A bus operation data driven intelligent scheduling system

Abstract

The application discloses a kind of public transport operation data-driven intelligent scheduling system, belong to computer technology field.System includes data acquisition module, situation awareness engine, collaborative decision engine and dispatching instruction engine.Data acquisition module collects vehicle position, passenger flow, road network topology, real-time traffic and operation plan data;Situation awareness engine contains graph neural network unit and time series prediction unit, based on road network topology and historical time series data to predict road congestion index.Collaborative decision engine models each bus as an agent, uses centralized and distributed hybrid communication architecture and Actor-Critic deep reinforcement learning algorithm, generates multi-vehicle collaborative scheduling decision;Dispatching instruction engine converts the decision into structured instruction message and issues to vehicle terminal, receives execution feedback closed-loop update;The application realizes road network situation active prediction, multi-agent collaborative decision and instruction automatic execution, improves bus punctuality and operation efficiency, reduces manual intervention.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, specifically relating to an intelligent scheduling system driven by public transportation operation data. Background Technology

[0002] Currently, most cities in my country still use the traditional operation mode of bus dispatching, which is based on fixed timetables and supplemented by human experience. The dispatch center arranges the departure plan according to the pre-compiled operating timetable and tracks the vehicle location through manual monitoring or GPS positioning. When abnormal situations such as road congestion, vehicle failure or sudden changes in passenger flow occur, dispatchers make manual adjustments to the operation.

[0003] Some existing solutions attempt to assist scheduling decisions by collecting GPS location data and passenger flow data, or optimize departure timetables based on historical passenger flow data. Other solutions use heuristic algorithms such as genetic algorithms to build optimization models for scheduling solutions. However, these solutions are still limited to static planning or post-event response in terms of technical approach, and fail to achieve proactive prediction of road network conditions and collaborative scheduling among vehicles.

[0004] The existing technology has the following main shortcomings: First, it lacks the ability to make forward-looking judgments on the road network operation status and cannot actively intervene before congestion or anomalies occur; Second, scheduling decisions are highly dependent on manual labor, and it takes a long time from the discovery of anomalies to the completion of adjustments. During this period, congestion often spreads and causes a chain of delays; Third, it is difficult to dynamically adjust strategies according to real-time road conditions after each vehicle departs, and multi-vehicle coordination cannot be achieved. Summary of the Invention

[0005] To address the technical problems in existing bus dispatching systems, such as lack of forward-looking road network situation prediction capabilities, inability to achieve multi-vehicle collaborative decision-making, and low on-time operation rates due to delayed dispatching response, this application provides a bus operation data-driven intelligent scheduling system, employing the following technical solution:

[0006] A data-driven intelligent scheduling system for public transportation operations includes:

[0007] The data acquisition module is used to collect multi-source operational data, including vehicle location data, passenger flow data, road network topology data, real-time traffic data, and operation plan data.

[0008] The situational awareness engine, connected to the data acquisition module, receives the structured data stream output by the data acquisition module. The situational awareness engine includes a graph neural network unit and a temporal prediction unit. The graph neural network unit performs spatial dependency modeling based on the road network topology, enabling each road segment node to perceive the traffic status of adjacent road segments. The temporal prediction unit predicts the travel time of each road segment within a preset time period based on a temporal convolutional network. The situational awareness engine outputs road network situational analysis results that include the congestion index prediction of each road segment.

[0009] The collaborative decision-making engine, connected to the situational awareness engine, models each bus as an agent, constructs a multi-agent collaborative decision-making model, and generates scheduling decisions for each vehicle based on deep reinforcement learning algorithms. The collaborative decision-making engine includes an agent modeling unit, a communication mechanism unit, a policy network unit, and a reward evaluation unit. The agent modeling unit creates agent instances with independent state and action spaces for each bus. The communication mechanism unit adopts a hybrid communication architecture combining centralized and distributed approaches to achieve information transmission between agents and collaborative information interaction based on a value iteration network. The policy network unit outputs the action policy of each agent based on an Actor-Critic architecture. The reward evaluation unit evaluates the decision-making effect based on a multi-objective reward function.

[0010] The scheduling command engine, connected to the collaborative decision engine, is used to convert scheduling decisions into structured command messages and send them to the corresponding vehicle terminal through the communication transmission layer. The scheduling command engine includes a command generation unit, a command sending unit, and an execution feedback unit. The command generation unit converts the collaborative decision results into structured command messages, the command sending unit sends the commands to the target vehicle through the mobile communication network, and the execution feedback unit receives the execution results and updates the system status.

[0011] In summary, this application includes at least one of the following beneficial technical effects:

[0012] 1. This application collects vehicle location data, road network topology data, and real-time traffic data through a data acquisition module. The graph neural network unit in the situational awareness engine performs spatial dependency modeling based on the road network topology, enabling each road segment node to perceive the traffic status of adjacent road segments through a message passing mechanism. At the same time, the temporal prediction unit predicts the travel time of each road segment within a preset time period based on a temporal convolutional network, generating a road network situational analysis result that includes the congestion index prediction of each road segment. This allows the dispatching system to perceive the formation and spread trend of road network congestion in advance, transforming dispatching decisions from reactive post-event response to proactive pre-event prediction.

[0013] 2. This application models each bus as an intelligent agent with an independent state space and action space through a collaborative decision engine, constructs a multi-agent collaborative decision model, and automatically generates scheduling decisions for each vehicle based on a deep reinforcement learning algorithm. At the same time, the scheduling command engine automatically converts the scheduling decisions into structured command messages and sends them to the corresponding on-board terminal through the communication transmission layer. The execution feedback unit receives the execution results and updates the system status, realizing an automated closed loop from situational awareness to scheduling decision and then to command execution. The scheduling plan can be generated and issued in the early stage of congestion without manual intervention.

[0014] 3. This application adopts a hybrid communication architecture that combines centralized and distributed communication in the communication mechanism unit of the collaborative decision engine. Each agent shares local state and intent information through point-to-point communication. The centralized component calculates the global value function and the collaborative factors of each agent based on the value iteration network and broadcasts them to all agents. The distributed component enables each agent to selectively receive and process neighbor information through an attention mechanism, realizing collaborative information interaction between agents. This allows each vehicle to dynamically adjust its driving strategy according to real-time road conditions and the status of neighboring vehicles, achieving multi-vehicle collaborative scheduling. Attached Figure Description

[0015] Figure 1 This is a flowchart of the overall closed-loop operation of the system of this invention.

[0016] Figure 2 This is an internal flowchart of the situational awareness engine of the present invention.

[0017] Figure 3 This is a flowchart of the multi-agent collaborative decision-making process of the present invention.

[0018] Figure 4 This is a flowchart of the three-level communication degradation process of the present invention.

[0019] Figure 5 This is a flowchart of the main state transition of the system of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, please refer to... Figures 1 to 5 The present invention will be further described in detail below with reference to specific embodiments.

