Train dynamic scheduling method

Abstract

The present application relates to the field of data processing technology and provides a method for dynamic train scheduling. By acquiring and processing heterogeneous data in real time, generating train priorities, and then generating and verifying scheduling plans, the method achieves refined management of the entire train scheduling process. Each step cooperates with each other and progresses step by step, and can comprehensively consider various complex factors in the train operation process. It changes the limitations of traditional scheduling methods that rely on experience or single indicators, and significantly improves the scientificity, accuracy and timeliness of train scheduling decisions. In the face of complex and changeable operating scenarios, resources can be dynamically and efficiently allocated, train operation arrangements can be optimized, transportation efficiency and service quality can be improved, and the growing diversified needs of modern railway transportation can be effectively met.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a method for dynamic train scheduling. Background Art

[0002] With the rapid development of the railway transportation industry, the density of train operations has continued to increase, and the transportation environment has become increasingly complex, which has put forward higher requirements for the accuracy, efficiency and flexibility of train scheduling. Early train scheduling mainly relied on the experience and manual operation of dispatchers, and train operation was directed through telephones and signal equipment. This method is inefficient, prone to errors, and difficult to cope with complex transportation scenarios; with the advancement of information technology, an automated scheduling system based on a fixed timetable has emerged. This scheduling system performs scheduling through pre-established train operation plans, which has improved scheduling efficiency and accuracy to a certain extent. However, this method lacks adaptability to real-time changing factors. When encountering emergencies, the scheduling plan cannot be adjusted in time, which easily leads to problems such as train delays and line congestion.

[0003] At the same time, most existing technologies use a single data type, simple resource status assessment methods, generate scheduling plans with a single optimization objective, and verify scheduling plans that cannot effectively evaluate and adjust them based on real-time changing environmental factors and operating conditions. Most of these technologies fail to address how to achieve dynamic optimization of train scheduling and rational resource allocation in complex operating environments. Summary of the Invention

[0004] In response to the deficiencies of the existing technology, this application provides a train dynamic scheduling method.

[0005] In a first aspect, the present application provides a train dynamic scheduling method, comprising: acquiring heterogeneous data in real time, the heterogeneous data including track circuit signals, platform passenger flow images, and meteorological environment data; performing spatiotemporal alignment of the heterogeneous data at an onboard edge node using a spatiotemporal attention mechanism; processing the heterogeneous data; and simultaneously monitoring a power supply network, standby trains, and signal machines to determine resource status, thereby outputting a train data matrix, the train data matrix including a signal feature vector, a passenger flow density coefficient, a meteorological correction parameter, and resource status.

[0006] An evaluation vector for train scheduling is constructed based on the train data matrix and the scheduling knowledge graph. The evaluation vector includes an urgency factor, an economic value factor, and a line impact factor. The dimensional weights of the evaluation vector are dynamically adjusted based on the fuzzy analytic hierarchy process to generate train priorities.

[0007] The scheduling plan is generated through reinforcement learning. The state space of reinforcement learning represents train priority and resource status, and the reward function of reinforcement learning represents time punctuality, energy consumption reduction rate and conflict risk value.

[0008] Monte Carlo simulation is performed on each scheduling scheme in parallel using GPUs. The conflict probability, energy consumption deviation, and delay margin of each scheduling scheme are verified based on thresholds to output a verified scheduling instruction set. The scheduling instruction set includes train path adjustment and resource allocation instructions. The scheduling instruction set is executed and feedback data during the execution of the scheduling instruction set is monitored in real time to determine whether to resample heterogeneous data and update resource status.

[0009] As an optional implementation, the resource status determination sub-strategy includes:

[0010] Deploy sensors to monitor the power supply network, backup trains, and signals in real time to obtain resource data, including the load rate of the power supply network, the equipment health and location of backup trains, and the failure probability of signals.

[0011] A topology graph is established using the power supply network, standby trains, and signal machines as nodes. Node attributes represent resource data, and edges represent the coupling relationship between resource data. Node feature vectors of the topology graph are extracted using a graph convolutional network.

[0012] A Bayesian network is embedded in the topology graph, the node feature vector is input into the Bayesian network, and the Bayesian network is trained with historical operation data to determine the state transition probability between nodes, and the state transition probability between nodes is integrated to determine the resource status.

[0013] As an optional implementation, the output strategy of the train data matrix includes:

[0014] Real-time acquisition of heterogeneous data, including track circuit signals, platform passenger flow images, and meteorological environment data. This data is then spatially and temporally aligned at the onboard edge node using a spatiotemporal attention mechanism, and then standardized.

[0015] The track circuit signal is extracted through a residual convolutional neural network, and the extracted features are mapped into a signal feature vector, which includes signal strength, signal frequency, and signal phase.

[0016] Perform semantic segmentation on platform passenger flow images and calculate the ratio of passenger pixel area to total platform area to obtain the initial passenger flow density. Combined with historical passenger flow data, the long short-term memory network is used to predict passenger flow trends, dynamically weighting the initial passenger flow density to generate a passenger flow density coefficient.

[0017] The random forest algorithm is used to analyze the impact of different meteorological environmental data on train operation data, braking distance, and energy consumption. Based on the analysis results, an impact weight is assigned to each meteorological environmental data. The standardized meteorological environmental data and the impact weight are weighted and summed to generate meteorological correction parameters.

[0018] Monitor the power supply network, standby trains and signals to determine resource status, integrate signal feature vectors, passenger flow density coefficients, meteorological correction parameters and resource status, and output a train data matrix.

[0019] As an optional implementation, the sub-strategy for constructing the evaluation vector includes:

[0020] Normalize the train data matrix and associate it with the scheduling knowledge graph;

[0021] The train's operating status is analyzed by combining the signal feature vectors and resource status in the train data matrix. Historical scheduling data from the scheduling knowledge graph is extracted to assess the delay risk value under the train's operating status. The passenger impact value of train delays is assessed based on the passenger flow density coefficient and platform carrying capacity. The delay risk value and passenger impact value are weighted and summed to obtain the train's urgency factor.

