time reliability, energy efficiency and passenger satisfaction; and selecting high-quality paths according to the evaluation results; Perform local search operations on the selected high-quality paths, and use variable neighborhood search strategies to design a variety of neighborhood structures, including 2-opt exchange, node insertion, and subpath reversal; obtain the optimized path set by changing the neighborhood structure; Based on the optimized path set, the global pheromone distribution is updated, the idea of ​​elite ant colony system is adopted, and the pheromone increments of the current iterative optimal path and the global optimal path are considered at the same time; the idea of ​​maximum and minimum ant colony system is introduced to limit the pheromone concentration within the maximum and minimum range of dynamic adjustment; The initial path set construction, path evaluation, local search optimization and global pheromone distribution update steps are repeated until the preset convergence condition is met or the maximum number of iterations is reached.

3. The method according to claim 2, characterized in that The probability of selection in the roulette wheel selection method is defined as: ； Among them, each parameter represents the meaning: p ij (t): The probability that the ant chooses to go to node j from the current node i at time t; τ ij (t): pheromone concentration on the path connecting node i and node j at time t; η ij (t): heuristic information from node i to node j; α: pheromone importance factor; β: heuristic information importance factor; l: index variable, used to represent all possible next nodes of node i; τ il (t): pheromone concentration on the path connecting node i and node l at time t; η il (t): heuristic information from node i to node l; N i (t): At time t, the candidate set of node i, that is, the set of next nodes that the ant can currently choose; j∈N i (t): indicates that node j is an element in the candidate set of node i; The numerator of the formula [τ ij (t)] α *[η ij (t)] β represents the total attraction of the path from node i to node j; The denominator of the formula Σ[τ il (t)] α *[η il (t)] β is a normalization factor, which is the sum of the total attractiveness of all possible choices. By dividing by this sum, we get a probability distribution that ensures that the sum of the probabilities of all possible choices is 1; By adjusting the values ​​of α and β, the balance between utilizing pheromones and heuristic information can be controlled.

4. The method according to claim 2, characterized in that: A local search operation is performed on the selected high-quality paths, and a variable neighborhood search strategy is used to design a variety of neighborhood structures, including 2-opt exchange, node insertion, and subpath reversal. The steps of obtaining the optimized path set by changing the neighborhood structure include: The selected high-quality path is used as the current solution; Design a variety of neighborhood structures and build a collection of neighborhood structures, including 2-opt exchange structure, node insertion structure and subpath inversion structure; A local search operation is performed on the current solution, wherein the local search operation adopts a variable neighborhood search strategy, including: Randomly select nodes or edges in the path in the current neighborhood structure, and generate new solutions according to the rules of the current neighborhood structure. The 2-opt exchange structure randomly selects two non-adjacent edges to disconnect and reconnect to form a new path, the node insertion structure randomly selects a node and inserts it to another position in the path, and the subpath reversal structure randomly selects a subpath and reverses its order. Performing local optimization on the rule-generated new solution of the current neighborhood structure to obtain a local optimal solution; Calculate the objective function value of the local optimal solution and compare it with the objective function value of the current solution; update the search state according to the comparison result. When the local optimal solution is better than the current solution, set the local optimal solution as the new current solution and reset the neighborhood structure index to the initial value; otherwise, increase the neighborhood structure index and switch to the next neighborhood structure; Determine whether the neighborhood structure index exceeds the size of the neighborhood structure set. If not, return to the step of randomly generating a new solution in the current neighborhood structure and continue searching in the current neighborhood structure. If it exceeds, enter the next round of search. The local search operation is repeated until a preset stopping criterion is met.

5. The method according to claim 1, characterized in that A multi-objective optimization framework is constructed, wherein the multi-objective optimization framework includes a passenger satisfaction objective function, an operating cost objective function, and an environmental impact objective function; the updated global pheromone distribution is used as prior information, and the steps of solving the Pareto optimal solution set using the improved NSGA-II algorithm under the multi-objective optimization framework include: The passenger satisfaction objective function takes into account the passenger waiting time; the operating cost objective function takes into account the vehicle's driving distance and energy consumption, and the energy consumption calculation involves the vehicle type, load and road conditions; the environmental impact objective function takes into account the vehicle emission factor and energy consumption, and the emission factor is dynamically determined according to the vehicle type and the use of new energy; Establish constraints, including vehicle capacity constraints, time window constraints, and power constraints; Creating and maintaining a global pheromone distribution matrix, wherein the global pheromone distribution matrix represents path attractiveness and is dynamically adjusted through pheromone update rules; The improved NSGA-II algorithm is used to solve the Pareto optimal solution set. The improved NSGA-II algorithm includes: generating an initial population based on the global pheromone distribution, and using a priority coding method to represent individuals; performing adaptive crossover and mutation operations, and the crossover and mutation probabilities are dynamically adjusted according to the individual fitness; using an improved crowding calculation method based on k-nearest neighbors to evaluate the diversity of solutions; performing simulated annealing local search on elite individuals to improve the quality of solutions; performing non-dominated sorting and environmental selection to generate a new generation of populations; updating the global pheromone distribution and strengthening the path corresponding to the high-quality solution; and using the hypervolume indicator to evaluate the quality and convergence of the Pareto frontier. Repeat the iteration until the termination condition is met; The density of the final Pareto solution set is controlled and the boundary is preserved, and the Pareto optimal solution set is output.

6. A fixed-point area route optimization system for responsive public transportation services, used to implement the method described in any one of claims 1 to 5, characterized in that: include: The first unit is used to construct a responsive public transportation service system model, including a passenger demand module, a vehicle status module and a road network information module, and to determine a mathematical expression of a fixed-point area path optimization problem based on the responsive public transportation service system model; The second unit is used to design a path optimization solver using an improved ant colony algorithm, wherein the improved ant colony algorithm includes an adaptive pheromone update mechanism and a dynamic candidate set strategy; Use the path optimization solver to build an initial path set, and evaluate the paths in the initial path set according to the preset evaluation index; perform local search operations, locally optimize the evaluated paths, and obtain an optimized path set; based on the optimized path set, update the global pheromone distribution; The third unit is used to construct a multi-objective optimization framework, which includes a passenger satisfaction objective function, an operating cost objective function, and an environmental impact objective function; using the updated global pheromone distribution as prior information, and using the improved NSGA-II algorithm to solve the Pareto optimal solution set under the multi-objective optimization framework; predicting short-term passenger demand based on historical data and real-time information; generating a candidate path set based on short-term passenger demand, the Pareto optimal solution set, and the updated global pheromone distribution; using a deep reinforcement learning algorithm to select the optimal path from the candidate path set, and perform vehicle allocation and scheduling; According to the actual operation results, the global pheromone distribution is updated, and the updated information is fed back to the path optimization solver, multi-objective optimization framework and dynamic path planning algorithm to continuously optimize the responsive bus service system model.

7. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 5.

8. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 5 is implemented.

Images