[0021] Example 1

[0022] This invention provides a data-driven intelligent scheduling system for public transportation operations, which includes four core components: a data acquisition module, a situational awareness engine, a collaborative decision-making engine, and a scheduling command engine.

[0023] The data acquisition module performs unified collection and preprocessing of multi-source operational data. The types of data collected include vehicle location data, passenger flow data, road network topology data, real-time traffic data, and operation plan data.

[0024] This module connects to external data sources through a multi-source heterogeneous data access interface. External data sources include GPS data servers, public transport IC card swiping systems, urban traffic information release platforms, electronic map service provider application interfaces, and public transport company operation and management information systems. The data acquisition module obtains public transport vehicle location information, passenger card swiping records, road network structure information, traffic congestion status information, and pre-defined operating timetable data.

[0025] The data cleaning unit performs noise reduction, outlier removal, and format standardization on the collected raw data to generate a structured data stream with a unified format.

[0026] The noise reduction process uses either moving average filtering or median filtering. The size of the filtering window ranges from 5 to 15 sampling points. Median filtering sorts the sampled values ​​within the window and takes the median as the output. Moving average filtering calculates the arithmetic mean of the sampled values ​​within the window as the output.

[0027] Outlier removal is based on the Laida criterion, also known as the 3σ criterion. The mean and standard deviation of the data within the current sliding window are calculated. When a sampled value deviates from the mean by more than three times the standard deviation, it is marked as an outlier and removed from the data stream. After removal, the missing positions are filled by the linear interpolation results of the previous and next valid sampled points.

[0028] The format standardization operation includes three steps: field name mapping, data type conversion, and timestamp format unification. Field name mapping maps fields with the same meaning but different names from different external data sources to standard field names. Data type conversion converts numeric types with different formats from various data sources to standard data types defined internally by the system. Timestamp format unification converts various time formats reported by various data sources to the UTC standard time format.

[0029] The situational awareness engine receives the structured data stream output by the data acquisition module, constructs a road network topology model based on a graph neural network, and predicts the travel time of each road segment within a preset time period based on a temporal convolutional network, ultimately generating road network situational analysis results. The situational awareness engine includes four sub-modules: a feature extraction unit, a graph neural network unit, a temporal prediction unit, and a fusion output unit.

[0030] The feature extraction unit performs preprocessing and spatiotemporal feature encoding on the collected multi-source data to generate road segment map features and time-series trend features. This unit first performs time synchronization processing on the multi-source data, aligning data from different data sources according to a unified time benchmark to eliminate asynchrony in the time dimension. Subsequently, it performs spatial association processing on various types of data, mapping vehicle positioning data to corresponding road network segments and associating real-time traffic data with road segment identifiers. The feature extraction unit uses embedding technology to map discrete road segment identifiers and continuous state values ​​to a unified high-dimensional vector space, generating a high-dimensional feature vector containing road segment attribute information and traffic state information.

[0031] The graph neural network unit employs a message-passing mechanism to iteratively propagate road segment status information among nodes in the road network topology graph. This allows each road segment node to perceive the traffic status of its adjacent road segments, thereby capturing the spatial patterns of congestion propagation in the road network. The unit models the road network as a directed graph structure, where nodes correspond to road segments, and edges between nodes correspond to topological connections between road segments. The message-passing process of the graph neural network unit follows the following mathematical description:

[0032] ;

[0033] in, Indicates the first After layer update, the road segment nodes The hidden state vector; Represents road segment nodes The set of adjacent nodes; Indicates from adjacent nodes Passed to the node The message vector is calculated using the following formula: ,in Let be the learnable message transformation matrix of the l-th layer.

[0034] The state update function is implemented using a gated recurrent unit (GRU), and its specific form is as follows:

[0035] ;

[0036] ;

[0037] ;

[0038] ;

[0039] in The weight matrix is ​​a learnable matrix. For element-wise multiplication, the state vector dimension is set to 128 dimensions, and the total number of message passing layers is set to 3.

[0040] The message aggregation function is implemented using attention-weighted aggregation, and its specific form is as follows: ;

[0041] Attention weight , The dimension of the key vector is 64. , Let be the linear transformation matrix of the query and key vectors at level l.

[0042] Through multi-layer message passing iteration, the traffic status of a single road segment can spread along the topological path to surrounding road segments, thereby capturing the spatial correlation of congestion propagation in the road network.

[0043] The training data for the graph neural network unit comes from historical travel time data and corresponding road network topology data collected by the data acquisition module. Each training sample contains the travel time series of each road segment within a time window and the corresponding road network topology adjacency matrix. The sample is labeled with the actual travel time value of each road segment in the next time step. The training dataset is divided into training, validation, and test sets in a 7:1.5:1.5 ratio. The training objective of the graph neural network unit is to minimize the mean squared error loss function between the predicted travel time and the actual travel time. The specific form of the loss function is as follows:

[0044] ;

[0045] in, It is the set of all road segments in the road network; For graph neural networks to analyze road segments The predicted travel time; For road section The actual travel time value; the Adam optimizer is used for parameter iterative updates during training, the initial learning rate is set to 0.001, the training rounds are set to 200 rounds, and training is terminated early when the validation set loss no longer decreases for 20 consecutive rounds.

[0046] The temporal prediction unit employs a multi-scale temporal convolutional kernel structure to simultaneously capture short-term fluctuations and long-term trends in road network conditions, predicting the travel time of each road segment within the next 15 to 30 minutes. The core structure of this unit uses a temporal convolutional network that combines causal convolution and dilated convolution. Causal convolution ensures the temporal causality of the prediction process, meaning that the output at the current moment depends only on the input at the previous moment, avoiding the leakage of future information. Dilated convolution expands the receptive field of the convolutional kernel without increasing the number of parameters by setting different dilation rates, enabling the network to capture dependencies at different time scales.

[0047] The multi-scale temporal convolutional kernel structure of the temporal prediction unit contains four parallel causal dilated convolutional branches, each corresponding to a different time scale, with baseline dilation rates of 1, 2, 4, and 8 respectively. Each branch employs a multi-layer stacked structure with L convolutional layers, where the actual dilation rate of the l-th layer is the baseline dilation rate multiplied by [missing value]. The kernel size is k, the stride is 1, and zero-padding is used to maintain the sequence length. By stacking multiple dilated convolutions, the maximum receptive field of the temporal prediction unit can reach:

[0048] ;

[0049] in The baseline dilation rate is the value of the corresponding branch; when the kernel size... Number of floors At that time, the maximum receptive field of the branch with a baseline inflation rate of 1 is 31 time steps, and the maximum receptive field of the branch with a baseline inflation rate of 8 can reach 241 time steps. The input sequence length T is set to 60 time steps, corresponding to a 60-minute historical window, and the sampling interval is 1 minute. Each branch captures the temporal dependencies at the 30-minute, 60-minute, 120-minute, and 240-minute scales respectively through inflation rate configuration, which can effectively cover the traffic state change patterns at different time scales. The outputs of multiple parallel branches are weighted and combined through a gating fusion mechanism.