[0022] Based on the energy consumption and meteorological correction parameters in the train data matrix, the energy consumption cost of the train is determined in combination with the dispatching knowledge graph. The operating cost of the train is obtained by combining the resource cost. The operating profit value is determined based on the passenger flow density coefficient and the dispatching knowledge graph. The economic value is determined by combining the operating profit value and the operating cost. The economic value factor of the train is obtained after normalizing the economic value.

[0023] The passenger flow density coefficient and resource status in the train data matrix are combined to analyze the train line load. The line load value is evaluated by combining the historical load data in the scheduling knowledge graph. The line load value is weighted and combined with the line importance in the scheduling knowledge graph to obtain the line impact factor.

[0024] The urgency factor, economic value factor and line impact factor are integrated to construct the evaluation vector of train scheduling.

[0025] As an optional implementation, the train priority generation strategy includes:

[0026] The fuzzy analytic hierarchy process is used to take the train priority as the target layer and the evaluation vector as the criterion layer. The importance between the criterion layers is characterized by fuzzy numbers and the fuzzy judgment matrix is determined.

[0027] The consistency check and correction of the fuzzy judgment matrix are performed, and the fuzzy weight vector of the evaluation vector is obtained by solving the fuzzy characteristic equation. The fuzzy weight vector is fuzzified to obtain the dimension weight of the evaluation vector.

[0028] The evaluation vector is weighted and summed with the dimension weight to obtain the comprehensive evaluation value of the train, and the train priority is generated according to the comprehensive evaluation value of the train.

[0029] As an optional implementation, the generation strategy of the scheduling solution includes:

[0030] Build a reinforcement learning model. The state space of reinforcement learning represents train priority and resource status. The action space of reinforcement learning represents train speed adjustment, train route selection, and resource allocation. The reward function of reinforcement learning represents on-time performance, energy consumption reduction rate, and conflict risk value.

[0031] Train the deep Q network through state space and action space, calculate the target Q value, and update the parameters of the deep Q network based on the mean square error loss function;

[0032] The action with the largest target Q value is selected through the deep Q network, and a scheduling plan is generated based on the action. The generated scheduling plan is evaluated by the time punctuality rate, energy consumption reduction rate and conflict risk value in the reward function to determine whether to adjust the reinforcement learning.

[0033] As an optional implementation, the verification sub-strategy of the scheduling instruction set includes:

[0034] Build a GPU parallel computing architecture and decompose the Monte Carlo simulation task into three layers of subtasks;

[0035] Based on historical operating data, the threshold is determined and the conflict probability, energy consumption deviation and delay margin of each scheduling scheme are analyzed in time series using a long short-term memory network.

[0036] The threshold is dynamically adjusted through the attention mechanism by combining meteorological environment data, passenger flow density coefficient and resource status;

[0037] The conflict probability, energy consumption deviation and delay margin of each scheduling scheme obtained by Monte Carlo simulation are fuzzy matched with the threshold through fuzzy comprehensive evaluation method;

[0038] Membership functions are set for conflict probability, energy consumption deviation and delay margin. The evaluation result of the scheduling scheme is determined by integrating the membership functions through the fuzzy synthesis operator. The evaluation result is compared with the evaluation threshold to determine whether the scheduling scheme has passed the verification, and the scheduling instruction set after verification is output.

[0039] As an optional implementation, the resource status update strategy includes:

[0040] Execute the dispatch instruction set and monitor the feedback data during the execution of the dispatch instruction set in real time. The feedback data includes the operation data of the power supply network, standby trains and signals, as well as the execution status data of the dispatch instruction set;

[0041] Extract indicator features of feedback data through convolutional neural networks;

[0042] The indicator features of the feedback data are learned through the long short-term memory network to predict the changing trend of resource status and update the resource status. At the same time, the node attributes of the topological graph, the parameters of the Bayesian network and the state transition probability are updated.

[0043] As an optional implementation, the resampling strategy for heterogeneous data includes:

[0044] Analyze the associations between track circuit signals, platform passenger flow images, and meteorological environment data through association rule mining algorithms, detect anomalies in these associations based on statistical methods, and analyze the causes of detected abnormal heterogeneous data.

[0045] Based on the results of anomaly detection and cause analysis, determine the sampling frequency adjustment mechanism and select the sampling method based on the type of abnormal heterogeneous data;

[0046] The quality of the resampled heterogeneous data is evaluated to determine whether to resample the heterogeneous data, and the resampled heterogeneous data is fused with the train data matrix to update the train data matrix.

[0047] In a second aspect, the present application provides a train dynamic resource allocation system, which includes: real-time acquisition of heterogeneous data, including track circuit signals, platform passenger flow images, and meteorological environment data; spatiotemporal alignment of the heterogeneous data through on-board edge nodes; processing the heterogeneous data; and obtaining the load rate of the power supply network, the equipment health and location of the standby train, and the failure probability of the signal to determine the resource status, and outputting a train data matrix.

[0048] An evaluation vector for train scheduling is constructed based on the train data matrix and the scheduling knowledge graph. The dimensional weights of the evaluation vector are dynamically adjusted based on the fuzzy analytic hierarchy process to generate train priorities. A scheduling plan is generated through reinforcement learning. Based on the train priorities in the scheduling plan, the power quota for each section of the power supply network is dynamically allocated, and resource allocation instructions are generated based on the resource status.

[0049] Monte Carlo simulation is performed on each scheduling scheme in parallel through GPU. The conflict probability, energy consumption deviation and delay margin of each scheduling scheme are verified based on the threshold. The scheduling instruction set that has passed the verification is output, and the scheduling instruction set is executed and the execution effect of the scheduling instruction set is monitored in real time to determine whether to resample heterogeneous data and update resource status.

[0050] Compared with the existing technology, the beneficial effects of this application are: through the real-time acquisition and processing of heterogeneous data, the generation of train priorities, and the generation and verification and execution of scheduling plans, the refined management of the entire train scheduling process is achieved. The various steps cooperate with each other and progress layer by layer, which can comprehensively consider various complex factors in the train operation process, change the limitations of traditional scheduling methods that rely on experience or single indicators, and significantly improve the scientificity, accuracy and timeliness of train scheduling decisions. In the face of complex and changeable operating scenarios, resources can be dynamically and efficiently allocated, train operation arrangements can be optimized, transportation efficiency and service quality can be improved, and the growing diversified needs of modern railway transportation can be effectively met.