[0050] The specific calculation process of the gating fusion mechanism is as follows: the output of each parallel branch is denoted as The gated network is a single-hidden-layer fully connected network, and its input is the concatenated vector of the outputs of each branch. The hidden layer has 16 neurons, using the ReLU activation function. The output layer has K neurons, which is the total number of branches. The gating weights are obtained after Softmax normalization. The final fusion result is the weighted sum of the outputs of each branch according to the gating weights:

[0051] ;

[0052] The weights and bias parameters of the fully connected network are automatically optimized through backpropagation during training, enabling the model to adaptively assign weights to prediction tasks at different time scales.

[0053] The input sequence length T is set to 60 time steps, corresponding to a 60-minute historical window, with a sampling interval of 1 minute. The predicted sequence length is... Set to 30 time steps, corresponding to a 30-minute prediction window.

[0054] The training dataset is divided into training, validation, and test sets in a 7:1.5:1.5 ratio. The training objective of the time series prediction unit is to minimize the mean absolute error loss function between the predicted sequence and the true sequence. The specific form of the loss function is as follows:

[0055] ;

[0056] in, To predict the number of time steps; It is the set of all road segments in the road network; The time prediction unit is the predicted travel time of road segment v at the t-th prediction time step; Let t be the actual travel time value of road segment v at the t-th prediction time step; during training, the Adam optimizer is used for parameter iterative updates, the initial learning rate is set to 0.001, the training rounds are set to 150 rounds, and training is terminated early when the validation set loss no longer decreases for 15 consecutive rounds.

[0057] The fusion output unit receives spatial features from the graph neural network unit and temporal features from the temporal prediction unit, and generates fused features through bilinear fusion or attention fusion. Bilinear fusion performs an outer product operation on the spatial feature vector and the temporal feature vector to obtain a bilinear feature matrix. After flattening the bilinear feature matrix, it is projected onto the target fused feature dimension through a fully connected layer. Attention fusion concatenates the spatial feature vector and the temporal feature vector after linear transformation, and calculates the attention weight of each feature channel through a single hidden layer attention network. The spatial feature and the temporal feature are then weighted and summed to obtain the fused feature vector.

[0058] The fusion feature dimension ranges from 64 to 256, and the optimal dimension configuration is determined through cross-validation before system deployment.

[0059] This unit inputs the fused features into a mapping network consisting of two fully connected layers. The first layer contains 128 neurons and uses the ReLU activation function. The second layer contains the same number of output units as the road segments. The output layer uses a linear activation function to map the fused features into a congestion index prediction for each road segment.

[0060] The road network situation heat map presents the traffic status of each road segment in a color gradient visualization, where green indicates smooth traffic, yellow indicates slow traffic, and red indicates congestion, providing intuitive situational awareness results for scheduling decisions.

[0061] Through the collaborative work of the data acquisition module and the various sub-modules of the situational awareness engine, this embodiment completes the entire process from multi-source data acquisition, spatiotemporal feature extraction, graph neural network spatial dependency modeling, temporal convolutional network trend prediction to the generation of road network situational heatmaps, providing a precise situational awareness data foundation for the subsequent multi-agent scheduling decision-making of the collaborative decision-making engine.

[0062] Through the collaborative work of the aforementioned data acquisition module and the situational awareness engine, a precise situational awareness data foundation is provided for the subsequent multi-agent scheduling decision-making of the collaborative decision-making engine.

[0063] The collaborative decision-making engine is connected to the situational awareness engine to model each bus as an intelligent agent, construct a multi-agent collaborative decision-making model, and generate scheduling decisions for each vehicle based on deep reinforcement learning algorithms. The collaborative decision-making engine includes an intelligent agent modeling unit, a communication mechanism unit, a policy network unit, and a reward evaluation unit. The intelligent agent modeling unit models each bus as an intelligent agent with an independent state space and action space. The communication mechanism unit realizes information transmission between intelligent agents and achieves collaborative information interaction based on a value iteration network. The policy network unit outputs the optimal action policy for each intelligent agent based on deep reinforcement learning algorithms. The reward evaluation unit constructs a multi-objective reward function based on punctuality rate, passenger waiting time, and operating costs.

[0064] The agent modeling unit creates an independent agent instance for each bus, and each agent has a complete state space and action space definition; the state space is represented in tuple form:

[0065] ;

[0066] in, It indicates the vehicle's current location, consisting of the road segment sign where the vehicle is currently located and its relative position within that road segment; It indicates the vehicle's current speed, which is collected in real time by onboard positioning equipment; This indicates the current passenger load of the vehicle, which is reported in real time by the passenger flow statistics module. The congestion index of the road ahead is obtained from the situational awareness engine. It indicates the deviation between the planned arrival time and the actual arrival time, and is used to assess the punctuality status of the vehicle; It indicates the location information of nearby vehicles, which is obtained through the vehicle-to-everything (V2X) communication module.

[0067] The motion space adopts a hybrid form combining discrete and continuous motion:

[0068] ;

[0069] in, This indicates an acceleration action, whose parameters include the target speed increment. This indicates a deceleration action, and its parameters include the target speed reduction. Indicates maintaining the current speed; no additional parameters are required. This indicates an action to request joining the cross-line support queue, and its action parameters include the target support line identifier; This indicates an action to request a change of driving route, and its action parameters include the target route identifier in the set of alternative routes.

[0070] For cross-line support actions, multiple agents achieve autonomous matching and coordination through distributed communication, without the need for full manual intervention from the scheduling center. The specific interaction process is as follows:

[0071] 1. Request Initiation: When a vehicle agent on a certain route detects that the number of people waiting at the next station on the route exceeds the capacity threshold (e.g., more than 80 people waiting at a single station) and there is no spare capacity on the route, it broadcasts a cross-line support request to all vehicle agents within the communication range. The request message includes the requesting station, the requesting time period, and the required capacity level.

[0072] 2. Willingness Feedback: Upon receiving a support request, the vehicle's intelligent agent calculates its support suitability based on its current passenger load, remaining operational tasks, and travel time to the requested station. The formula for calculating support suitability is:

[0073] ;

[0074] in Based on the current passenger capacity, Estimated arrival time at the support station (in minutes). This represents the percentage of remaining operational tasks. The normalized coefficient of each indicator is the maximum value, and the weighting coefficient is set to [value]. .

[0075] Vehicle agents with a suitability higher than the preset threshold of 0.6 provide feedback to the requester regarding their willingness to provide support and their estimated arrival time. The threshold of 0.6 was determined through offline simulation optimization to balance support response speed and the vehicle's inherent operational efficiency. When multiple requesters simultaneously apply for the same support vehicle, the vehicles are allocated according to the priority of the requesters' capacity gaps, and requesters with higher suitability are selected under the same priority.