[0051] It acquires and processes heterogeneous data in real time, determines the resource status, and outputs a train data matrix. By acquiring heterogeneous data such as track circuit signals, platform passenger flow images, and meteorological environment data, it comprehensively covers the key factors affecting train operation, enabling scheduling decisions to be based on a richer and more comprehensive information foundation, avoiding decision-making biases caused by data loss. By monitoring the power supply network, spare trains, and signal lights, it can accurately determine the resource status, deeply explore the potential relationships and changing trends between resources, and provide an accurate basis for the rational allocation of resources. The processed data is integrated to form a train data matrix, presenting various types of information in a structured form to facilitate subsequent analysis and application, and providing standardized and normalized data input for constructing evaluation vectors and generating scheduling plans.

[0052] An evaluation vector is constructed based on the train data matrix and the scheduling knowledge graph to generate train priorities. The evaluation vector is constructed from multiple dimensions such as urgency, economic value and line impact. Key factors such as train operation safety, economy and line resource utilization are comprehensively considered, making the generation of train priorities more comprehensive and objective, able to balance the needs of different aspects and optimize resource allocation. The dimensional weights of the evaluation vector are dynamically adjusted based on the fuzzy hierarchical analysis method. It can flexibly reflect the relative importance of each evaluation factor according to different operating scenarios and needs, making the train priority more in line with the actual situation and enhancing the flexibility and adaptability of scheduling decisions.

[0053] Scheduling plans are generated through reinforcement learning. Using reinforcement learning, based on train priority and resource status, through continuous trial and error and learning, the optimal scheduling plan is automatically searched for based on multiple objectives such as punctuality, energy consumption reduction, and conflict risk value. This achieves intelligent and automated generation of scheduling plans. Compared with traditional methods, it can more efficiently find better solutions and improve the quality of scheduling plans. The reward function covers multiple key indicators such as punctuality, energy consumption reduction, and conflict risk value, prompting reinforcement learning to comprehensively consider multiple important aspects of train operation when generating scheduling plans, effectively balancing the safety, punctuality, and economy of train operation, and improving the overall efficiency of railway transportation.

[0054] Through GPU parallelization, Monte Carlo simulation of scheduling plans is performed to verify and output scheduling instruction sets, and execution feedback is monitored. Leveraging the GPU parallel computing architecture, Monte Carlo simulation of scheduling plans significantly improves simulation computing efficiency and can quickly verify a large number of scheduling plans in a short period of time, ensuring that feasible scheduling plans are screened out within a limited time to meet the real-time requirements of train scheduling. Based on thresholds, the conflict probability, energy consumption deviation, and delay margin of scheduling plans are verified. Combined with fuzzy comprehensive evaluation methods and other means, scheduling plans are scientifically and comprehensively evaluated, effectively eliminating plans with potential risks or that do not meet the requirements, and ensuring the high feasibility and reliability of the output scheduling instruction sets. Feedback data from the execution of scheduling instruction sets is monitored in real time. Based on the execution results, it is determined whether to resample heterogeneous data and update resource status. This enables timely perception of changes in actual operation and dynamic adjustment and optimization of scheduling decisions based on new information, forming a closed-loop management system, continuously improving the accuracy and adaptability of train scheduling, and ensuring the stability and efficiency of train operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be derived from these drawings without inventive work. Among them:

[0056] Figure 1 A flow chart of a method for dynamic train scheduling provided by an embodiment of the present application;

[0057] Figure 2 A sub-strategy diagram for determining resource status of the train dynamic scheduling method provided in an embodiment of the present application;

[0058] Figure 3 This is a verification sub-strategy diagram of the scheduling instruction set of the train dynamic scheduling method provided in an embodiment of the present application. DETAILED DESCRIPTION

[0059] In order to make the objectives, technical solutions and advantages of the embodiments of the present application more obvious and easy to understand, the technical solutions in the embodiments of the present application are clearly and completely described below in conjunction with the drawings in the specification. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments.

[0060] Example 1

[0061] like Figure 1 As shown, a flow chart of a method for dynamic train scheduling is provided for an embodiment of the present application. The dynamic train scheduling method includes:

[0062] S1. Acquire heterogeneous data in real time. The heterogeneous data includes track circuit signals, platform passenger flow images, and meteorological environment data. The heterogeneous data are spatiotemporally aligned at the onboard edge node through the spatiotemporal attention mechanism and processed. At the same time, the power supply network, spare trains, and signal lights are monitored to determine the resource status. The train data matrix is output. The train data matrix includes signal feature vectors, passenger flow density coefficients, meteorological correction parameters, and resource status.

[0063] like Figure 2 As shown, the sub-strategies for determining resource status include:

[0064] Deploy sensors to monitor the power supply network, backup trains, and signals in real time to obtain resource data, including the load rate of the power supply network, the equipment health and location of backup trains, and the failure probability of signals.

[0065] A topology graph is established using the power supply network, standby trains, and signal machines as nodes. Node attributes represent resource data, and edges represent the coupling relationship between resource data. Node feature vectors of the topology graph are extracted using a graph convolutional network.

[0066] A Bayesian network is embedded in the topology graph, the node feature vector is input into the Bayesian network, and the Bayesian network is trained with historical operation data to determine the state transition probability between nodes, and the state transition probability between nodes is integrated to determine the resource status.

[0067] Train operation depends on the normal operation of resources such as the power supply network, spare trains and signal lights. Real-time acquisition of their status data is the basis for understanding the resource status and providing a basis for subsequent scheduling decisions. Voltage sensors, current sensors and load sensors are deployed at key nodes of the power supply network to obtain data such as the load rate of the power supply network in real time. Equipment health monitoring sensors are installed on spare trains to monitor equipment operating parameters to obtain equipment health. At the same time, positioning devices are used to obtain the train position. Fault detection modules are set at signal lights to monitor status information such as the failure probability of signal lights in real time. These sensors and monitoring modules transmit data to on-board edge nodes through wireless networks. This realizes the comprehensive and real-time collection of resource data, ensures the timeliness and accuracy of the data, enables the resource allocation system to perceive changes in resource status in a timely manner, provides raw data for constructing the resource topology map, and ensures that the node feature vectors subsequently extracted by the graph convolutional network can truly reflect the resource status.