[0076] 3. Matching Confirmation: The requesting agent selects 1-2 vehicles with the highest compatibility from all willing vehicles, sends a confirmation command, and synchronizes it to the dispatch center for record-keeping; the selected vehicle agent triggers a route adjustment and heads to the support station; the unselected vehicles resume their original operation plan.

[0077] 4. Task Completion: After the support vehicle completes the support task, it sends a task completion message to the original route dispatching agent and dispatching center, and the reward evaluation unit calculates the cross-line cooperation reward.

[0078] The communication mechanism unit adopts a hybrid architecture that combines centralized and distributed approaches. The scheduling center acts as the central coordination node, aggregating global situational information and distributing it to each agent. Meanwhile, each agent shares local state and intent information through point-to-point communication. In this hybrid architecture, the centralized component is responsible for maintaining the global state view and calculating the global value function based on the value iteration network, providing a benchmark reference for collaborative decision-making among agents.

[0079] The distributed component employs an attention-based message aggregation mechanism, enabling each agent to selectively receive and process neighbor information based on task relevance; for the i-th agent, let its communication radius be... There are M neighboring agents in total, and the message vector sent by neighbor j is... Generated by splicing encoding: .

[0080] in As a position embedding vector, the position coordinates are mapped to a 32-dimensional space through a two-layer fully connected network. It is an 8-dimensional velocity encoding vector. The message vector is a one-hot encoded vector of the action intent, with a total dimension of 56. When the number of neighbors changes, the aggregated vector is padded with zeros to a fixed dimension of 64 before being input into the policy network. It encodes information about up to 10 neighboring vehicles, and fills the vector with zeros when there are fewer than 10 neighboring vehicles.

[0081] Agent i through linear transformation matrix My own state Mapped to query vector ; through linear transformation matrix Messages from each neighbor Mapped to key vectors: ; Calculate the attention scores between the query vector and each key vector:

[0082] ;

[0083] in The key vector dimension has a value of 64, which is the scaling factor. This is used to prevent the softmax gradient from saturating due to an excessively large inner product.

[0084] The global baseline provided by centralized components and the local awareness provided by distributed components are weighted and combined through a gating fusion mechanism during the action selection process of policy network units. The fusion weights are determined by a global coordination factor. With local confidence This is calculated collaboratively, ensuring that each agent's decisions both follow the global optimization direction and can respond quickly to changes in the local environment. Among these factors is the global collaborative factor. The value is calculated by the centralized component based on the global value function, and the calculation formula is as follows:

[0085] ;

[0086] In the formula, This is an estimate of the total system value in a global state. This is the local value estimate obtained by the i-th agent based solely on its local state. The value range is limited to [0.5, 2.0], used to characterize the importance of the global optimization objective relative to local decisions. Local confidence score. The calculation is obtained by the agent based on the quality of neighbor communication and the integrity of local data. The calculation formula is as follows:

[0087] ;

[0088] Where PRR is the packet reception rate of neighbor messages in the most recent 10 communication cycles, and DSR is the local sensor data integrity rate, which is the ratio of the effective sampled data volume to the expected total sampled data volume in the statistical period, and the value range is [0,1]. The more stable the communication link, the more complete the local sensing data. The higher the value, the greater the final global information fusion weight.

[0089] ;

[0090] Local information fusion weights:

[0091] ;

[0092] The final action value of the agent is obtained by weighting the global action value and the local action value according to the above weights.

[0093] The message passing of the communication mechanism unit follows the following protocol specifications: After each communication cycle Each intelligent agent reports its local status to the scheduling center. and intention of action The dispatch center summarizes the global status. Then, calculate the global value estimate. and the cooperating factors of each agent It also broadcasts global situational information and cooperation factors to all agents; neighboring agents communicate based on communication radius. Establish point-to-point connections to exchange local status and short-term action plans.

[0094] The communication radius The value is based on the fact that the PC5 direct communication interface based on cellular vehicle-to-everything (C-V2X) has an effective communication distance of 300 to 500 meters in a typical urban environment, which is sufficient to cover the information exchange needs between adjacent buses.

[0095] As a further improvement to the above technical solution, the communication mechanism unit also incorporates a three-level communication degradation processing strategy. The degradation switching is based on a dual-dimensional judgment of continuous packet loss rate and heartbeat timeout time. The specific rules and switching process are as follows:

[0096] Level 1 Mode (Fully Collaborative Mode): The continuous packet loss rate between the agent and the scheduling center is no higher than 20%, and this occurs over three consecutive communication cycles. All receive global broadcast information. At this time, the agent simultaneously accesses centralized global information and distributed neighbor information, and outputs collaborative decisions according to the fusion weight. Among them, the optimal communication state is when the continuous packet loss rate is less than 5%, and the first-level full collaborative mode is maintained within the range of 5% to 20%.

[0097] Level 2 mode (neighbor collaboration mode): When the continuous packet loss rate between the agent and the scheduling center exceeds 20%, or for 5 consecutive communication cycles... If no global broadcast message is received, a first-level degradation is automatically triggered, and the agent switches to a local decision-making mode that relies solely on distributed neighbor communication; in this mode, the agent only utilizes the communication radius. Action selection is based on the state information of neighboring agents, and regional collaboration is achieved through local value iteration among neighbors, eliminating the need to wait for global scheduling instructions.

[0098] Level 3 Mode (Autonomous Decision-Making Mode): When the continuous packet loss rate of PC5 direct communication between the agent and neighboring vehicles exceeds 30%, or after 10 consecutive communication cycles... If no heartbeat information is received from a neighbor, a second-level degradation is triggered, and the agent reverts to a state based on its local state. The single-agent decision-making mode; in this mode, the agent only relies on the information in its own state space S, and outputs basic scheduling actions based on the pre-trained single-agent policy network to prioritize the on-time operation of vehicles.

[0099] When the communication link is restored and the packet loss rate remains below 80% of the target threshold for six consecutive cycles, the system automatically upgrades to a higher level of collaborative mode to ensure the system's basic operational capability under communication network degradation.

[0100] The policy network unit adopts an Actor-Critic architecture, where the Actor network outputs the action probability distribution based on the current state, and the Critic network evaluates the state value function. The two work together iteratively to achieve stable policy convergence. The Actor network is implemented using a deep neural network, and its input is the agent's state vector. The output is a probability distribution in the action space. Its network structure contains multiple fully connected layers, each using the ReLU activation function, and the output layer uses the Softmax function to generate the action probability distribution.

[0101] The training data for the Actor network comes from state-action-reward trajectory samples collected during the interaction between the agent and the environment. Each sample contains the current state. Take action Receive rewards and the next state During training, an experience replay mechanism is used to store the samples generated by the interaction in an experience pool. During training, random batch sampling is performed from the experience pool to break the temporal correlation between samples and improve training stability.

[0102] This multi-agent decision-making system adopts a centralized training and distributed execution (CTDE) training paradigm: during the training phase, a centralized Critic network accesses the global state. With the set of actions of all agents A={ The global value function and the advantage function of each agent are calculated; during the execution phase, each agent relies solely on its local state. The received neighbor information and global coordination factors are used to independently output actions through the local Actor network without relying on the global complete state.

[0103] The training process employs a centralized experience replay pool, uniformly storing state-action-reward-next-state trajectory samples for all agents. In each training round, a batch of samples containing the synchronized trajectories of N agents is sampled from the experience pool, simultaneously updating the global Critic network and the Actor networks for each agent. To mitigate the environmental non-stationarity problem in multi-agent training, the policy network unit adopts a policy delayed update mechanism: the Critic network updates twice, and the Actor network updates once. Furthermore, the target network uses a soft update method, with the soft update coefficient τ set to 0.005.

[0104] The training objective of the Actor network is to maximize the cumulative expected reward. Its parameter updates follow the policy gradient theorem:

[0105] ;

[0106] in, Expected cumulative reward; For strategy State access distribution The advantage function is used to measure the advantage in a given state. Take action below The degree of advantage compared to the benchmark strategy; Indicates the network parameters of the policy Find the gradient.

[0107] The Critic network is also implemented using a deep neural network, with the agent's state vector as input. The output is a state value function estimate. The training data for the Critic network comes from state-reward trajectory samples collected during the interaction between the agent and the environment. Each sample contains the current state. Instant rewards and the next state The training objective of the Critic network is to minimize the temporal difference error. By fitting a state-value function, it provides an accurate value evaluation benchmark for the Actor network. Its loss function is:

[0108] ;

[0109] in, For instant rewards; This is a discount factor used to balance the importance of immediate rewards and future rewards; The state at the next moment; The parameters of the Critic network before the update; These are the current Critic network parameters.

[0110] The reward evaluation unit constructs a multi-objective reward function, which includes on-time arrival rewards, passenger waiting time penalties, operating cost penalties, cross-line cooperation rewards, and conflict avoidance penalties. After performing min-max normalization on each sub-item, multi-objective optimization is achieved through weighted summation. The specific form of the reward function is as follows:

[0111] ;

[0112] in , , , , These are the weighting coefficients for each reward, with a default value of [value]. =0.35、 =0.3、 =0.15、 =0.1、 =0.1, and the value range of each weight coefficient is [0.05, 0.5]. The dispatch administrator can adjust the weight allocation within this range according to operational needs, and must maintain... and The relative proportions are prioritized to ensure on-time performance remains the core optimization objective; the quantitative calculation and normalization methods for each sub-reward item are as follows:

[0113] 1. Reward for arriving on time Let the planned arrival time of the vehicle be... The actual arrival time is The tolerance window is [-60s, 120s]; when When it falls within the tolerance window, =+10; when >300s, =-20; the reward value for the remaining intervals is calculated using linear interpolation; the normalization method is:

[0114] ;

[0115] in , After normalization, the numerical range is [0,1].

[0116] 2. Passenger waiting time penalty Let the number of passengers waiting at station k be... The average waiting time is If the unit is min, then:

[0117] ;

[0118] The number of waiting passengers is calculated as a weighted average of the cumulative number of card swipes and infrared detection values ​​at the station within the last 5 minutes. The average waiting time is the difference between the current time and the earliest card swipe time at the station. The normalization method is as follows:

[0119] ;

[0120] in The baseline value for the maximum number of people waiting at a single station is 200. The baseline value for the longest waiting time is 30 minutes. After normalization, the value ranges from [0,1]. The penalty value increases linearly with the product of the number of people waiting and the waiting time.

[0121] 3. Operating cost penalty Assume the energy cost per unit mileage of the vehicle is The labor cost per unit time is The mileage driven during the statistical period was The runtime is ,but:

[0122] ;

[0123] The normalization method is as follows:

[0124] ;

[0125] in The maximum operating cost benchmark for a single cycle is determined based on the hourly statistics of energy consumption per unit mileage and labor cost per unit time of the route. After normalization, the value range is [0,1]. Negative rewards are used to guide vehicles to reduce ineffective driving and empty driving.

[0126] 4. Cross-line collaboration rewards When a vehicle responds to a cross-line support request and completes passenger transport at least one support station, it receives a base reward of +15. If the average departure interval deviation of the target route decreases by more than 20% after support, an additional reward of +10 is awarded. The above reward values ​​are determined based on the statistical data of the average fare and unit operating cost of urban bus routes. The normalized values ​​range from [0,1], and are determined by weighting coefficients. The weighted summation is incorporated into the reward function.

[0127] 5. Conflict Avoidance and Punishment When two or more vehicles simultaneously request to enter the station or simultaneously request the same cross-line support mission within a 30-second time window at the same station, all agents involved in the conflict will receive [benefits / rewards] for each conflict. A penalty of -8 is applied to guide the agent to autonomously avoid resource competition through information interaction; after normalization, the corresponding numerical range is [0,1], which is determined by weighting coefficients. The weighted summation is incorporated into the reward function.

[0128] Through the collaborative work of the aforementioned collaborative decision-making engine and dispatching command engine, a complete intelligent solution for bus operation dispatching is provided, from situational awareness to multi-agent decision-making and command execution.

[0129] The scheduling command engine is connected to the collaborative decision engine to convert scheduling decisions into specific execution commands and send them to the corresponding vehicle terminals through the communication transmission layer. The scheduling command engine consists of three parts: a command generation unit, a command issuance unit, and an execution feedback unit. The command generation unit is responsible for converting the collaborative decision results into structured command messages, the command issuance unit is responsible for sending the commands to the target vehicles through the mobile communication network, and the execution feedback unit is responsible for receiving the execution results and updating the system status.

[0130] The instruction generation unit receives the action suggestions output by the collaborative decision engine and converts them into structured scheduling instruction messages. The scheduling instruction message contains the following fields: the instruction number is a globally unique identifier, which is a string and has the format "INST-YYYYMMDD-XXXXXX", where YYYYMMDD represents the instruction generation date and XXXXXX is a six-digit sequential number.

[0131] The target vehicle identifier is a string type, recording the license plate number or vehicle ID; the instruction type is an enumeration type, with values ​​including 1 for departure time adjustment, 2 for cross-line support dispatch, 3 for temporary vehicle addition, and 4 for detour route suggestion; the instruction parameters are JSON objects, defining corresponding parameter sets according to different instruction types; the execution time window is a time interval type, representing the effective execution period of the instruction and including start and end times; the priority identifier is an integer type, divided into four levels: 1 for urgent, 2 for high, 3 for normal, and 4 for low; the creation timestamp and expiration timestamp are both date and time types, recording the specific time of instruction generation and expiration, respectively.