[0068] There are mutual correlations and influences between resources. The relationship between resources can be intuitively displayed through a topological graph. The graph convolutional network can effectively extract the characteristic vectors of the implicit relationship, laying the foundation for accurate evaluation of resource status; the obtained resource data is used as node attributes, and the edges are determined according to the actual connection and mutual influence relationship of the resources. A directed weighted topological graph is constructed, and the topological graph is processed by the graph convolutional network. Through multi-layer convolution operations, the node and its neighborhood information are aggregated to extract the node characteristic vector containing the coupling relationship between resources; thereby, complex resource relationships are converted into topological structures and characteristic vectors that are easy for computers to process, improving the efficiency and accuracy of resource status analysis, being able to explore the potential influence patterns between resources, and providing effective input data for the Bayesian network, so that the Bayesian network can more accurately determine the state transition probability between nodes based on these characteristic vectors.

[0069] Bayesian networks are good at dealing with uncertainty problems. After training with historical data, they can predict resource status changes based on current feature vectors, providing more reliable resource status information for scheduling decisions. The Bayesian network is embedded in the constructed topology map, and the extracted node feature vectors are used as input. A large amount of historical operation data is collected, and the Bayesian network is trained using methods such as maximum likelihood estimation to determine the state transition probability between nodes. Finally, these state transition probabilities are integrated to comprehensively judge the current state of resources and future change trends. This enables dynamic prediction of resource status, early detection of potential problems, and the purchase of time for scheduling decisions, thereby improving the foresight and accuracy of scheduling. The determined resource status, as an important component of the train data matrix, affects the construction of subsequent train scheduling evaluation vectors and the generation of scheduling plans.

[0070] The output strategies of the train data matrix include:

[0071] Real-time acquisition of heterogeneous data, including track circuit signals, platform passenger flow images, and meteorological environment data. This data is then spatially and temporally aligned at the onboard edge node using a spatiotemporal attention mechanism, and then standardized.

[0072] The track circuit signal is extracted through a residual convolutional neural network, and the extracted features are mapped into a signal feature vector, which includes signal strength, signal frequency, and signal phase.

[0073] Perform semantic segmentation on platform passenger flow images and calculate the ratio of passenger pixel area to total platform area to obtain the initial passenger flow density. Combined with historical passenger flow data, the long short-term memory network is used to predict passenger flow trends, dynamically weighting the initial passenger flow density to generate a passenger flow density coefficient.

[0074] The random forest algorithm is used to analyze the impact of different meteorological environmental data on train operation data, braking distance, and energy consumption. Based on the analysis results, an impact weight is assigned to each meteorological environmental data. The standardized meteorological environmental data and the impact weight are weighted and summed to generate meteorological correction parameters.

[0075] Monitor the power supply network, standby trains and signals to determine resource status, integrate signal feature vectors, passenger flow density coefficients, meteorological correction parameters and resource status, and output a train data matrix.

[0076] Train operation is affected by a variety of heterogeneous data. These heterogeneous data differ in time and space and need to be aligned and standardized to provide a unified and accurate data foundation for subsequent analysis. Through sensors and cameras installed on trains and along the tracks, track circuit signals, platform passenger flow images and meteorological environment data are acquired in real time. A spatiotemporal attention mechanism is deployed at the onboard edge nodes. This spatiotemporal attention mechanism automatically learns the correlation features of different heterogeneous data in the spatiotemporal dimensions based on the timestamp and spatial location information of the data, achieving spatiotemporal alignment of the heterogeneous data. Then, normalization and other methods are used to standardize the heterogeneous data to eliminate dimensional effects. This solves the problems of spatiotemporal inconsistency and dimensional differences in heterogeneous data, making heterogeneous data comparable and compatible, improving the efficiency and accuracy of heterogeneous data processing, and providing high-quality data for the subsequent extraction of signal feature vectors, calculation of passenger flow density coefficients and meteorological correction parameters, ensuring the accuracy of these parameter calculations.

[0077] Track circuit signals contain key information about train operation, but the original track circuit signals are complex. The residual convolutional neural network can effectively extract their key features and form signal feature vectors for analyzing the train operation status. The standardized track circuit signals are input into the residual convolutional neural network. The residual convolutional neural network automatically extracts the deep features of the track circuit signals in the time series through a multi-layer residual block structure, including key information such as signal strength, frequency, and phase, and maps them into signal feature vectors. This can accurately extract key features from complex track circuit signals, avoid information loss, and provide accurate signal analysis basis for train scheduling. As part of the train data matrix, the signal feature vector participates in the construction of the subsequent train scheduling evaluation vector, affecting the determination of train priority.

[0078] Passenger flow density affects train scheduling strategies. Through semantic segmentation and combined with historical data prediction, it can more accurately reflect current and future passenger flow conditions and provide a basis for scheduling decisions. The semantic segmentation algorithm is used to process platform passenger flow images, distinguish between passengers and background, and calculate the ratio of passenger pixel area (that is, the number of passenger pixels) to the total platform area to obtain the initial passenger flow density. At the same time, historical passenger flow data is collected, and a long short-term memory network is used to predict passenger flow change trends. Based on the prediction results, the initial passenger flow density is dynamically weighted and corrected to generate a passenger flow density coefficient. Not only the current passenger flow situation is taken into account, but also historical data is combined to predict future trends, making the passenger flow density coefficient more forward-looking and accurate, providing support for the reasonable arrangement of train capacity. As a component of the train data matrix, the passenger flow density coefficient is used to construct a train scheduling evaluation vector, which affects the formulation of train priorities and scheduling plans.

[0079] The meteorological environment has a significant impact on train operation. By analyzing its relationship with train operating parameters and generating meteorological correction parameters, the impact of meteorological factors can be considered in scheduling decisions and scheduling plans can be optimized. Standardized meteorological environment data and train operation data are input into the random forest algorithm to analyze the impact of different meteorological factors on train operating parameters such as braking distance and energy consumption. Based on the analysis results, an impact weight is assigned to each meteorological environment data. The standardized meteorological environment data and the weight are weighted and summed to obtain the meteorological correction parameters. This quantifies the impact of the meteorological environment on train operation, enabling scheduling decisions to fully consider meteorological factors, improve the safety and efficiency of train operation, and reduce operating costs. As part of the train data matrix, the meteorological correction parameters participate in the construction of subsequent train scheduling evaluation vectors, affecting train priorities and the generation of scheduling plans.

[0080] Train scheduling needs to comprehensively consider multiple factors and integrate these key data to form a train data matrix to provide comprehensive data support for subsequent scheduling decisions; determine the resource status according to the logical steps of the above resource status determination, and splice and integrate the extracted signal feature vectors, calculated passenger flow density coefficients, generated meteorological correction parameters and determined resource status according to specific dimensions and sequences to form a train data matrix; thereby integrating various key data related to train operation into a matrix, making the data structure clear, facilitating the subsequent scheduling decision system to quickly obtain and analyze, and improving the efficiency and accuracy of scheduling decisions. The output train data matrix is used to construct the evaluation vector for train scheduling, which in turn affects the generation of train priorities and the formulation of scheduling plans.

[0081] S2. Construct an evaluation vector for train scheduling based on the train data matrix and the scheduling knowledge graph. The evaluation vector includes an urgency factor, an economic value factor, and a line impact factor. Dynamically adjust the dimensional weights of the evaluation vector based on the fuzzy analytic hierarchy process to generate train priorities.

[0082] The sub-strategies for constructing the evaluation vector include:

[0083] Normalize the train data matrix and associate it with the scheduling knowledge graph;

[0084] The train's operating status is analyzed by combining the signal feature vectors and resource status in the train data matrix. Historical scheduling data from the scheduling knowledge graph is extracted to assess the delay risk value under the train's operating status. The passenger impact value of train delays is assessed based on the passenger flow density coefficient and platform carrying capacity. The delay risk value and passenger impact value are weighted and summed to obtain the train's urgency factor.

[0085] Based on the energy consumption and meteorological correction parameters in the train data matrix, the energy consumption cost of the train is determined in combination with the dispatching knowledge graph. The operating cost of the train is obtained by combining the resource cost. The operating profit value is determined based on the passenger flow density coefficient and the dispatching knowledge graph. The economic value is determined by combining the operating profit value and the operating cost. The economic value factor of the train is obtained after normalizing the economic value.

[0086] The passenger flow density coefficient and resource status in the train data matrix are combined to analyze the train line load. The line load value is evaluated by combining the historical load data in the scheduling knowledge graph. The line load value is weighted and combined with the line importance in the scheduling knowledge graph to obtain the line impact factor.

[0087] The urgency factor, economic value factor and line impact factor are integrated to construct the evaluation vector of train scheduling.

[0088] The data dimensions and scales in the train data matrix are different. Normalization can eliminate the differences and facilitate comparison and analysis. The dispatching knowledge graph contains information such as historical dispatching experience and rules. Associating the two can provide more comprehensive data support and knowledge reference for the construction of the evaluation vector. The minimum-maximum normalization method is used to map the signal feature vectors, passenger flow density coefficients, meteorological correction parameters, and resource status in the train data matrix to a unified interval. Using knowledge graph embedding technology, the key data in the train data matrix is semantically matched with the entities and relationships in the dispatching knowledge graph, and a data association index is established to achieve rapid interactive calling of data and knowledge. This ensures data consistency and comparability, improves data analysis efficiency, integrates historical experience and real-time data, and mines potential information, providing a richer and more accurate basis for the subsequent calculation of evaluation factors. It also provides a standardized and associated knowledge data foundation for the calculation of urgency factors, economic value factors, and line impact factors, making the calculation of each factor more accurate and thus ensuring the reliability of the evaluation vector.

[0089] Combine the signal feature vectors and resource status in the train data matrix to analyze the train operation status, extract the historical scheduling data of the scheduling knowledge graph to evaluate the delay risk value under the train operation status, and evaluate the passenger impact value under the train delay based on the passenger flow density coefficient and platform carrying capacity. The delay risk value and passenger impact value are weighted and summed to obtain the train urgency factor.

[0090] Train delays can affect operational order and passenger experience. By comprehensively considering factors such as operating status, historical data, and passenger flow, the delay risk and passenger impact are assessed, accurately quantifying the train's urgency and providing a basis for dispatching decisions to prioritize emergency trains. Signal feature vectors are analyzed to determine whether track circuit signals are abnormal. The normal operation of power supply and signal equipment is determined based on resource status. A comprehensive assessment of the train's operating status is then conducted. Historical delay data from similar operating conditions is extracted from the dispatching knowledge graph. Case-based reasoning techniques are used to predict the current train's delay risk. The impact of passengers stranded on the platform itself is determined based on the passenger flow density coefficient and platform capacity. The passenger impact of train delays is then assessed. Finally, weights are determined based on historical data statistics. The delay risk and passenger impact are weighted and summed to produce an urgency factor. This comprehensive consideration of multiple factors allows for an accurate assessment of train urgency, enabling dispatching decisions to prioritize emergency situations, reduce the adverse impact of delays on operations and passengers, and improve railway transportation's emergency response capabilities. The urgency factor, as an important component of the evaluation vector, directly influences the generation of train priorities. Trains with higher urgency levels are given higher priority in dispatching.

[0091] Railway operations need to consider costs and benefits. Accurately assessing the economic value of trains will help optimize resource allocation, improve operational efficiency, and achieve the economic goals of railway transportation. Based on energy consumption and meteorological correction parameters, combined with the empirical consumption relationship of energy consumption under different meteorological conditions in the scheduling knowledge graph, the train energy consumption cost is estimated, and resource costs such as energy consumption cost, equipment maintenance cost, and labor cost are summarized to obtain the train operating cost. Based on the passenger flow density coefficient, combined with the ticket sales pattern under different passenger flow conditions in the scheduling knowledge graph, the operating income value is predicted, and the operating cost is subtracted from the operating income value to obtain the economic value. The normalization method is then used to map the economic value to a specific interval to obtain the economic value factor. This quantifies the economic indicators of train operation and provides an economic reference for railway operation decision-making, helping to reasonably arrange train operations, reduce costs, increase revenue, and maximize economic benefits. The economic value factor is one of the key elements of the evaluation vector and participates in the train priority calculation, so that scheduling decisions take into account both urgency and economic factors and optimize resource allocation.