[0132] The parameter structures for different instruction types are defined as follows: The parameters for departure time adjustment instructions include the original departure time, the adjusted departure time, and the reason for adjustment; the parameters for cross-line support dispatch instructions include the supporting vehicle identifier, the supporting origin station, the supporting destination station, and the supporting route identifier; the parameters for temporary vehicle addition instructions include the number of vehicles added, the time period for adding vehicles, and the route for adding vehicles; the parameters for detour route suggestion instructions include the original route set, the suggested detour route, and an explanation of the detour reason. Through the definitions of the above four instruction types and their corresponding parameter structures, the dispatch instruction engine can cover common adjustment needs in public transport operation dispatching, including four scenarios: departure time adjustment, cross-line resource allocation, temporary capacity supplementation, and route change.

[0133] The instruction issuing unit sends instructions to the vehicle-mounted terminal of the target vehicle through the mobile communication network. The communication transmission layer uses a fifth-generation mobile communication network or a fourth-generation mobile communication network as the data transmission channel between the cloud dispatch center and the vehicle-mounted terminal. The instruction issuing unit has a built-in message queue management module to prioritize and optimize the batch issuance of instructions. When the network connection is interrupted, the instruction issuing unit caches the instructions to local persistent storage and automatically retransmits them after the network is restored, ensuring the reliability of instruction delivery.

[0134] The execution feedback unit receives execution feedback messages from the vehicle terminal and updates the system status. The execution feedback messages include information such as instruction execution status, actual execution time, vehicle location updates, and abnormal event reports. The execution feedback unit evaluates the instruction execution results based on the feedback information. If the instruction is successfully executed, the scheduling plan is updated and the execution log is recorded. If the instruction fails, an exception handling process is triggered and an alternative scheduling scheme or manual intervention request is generated. The execution feedback unit also sends the execution results back to the collaborative decision engine as environmental feedback for the reinforcement learning agent, which is used for offline training and online fine-tuning of the policy network.

[0135] Through the complete working mechanism of the aforementioned scheduling instruction engine, this embodiment realizes a complete closed loop from scheduling decision-making to structured instruction generation, from priority sorting and batch issuance to network interruption caching and retransmission, from execution result feedback and status update to exception handling and policy network feedback, providing automated execution and closed-loop management capabilities for public transportation operation scheduling decisions from instruction generation to execution feedback.

[0136] Example 2

[0137] Based on Embodiment 1, this embodiment also provides the vehicle-mounted terminal part of the intelligent scheduling system driven by public transportation operation data. The vehicle-mounted terminal includes four components: a positioning module, a passenger flow statistics module, a dispatch display screen, and a vehicle-to-everything (V2X) communication module. The positioning module is used to obtain the real-time location and driving status of the vehicle, the passenger flow statistics module is used to collect the number of passengers getting on and off the bus and the distribution information of stations, the dispatch display screen is used to display the received dispatch instructions to the driver, and the V2X communication module is used to realize information interaction between the vehicle and the dispatch center and other vehicles.

[0138] The vehicle-mounted terminal's positioning module employs multimodal positioning technology, integrating data from GPS, BeiDou Navigation Satellite System (BDS), and Inertial Measurement Unit (IMU) to provide high-precision vehicle position and driving status information. The positioning module's data update frequency is no less than 1Hz, and its position accuracy is better than 5m in open environments. In multipath environments such as urban canyons, the accuracy can be maintained within 15m with inertial navigation assistance. The positioning module also outputs motion state parameters such as vehicle heading angle, speed, and acceleration, providing data input for the situational awareness engine and collaborative decision-making engine.

[0139] The passenger flow statistics module of the vehicle terminal adopts a dual-mode detection scheme combining infrared beam sensors and pressure sensors. The infrared beam sensors are installed on both sides of the door to detect occlusion events when passengers pass through the door. The pressure sensors are installed in the door sill area to help determine the passenger's boarding and alighting status. The passenger flow statistics module has a built-in passenger counting algorithm. By fusing the results of infrared detection and pressure detection, it eliminates false judgments of occlusion and count omissions, and realizes accurate statistics of the number of passengers boarding and alighting.

[0140] The fusion logic of the passenger counting algorithm is as follows: When the infrared beam sensor detects an obstruction event, the pressure change sequence of the pressure sensor is sampled simultaneously; if the pressure change sequence meets the two-step feature (corresponding to the pressure change of the passenger's two feet when getting on and off the vehicle), it is determined to be a valid boarding or alighting event; if the infrared obstruction lasts for more than 2 seconds but the pressure sensor has no valid pressure change, or the pressure change sequence does not meet the two-step feature, it is determined to be an obstruction misjudgment and the infrared event is ignored; the number of valid boarding events and valid alighting events is accumulated, and the number of boarding events and the number of alighting events are output respectively.

[0141] The passenger flow statistics module outputs passenger count results by station, including three indicators: the number of passengers boarding, the number of passengers alighting, and the current passenger capacity, and reports them to the dispatch center in real time through the vehicle network communication module.

[0142] The vehicle-mounted terminal's dispatch display screen uses a high-brightness LCD screen with a size of no less than 7 inches and a resolution of no less than 1024×600 pixels. The dispatch display screen receives dispatch command messages from the dispatch command engine, parses the command content, and presents it to the driver in a visual manner. The interface layout of the dispatch display screen includes three parts: a top status bar, a central main display area, and a bottom operation bar.

[0143] The top status bar displays the current time, vehicle number, and current route information; the central main display area shows different instruction content according to the instruction type, such as departure time adjustment instructions displaying the countdown to the adjusted departure time, cross-line support dispatch instructions displaying the support route and stops along the way, and detour route suggestion instructions displaying an electronic map and recommended routes; the bottom operation bar provides three operation buttons: confirm, reject, and request manual assistance.

[0144] The vehicle-to-everything (V2X) communication module of the in-vehicle terminal supports the cellular V2X communication protocol, enabling information interaction between vehicles, the dispatch center, and other vehicles. This communication module supports two communication modes: Uu interface communication based on the cellular network for high-bandwidth, low-latency data transmission with the dispatch center, and PC5 interface communication based on direct communication for low-latency, real-time information exchange with nearby vehicles. The V2X communication module also supports multiple protocol stacks, including Transmission Control Protocol (TCP), User Datagram Protocol (UDP), and Message Queuing Telemetry Transport Protocol (MQTT), allowing the selection of the optimal transmission protocol based on the communication scenario.

[0145] With the complete configuration of the aforementioned vehicle-mounted terminal, the vehicle-mounted terminal provides reliable vehicle-side data collection, instruction presentation, and communication support for the intelligent scheduling system driven by public transportation operation data, ensuring a complete closed loop of dispatch instructions from the cloud to the vehicle.