[0092] Train operation generates load on the line, and different lines have different importance. Comprehensively considering the line load and importance can accurately assess the impact of train operation on the line, rationally plan train operation, and ensure efficient line operation. The pressure of passenger flow on the line is judged based on the passenger flow density coefficient, and the equipment carrying capacity is analyzed in combination with the resource status to evaluate the train line load. Historical load data is extracted from the scheduling knowledge graph, and the current line load value is predicted using the time series analysis method. Relevant indicators of line importance in the scheduling knowledge graph are obtained, such as the economic status of the city connected by the line and the proportion of passenger flow. The weight is determined according to the hierarchical analysis method, and the line load value and line importance are weightedly combined to obtain the line impact factor. In this way, the impact of trains on the line is comprehensively assessed, so that scheduling decisions can reasonably arrange trains according to the actual situation of the line, avoid line congestion, and improve the overall operation efficiency and reliability of the line. As a component of the evaluation vector, the line impact factor plays a role in the generation of train priorities, affecting the formulation of scheduling plans, so that train operation is more in line with the actual needs of the line.

[0093] A single factor cannot fully reflect train scheduling needs. Integrating multiple factors to construct an evaluation vector can comprehensively evaluate train scheduling from multiple dimensions and provide a comprehensive basis for generating reasonable train priorities. The calculated urgency factor, economic value factor, and line impact factor are combined in a preset order and format to form a multidimensional vector, namely the evaluation vector for train scheduling. This achieves a comprehensive quantitative evaluation of train scheduling, provides comprehensive and accurate data support for subsequent train priority generation, and makes scheduling decisions more scientific and reasonable. The evaluation vector is the direct basis for generating train priorities. Its accuracy and comprehensiveness determine the rationality of train priorities, which in turn affects the formulation and implementation of scheduling plans.

[0094] The strategies for generating train priorities include:

[0095] The fuzzy analytic hierarchy process is used to take the train priority as the target layer and the evaluation vector as the criterion layer. The importance between the criterion layers is characterized by fuzzy numbers and the fuzzy judgment matrix is determined.

[0096] The consistency check and correction of the fuzzy judgment matrix are performed, and the fuzzy weight vector of the evaluation vector is obtained by solving the fuzzy characteristic equation. The fuzzy weight vector is fuzzified to obtain the dimension weight of the evaluation vector.

[0097] The evaluation vector is weighted and summed with the dimension weight to obtain the comprehensive evaluation value of the train, and the train priority is generated according to the comprehensive evaluation value of the train.

[0098] The factors affecting train scheduling are complex and ambiguous, making them difficult to accurately handle with traditional methods. The fuzzy hierarchical analysis method can effectively process fuzzy information. By constructing a hierarchical structure and a fuzzy judgment matrix, the relative importance of each evaluation factor is quantified, laying the foundation for generating accurate train priority. Train priority is set as the target layer, and the urgency factor, economic value factor, and line impact factor in the evaluation vector are used as the criterion layer. Based on experience and actual conditions, triangular fuzzy numbers or trapezoidal fuzzy numbers are used to compare the importance of each factor in the criterion layer, and a fuzzy judgment matrix is constructed. This fully considers the fuzziness and uncertainty in train scheduling decisions, more realistically reflects the actual situation, makes the quantification of the importance of each evaluation factor more accurate, and improves the scientific nature of train priority generation. The fuzzy judgment matrix is the basis for calculating the dimensional weights of the evaluation vector. Its accuracy directly affects the weight calculation results, and thus affects the rationality of train priority.

[0099] Ensure the consistency of the fuzzy judgment matrix, avoid logical contradictions, ensure the accuracy of weight calculation, solve the fuzzy characteristic equation to obtain the fuzzy weight vector and process it, so that the fuzzy information can be converted into accurate weights that can be used for calculation; use fuzzy consistency index and fuzzy random consistency index to check the consistency of the fuzzy judgment matrix. If the consistency requirements are not met, the matrix is corrected through expert feedback. Using fuzzy mathematics theory, the fuzzy characteristic equation is solved to obtain the fuzzy weight vector of the evaluation vector, and then the fuzzy weight vector is converted into accurate dimensional weights through defuzzification methods such as the centroid method and the maximum membership method; thereby ensuring the rationality of the fuzzy judgment matrix and the accuracy of weight calculation, so that the train priority generation is based on reliable weight distribution, improving the reliability and effectiveness of scheduling decisions, and the determined dimensional weights are used to calculate the comprehensive evaluation value of the train, which directly affects the train priority sorting. Reasonable weights can make the priority more in line with actual scheduling needs.

[0100] The evaluation vector and the dimension weight are combined through weighted summation to obtain a comprehensive evaluation value, which is used as a basis to generate train priority, realizing the transformation from multi-dimensional evaluation to specific priority sorting, and providing a clear priority reference for scheduling decisions; the urgency factor, economic value factor and line impact factor in the evaluation vector are multiplied by the corresponding dimension weights respectively, and then the products are summed to obtain a comprehensive evaluation value of the train, and the trains are sorted in descending order according to the comprehensive evaluation value to generate the train priority; thereby, the complex multi-dimensional evaluation results are converted into an intuitive priority sorting, providing dispatchers with a clear decision-making basis, making scheduling decisions more efficient and accurate, and optimizing the allocation of train scheduling resources. The generated train priority is an important basis for generating scheduling plans through reinforcement learning. Trains with high train priority will be given priority in the generation of scheduling plans, affecting the overall planning and implementation of the scheduling plan.

[0101] S3. Generate a scheduling plan through reinforcement learning. The state space of reinforcement learning represents train priority and resource status, and the reward function of reinforcement learning represents time punctuality, energy consumption reduction rate and conflict risk value.

[0102] The generation strategies of scheduling plans include:

[0103] Build a reinforcement learning model. The state space of reinforcement learning represents train priority and resource status. The action space of reinforcement learning represents train speed adjustment, train route selection, and resource allocation. The reward function of reinforcement learning represents on-time performance, energy consumption reduction rate, and conflict risk value.

[0104] Train the deep Q network through state space and action space, calculate the target Q value, and update the parameters of the deep Q network based on the mean square error loss function;

[0105] The action with the largest target Q value is selected through the deep Q network, and a scheduling plan is generated based on the action. The generated scheduling plan is evaluated by the time punctuality rate, energy consumption reduction rate and conflict risk value in the reward function to determine whether to adjust the reinforcement learning.

[0106] In order to enable train scheduling to adapt to the complex and changing operating environment, the train scheduling process is simulated through reinforcement learning. With train priority and resource status as the core decision-making basis, the optimal scheduling plan is automatically learned and generated based on key indicators such as time punctuality, energy consumption reduction rate and conflict risk value. Reinforcement learning is constructed to integrate train priority and resource status information into a state space. The action space clearly covers actual scheduling actions such as train speed adjustment (such as acceleration, deceleration and constant speed), train path selection (choosing among multiple optional routes) and resource allocation (dispatching spare trains and adjusting power supply network resources, etc.). At the same time, a reward function is defined. Taking punctuality, energy consumption reduction, and conflict risk value as reward evaluation indicators, and clarifying the reward or penalty rules corresponding to each indicator in different situations, this guides the reinforcement learning process; thus, a reinforcement learning model that meets the actual needs of train scheduling is constructed, providing a structured learning framework for the subsequent automatic generation of efficient scheduling plans, enabling the model to learn and optimize based on actual operating data and target indicators. At the same time, the determined state space, action space, and reward function provide clear rules and goals for the training of the deep Q network, enabling the deep Q network to effectively learn the train scheduling problem and generate reasonable scheduling plans.

[0107] The deep Q-network continuously learns within a given state and action space, evaluating the value of different actions by calculating target Q-values and adjusting network parameters based on the mean squared error loss function, gradually optimizing the scheduling strategy to adapt to complex train scheduling scenarios. The constructed state space is sequentially input into the deep Q-network, which evaluates each action in the action space based on the current state and outputs the corresponding Q-value. Simultaneously, an experience replay mechanism is employed to store the state, action, reward, and next state from each decision-making process in an experience pool. A batch of data is periodically randomly sampled from the experience pool, and the target Q-value is calculated using the Bellman equation. The mean squared error loss function is then used to calculate the error between the predicted Q-value and the target Q-value. The deep Q-network parameters are then updated using the backpropagation algorithm, continuously optimizing the network's ability to assess action value. This training method allows the deep Q-network to learn from a large amount of historical experience, avoiding learning bias caused by the correlation of consecutive decisions. This effectively improves the network's learning efficiency and accuracy in train scheduling strategies, gradually approaching the optimal scheduling solution. The trained and optimized deep Q-network can more accurately evaluate the value of different actions under various states, providing a reliable basis for selecting the action with the highest target Q-value, thereby generating a more reasonable scheduling solution.

[0108] After the deep Q network is trained, it is necessary to select the optimal action from the evaluated actions to generate a specific scheduling plan. At the same time, the plan is evaluated through the reward function to determine whether the current reinforcement learning strategy is effective so that timely adjustments and optimizations can be made. For the state space composed of the current train priority and resource status, the deep Q network outputs the Q value corresponding to each action, and selects the action with the largest Q value. If the action corresponding to the maximum Q value is to adjust the running path of a certain train, the corresponding train path adjustment plan is generated according to the action, and a complete scheduling plan is generated in combination with the resource allocation requirements. The scheduling plan is then simulated and executed. According to the time punctuality rate, energy consumption reduction rate and conflict risk value generated during the actual operation, the reward function is used to calculate the response. If the reward value does not meet expectations or is significantly different from historically excellent solutions, it is considered that current reinforcement learning needs to be adjusted, and the parameters of the deep Q network should be re-optimized or the rules of the reward function should be adjusted. This achieves a complete closed loop from learning to decision-making. By selecting the action with the largest Q value to generate a scheduling plan and evaluating it based on the reward function, problems in the scheduling plan and learning strategy can be discovered in a timely manner, ensuring that the generated scheduling plan has high practicality and optimization. The evaluated scheduling plan will be used as a candidate plan for simulation. Through GPU parallel Monte Carlo simulation and threshold verification, the final scheduling instruction set that meets the requirements is further screened out to ensure that the scheduling plan actually executed is safe and efficient.

[0109] S4. Perform Monte Carlo simulation on each scheduling scheme in parallel through GPUs. Verify the conflict probability, energy consumption deviation, and delay margin of each scheduling scheme based on thresholds to output a verified scheduling instruction set. The scheduling instruction set includes train path adjustment and resource allocation instructions. Execute the scheduling instruction set and monitor the feedback data during the execution of the scheduling instruction set in real time to determine whether to resample heterogeneous data and update resource status.

[0110] like Figure 3 As shown, the verification sub-strategies for the scheduling instruction set include:

[0111] Build a GPU parallel computing architecture and decompose the Monte Carlo simulation task into three layers of subtasks;

[0112] Based on historical operating data, the threshold is determined and the conflict probability, energy consumption deviation and delay margin of each scheduling scheme are analyzed in time series using a long short-term memory network.

[0113] The threshold is dynamically adjusted through the attention mechanism by combining meteorological environment data, passenger flow density coefficient and resource status;

[0114] The conflict probability, energy consumption deviation and delay margin of each scheduling scheme obtained by Monte Carlo simulation are fuzzy matched with the threshold through fuzzy comprehensive evaluation method;

[0115] Membership functions are set for conflict probability, energy consumption deviation and delay margin. The evaluation result of the scheduling scheme is determined by integrating the membership functions through the fuzzy synthesis operator. The evaluation result is compared with the evaluation threshold to determine whether the scheduling scheme has passed the verification, and the scheduling instruction set after verification is output.