[0146] Example 3

[0147] Based on Example 2, this example also provides the working states of a bus operation data-driven intelligent scheduling system. The system's working states include eight states: normal monitoring state, situation prediction state, collaborative decision-making state, instruction issuance state, instruction execution state, effect evaluation state, emergency dispatch state, and cross-line support state. The normal monitoring state is the system's baseline operating state. Under the normal monitoring state, the data acquisition module continuously collects multi-source data, and the situation awareness engine updates the road network situation heat map at a fixed cycle, which is set to 5 minutes. In this example, under the normal monitoring state, only data acquisition and situation awareness are performed without triggering scheduling adjustments.

[0148] When the congestion index of any road segment exceeds the first preset threshold At this time, this embodiment transitions from normal monitoring state to situation prediction state, with the first preset threshold. The value is 0.7, which corresponds to a moderate congestion state. The congestion index is defined as the ratio of the current travel time of a road segment to the historical free-flow travel time of that road segment. The situation prediction state starts the time-series prediction unit to predict the evolution of the road network situation in the next 15 to 30 minutes. The future road network situation is predicted based on graph neural networks and temporal convolutional networks.

[0149] If the predicted congestion impact exceeds the second preset threshold Then it transitions to a collaborative decision-making state, with the second preset threshold. The criteria for determining whether a bus is affected are: the number of vehicles affected within the next 15 minutes is predicted to exceed 10% of the baseline number of vehicles on the route, or the number of passengers affected is predicted to exceed 15% of the average hourly passenger flow on the route. The baseline number of vehicles and the average hourly passenger flow are determined based on the historical operating data of each route. Affected vehicles refer to buses whose predicted travel time delay exceeds 3 minutes.

[0150] The collaborative decision-making process invokes a multi-agent deep reinforcement learning algorithm to generate scheduling decision schemes for all buses in the affected area, with a decision generation cycle of 30 seconds to 1 minute.

[0151] After the vehicle terminal returns feedback, it enters the effect evaluation state to evaluate the execution effect of the dispatching instructions; the effect evaluation is based on the following quantitative indicators:

[0152] ①Amount to increase in on-time performance:

[0153] ;

[0154] in The on-time rate of the affected lines within 1 hour after the dispatch instruction is executed. The on-time rate for the hour preceding execution.

[0155] ② Passenger waiting time improvement rate:

[0156] .

[0157] ③Number of passengers affected .

[0158] The judgment rules are as follows: If and and If the effect is deemed satisfactory, return to normal monitoring status; if or If it is determined that an emergency dispatch state needs to be entered; Then, it re-enters the collaborative decision-making state to generate an optimized solution.

[0159] In the event of a major emergency, this embodiment can directly switch from any state to emergency dispatch state and activate the emergency plan. Major emergencies include road closures, vehicle breakdowns, and large-scale congestion with a congestion index > 0.9. If the capacity of a local route is insufficient, it can switch to cross-line support state and coordinate vehicle resources across routes for support dispatch.

[0160] The preset thresholds are set according to the principle of configurability. The dispatch administrator can adjust the threshold parameters according to actual operational needs. The first preset threshold is... The value ranges from 0.5 to 0.9. A lower value indicates greater sensitivity to congestion and more frequent scheduling responses. The second preset threshold... The value needs to be determined by combining the line capacity and passenger flow to ensure that the benefits of scheduling decisions outweigh the execution costs.

[0161] Through the complete definition of the above eight working states and their transition rules, this embodiment realizes a closed loop of state management for the bus operation dispatching system, from routine monitoring to situation prediction, from collaborative decision-making to instruction issuance and execution, and from effect evaluation to emergency response. It provides a state management framework covering all scenarios of normal operation and abnormal handling for the bus operation data-driven intelligent scheduling system.

[0162] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0163] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A smart scheduling system driven by public transportation operation data, characterized in that: include: The data acquisition module is used to collect multi-source operational data, including vehicle location data, passenger flow data, road network topology data, real-time traffic data, and operation plan data. A situational awareness engine, connected to the data acquisition module, is used to receive the structured data stream output by the data acquisition module. The situational awareness engine includes a graph neural network unit and a temporal prediction unit. The graph neural network unit performs spatial dependency modeling based on the road network topology, enabling each road segment node to perceive the traffic status of adjacent road segments. The temporal prediction unit predicts the travel time of each road segment within a preset time period based on a temporal convolutional network. The situational awareness engine outputs a road network situational analysis result containing the congestion index prediction of each road segment. A collaborative decision-making engine, connected to the situational awareness engine, is used to model each bus as an agent, construct a multi-agent collaborative decision-making model, and generate scheduling decisions for each vehicle based on a deep reinforcement learning algorithm. The collaborative decision-making engine includes an agent modeling unit, a communication mechanism unit, a policy network unit, and a reward evaluation unit. The agent modeling unit creates agent instances with independent state and action spaces for each bus. The communication mechanism unit adopts a hybrid communication architecture combining centralized and distributed approaches to realize information transmission between agents and collaborative information interaction based on a value iteration network. The policy network unit outputs the action policy of each agent based on an Actor-Critic architecture. The reward evaluation unit evaluates the decision-making effect based on a multi-objective reward function. A scheduling instruction engine, connected to the collaborative decision engine, is used to convert scheduling decisions into structured instruction messages and send them to the corresponding vehicle terminal through the communication transmission layer. The scheduling instruction engine includes an instruction generation unit, an instruction sending unit, and an execution feedback unit. The instruction generation unit converts the collaborative decision results into structured instruction messages. The instruction sending unit sends the instructions to the target vehicle through the mobile communication network. The execution feedback unit receives the execution results and updates the system status.

2. The intelligent scheduling system driven by public transportation operation data according to claim 1, characterized in that: The graph neural network unit adopts a message passing mechanism to iteratively propagate road segment status information among nodes in the road network topology graph, enabling each road segment node to perceive the traffic status of its adjacent road segments. In the aforementioned message passing mechanism, when a node updates at each layer, it aggregates the message vectors transmitted by all its neighboring nodes. The state update function of the gated cyclic unit structure generates the hidden state vector of the current layer. The message aggregation adopts the attention-weighted aggregation method, which calculates the attention weights between nodes and weights the adjacent message vectors to sum them. After multiple iterations, the spatial correlation of congestion propagation in the road network is captured.

3. The intelligent scheduling system driven by public transportation operation data according to claim 1, characterized in that: The temporal prediction unit adopts a structure that combines causal convolution and dilated convolution. Causal convolution ensures that the output at the current time depends only on the input at the previous time, while dilated convolution expands the receptive field of the convolution kernel by setting the dilation rate. The temporal prediction unit contains four parallel causal dilated convolutional branches, with the baseline dilation rate of each branch configured as 1, 2, 4, and 8, respectively. The convolutional stride is 1 and zero padding is used to preserve the sequence length. The outputs of multiple parallel branches are weighted and combined through a gating fusion mechanism. The gating weights are learned by a single hidden layer fully connected attention network based on the outputs of each branch. After normalization, the features of each branch are weighted and summed to obtain the fusion result, which adaptively captures traffic state changes at different time scales.

4. The intelligent scheduling system driven by public transportation operation data according to claim 1, characterized in that: In the intelligent agent instance created by the intelligent agent modeling unit for each bus, the state space is in tuple form and includes the vehicle's current position, current speed, current passenger load, congestion index of the road ahead, deviation between planned arrival time and actual arrival time, and location information of nearby vehicles. The action space adopts a hybrid form combining discrete action categories and continuous action parameters, including acceleration action, deceleration action, maintaining current speed action, requesting to join cross-line support queue action, and requesting to change driving route action. The request to join cross-line support queue action includes target support line identification parameters, and the request to change driving route action includes alternative route identification parameters.

5. The intelligent scheduling system driven by public transportation operation data according to claim 1, characterized in that: The communication mechanism unit adopts a hybrid communication architecture that combines centralized and distributed approaches. The scheduling center acts as the central coordination node, which aggregates global situational information and distributes it to each agent. At the same time, each agent shares local state and intent information through point-to-point communication. The centralized component of the communication mechanism unit calculates the global value function and the cooperation factor of each agent based on the value iteration network and broadcasts them to all agents. In the distributed components of the communication mechanism unit, each agent maps its own state to a query vector through a linear transformation matrix, maps each neighbor message to a key vector through a linear transformation matrix, calculates the inner product of the query vector and each key vector and divides it by a scaling factor in the form of the square root of the key vector dimension to obtain an attention score, processes the attention score through a normalization function to obtain an attention weight, sums the neighbor messages according to the attention weight to obtain aggregate information, and concatenates the aggregate information with its own state and inputs it into the policy network for action selection. The global baseline output from the centralized component and the local perception output from the distributed component are weighted and combined through a gating fusion mechanism. The fusion weight is calculated by the global coordination factor and the local confidence, where the local confidence is quantified based on the neighbor communication packet reception rate and the local sensor data integrity rate.

6. The intelligent scheduling system driven by public transportation operation data according to claim 5, characterized in that: The communication mechanism unit incorporates a three-level communication degradation processing strategy, with degradation switching based on a dual-dimensional determination of continuous packet loss rate and heartbeat timeout time. The first level is the fully collaborative mode, where the intelligent agent simultaneously accesses centralized global information and distributed neighbor information; When the continuous packet loss rate between the agent and the scheduling center exceeds the first preset threshold of 20%, or when no global broadcast information is received for 5 consecutive global communication cycles, the second-level neighbor collaboration mode is triggered, and the agent switches to a local decision-making mode that relies solely on distributed neighbor communication. When the packet loss rate of the direct communication between the agent and the neighboring vehicle exceeds the second preset threshold of 30%, or when no heartbeat information from the neighbor is received for 10 consecutive local communication cycles, the third-level autonomous decision-making mode is triggered, and the agent reverts to the single agent decision-making mode based on the local state. When the communication link is restored and the packet loss rate remains below 80% of the target threshold for six consecutive cycles, the system will automatically upgrade to a higher level of collaborative mode.

7. The intelligent scheduling system driven by public transportation operation data according to claim 1, characterized in that: The policy network unit adopts a training paradigm of centralized training and distributed execution. During the training phase, the centralized Critic network accesses the global state and the action set of all agents to calculate the global value function and the advantage function of each agent. During the execution phase, each agent independently outputs actions through the local Actor network, relying solely on its local state, received neighbor information, and global coordination factors. The Actor network outputs an action probability distribution based on the current state, and the Critic network evaluates the state value function. The parameters of the Actor network are updated according to the policy gradient, and the training objective of the Critic network is to minimize the temporal difference error.

8. The intelligent scheduling system driven by public transportation operation data according to claim 1, characterized in that: The multi-objective reward function constructed by the reward evaluation unit includes on-time arrival reward, passenger waiting time penalty, operating cost penalty, cross-line cooperation reward and conflict avoidance penalty. After min-max normalization, the multi-objective optimization is achieved by weighted summation. The cross-line cooperation reward is a basic reward when a vehicle responds to a cross-line support request and completes the passenger transport task, and an additional reward is given when the average departure interval deviation of the target line is reduced by more than a preset proportion after the support. The conflict avoidance penalty is triggered when multiple vehicles simultaneously apply to enter the station or simultaneously apply for the same cross-line support mission within a 30-second time window at the same station. The weight coefficients for each reward item range from 0.05 to 0.5, with priority given to maintaining the relative weight of on-time arrival rewards and passenger waiting time penalties.

9. The intelligent scheduling system driven by public transportation operation data according to claim 1, characterized in that: The structured instruction message generated by the instruction generation unit includes instruction number, target vehicle identifier, instruction type, instruction parameters, execution time window, and priority identifier. The instruction types include four types: departure time adjustment, cross-line support scheduling, temporary vehicle addition, and detour route suggestion. Different instruction types correspond to different instruction parameter structures. The instruction issuing unit has a built-in message queue management module that prioritizes and optimizes the batch issuance of instructions. When the network connection is interrupted, the instructions are cached to local persistent storage and automatically retransmitted after the network is restored.

10. The intelligent scheduling system driven by public transportation operation data according to claim 1, characterized in that: The system's working states include eight states: normal monitoring state, situation prediction state, collaborative decision-making state, command issuance state, command execution state, effect evaluation state, emergency dispatch state, and cross-line support state. Under normal monitoring conditions, the data acquisition module continuously collects multi-source data, and the situational awareness engine updates the road network situational heat map at a fixed period. When the congestion index of any road segment exceeds the first preset threshold, the system switches from normal monitoring to situation prediction and activates the time-series prediction unit to predict the future evolution of the road network situation. If the predicted congestion impact exceeds the second preset threshold, the system switches to a collaborative decision-making state, invoking a multi-agent deep reinforcement learning algorithm to generate a scheduling decision scheme. The second preset threshold is determined when the predicted number of affected vehicles in the next 15 minutes exceeds 10% of the baseline number of vehicles on the route, or the number of affected passengers exceeds 15% of the average hourly passenger flow of the route. After the collaborative decision-making phase ends, the system transitions to the instruction issuance phase, where the scheduling instruction engine converts the decision plan into scheduling instructions and issues them. After the instruction is issued, the system transitions to the instruction execution state, waiting for the onboard terminal to provide the execution result. After the vehicle terminal returns feedback, it enters the effect evaluation state to evaluate the execution effect of the dispatching instructions; If the on-time rate improvement is ≥5%, the passenger waiting time improvement rate is ≥10%, and the number of affected passengers is ≤100, then the effect is considered to have met the target and the system returns to normal monitoring status. If the on-time rate declines or the number of affected passengers exceeds 500, the system will switch to emergency dispatch mode; otherwise, it will revert to collaborative decision-making mode. When road closures, vehicle breakdowns, or large-scale congestion with a congestion index greater than 0.9 occur, the system can be directly switched from any state to emergency dispatch state.

Images