[0116] There are numerous train scheduling scenarios, and Monte Carlo simulations are computationally intensive, while traditional computing methods are inefficient. Building a GPU parallel computing architecture and decomposing tasks can fully utilize the parallel computing capabilities of the GPU, significantly shortening simulation time and meeting the real-time requirements of train scheduling. Using the CUDA programming model to build the GPU parallel computing architecture, the Monte Carlo simulation tasks are functionally divided into three levels: scenario-level, scenario-level, and computation-level subtasks. The scenario-level task is responsible for assigning different scheduling scenarios to different GPU thread blocks. Within each thread block, the scenario-level task launches multiple threads in parallel to simulate different operating scenarios, such as weather conditions and passenger flow. The computation-level task, on the other hand, involves each thread performing specific calculations, such as train trajectory and resource consumption calculations. Shared memory and synchronization mechanisms enable efficient data interaction and collaboration between threads. This significantly improves Monte Carlo simulation efficiency, ensuring that a large number of scheduling scenarios can be processed in a short period of time. This supports the rapid screening of feasible scheduling scenarios, improves the responsiveness of train scheduling, and rapidly completes simulation calculations, providing timely data for subsequent threshold-based scheduling plan verification, ensuring a smooth verification process and accelerating the generation of the entire scheduling instruction set.

[0117] Historical operating data contains the laws and experience of train scheduling under different conditions. The thresholds determined based on this data can be used as a benchmark for evaluating the feasibility of scheduling plans. The long-short-term memory network can effectively process time series data and analyze the changing trends of scheduling plan-related indicators, providing a basis for accurate evaluation of plans. A large amount of historical train operation data is collected, including conflict probability, energy consumption deviation, and delay margin data under different time periods, weather conditions, and passenger flow conditions. Statistical analysis methods are used, combined with expert experience, to determine the reasonable threshold ranges for each indicator. The conflict probability, energy consumption deviation, and delay margin data for each scheduling plan are organized into a chronological sequence and input into the long-short-term memory network. The long-short-term memory network automatically learns the time dependencies and trend characteristics in the data. The thresholds determined based on historical data are consistent with actual operating conditions. The long-short-term memory network can accurately capture the changing trends of scheduling plan indicators, improve the accuracy and reliability of scheduling plan evaluations, prevent unreasonable plans from entering actual scheduling, and provide basic data and analysis results for dynamic threshold adjustment and fuzzy matching, making the subsequent verification process more scientific and reasonable, and ensuring that the selected scheduling plans meet actual operating requirements.

[0118] The train operating environment is complex and changeable, with weather, passenger flow and resource status changing in real time. Fixed thresholds cannot adapt to different situations. By dynamically adjusting the thresholds through the attention mechanism combined with multi-source data, the thresholds can be made more in line with actual operating conditions and ensure the accuracy of scheduling plan evaluation. The real-time meteorological environment data, passenger flow density coefficient and resource status are integrated with historical threshold data. The attention mechanism is used to automatically assign weights based on the different degrees of impact of each data on train operation, highlighting key influencing factors. For example, in extreme weather, meteorological environment data is given a higher weight, and the thresholds of indicators such as conflict probability and energy consumption deviation are adjusted accordingly. This achieves dynamic adaptive adjustment of the thresholds, enabling scheduling plan evaluation to better adapt to different operating environments, improve the flexibility and accuracy of the evaluation, and ensure that the scheduling plan can meet operating requirements under various circumstances. The adjusted thresholds provide more accurate standards for fuzzy matching and plan evaluation, making subsequent verification results more reliable and helping to screen out better scheduling plans.

[0119] There is a certain degree of ambiguity in the scheduling scheme evaluation indicators, and precise matching is difficult to reflect the actual situation. The fuzzy comprehensive evaluation method can effectively process fuzzy information and more reasonably evaluate the degree of compliance of the scheduling scheme with the threshold through fuzzy matching. Fuzzy sets and membership functions are defined for indicators such as conflict probability, energy consumption deviation and delay margin. For example, the conflict probability is divided into fuzzy levels such as "low", "medium" and "high", and the membership function corresponding to each level is determined. The scheme indicator values obtained by Monte Carlo simulation are substituted into the membership function, and its membership at different fuzzy levels is calculated. Then, fuzzy matching is performed according to the fuzzy range set by the threshold to determine the degree of compliance of the scheme indicators with the threshold. This more realistically reflects the ambiguity and uncertainty in the scheduling scheme evaluation, improves the rationality and credibility of the evaluation results, avoids misjudgment caused by precise matching, ensures that the evaluation of the scheduling scheme is more in line with reality, and provides a basis for determining the scheduling scheme evaluation results, making it more scientific to judge whether the scheme has passed verification by setting evaluation thresholds, and ensuring that the output scheduling instruction set has high feasibility.

[0120] Taking into account multiple evaluation indicators, the fuzzy evaluation results of each indicator are synthesized by setting the membership function and fuzzy synthesis operator to obtain a comprehensive scheme evaluation result, which is compared with the evaluation threshold to determine whether the scheme is feasible and ensure the output scheduling instruction set is safe and reliable; according to actual operation requirements and experience, appropriate membership functions are set for conflict probability, energy consumption deviation and delay margin respectively, and appropriate fuzzy synthesis operators are selected, such as the maximum-minimum synthesis operator and the weighted average synthesis operator, and the membership of each indicator is comprehensively calculated to obtain the comprehensive evaluation result of the scheduling scheme, and the evaluation threshold is set to comprehensively evaluate the scheduling scheme. The results are compared with the evaluation threshold. If the evaluation result is greater than or equal to the evaluation threshold, the scheduling plan is judged to have passed the verification, and a scheduling instruction set containing train path adjustment and resource allocation instructions is output. The scheduling plan is comprehensively evaluated from multiple dimensions, and the impact of various indicators is fully considered. Through scientific evaluation methods, it is ensured that the selected scheduling plan meets actual operation needs and improves the safety and efficiency of train scheduling. The output scheduling instruction set that has passed the verification will enter the execution stage. At the same time, the feedback data during the execution process will be used to determine whether it is necessary to resample heterogeneous data and update resource status, providing a reference for subsequent scheduling decisions.

[0121] Resource status update strategies include:

[0122] Execute the dispatch instruction set and monitor the feedback data during the execution of the dispatch instruction set in real time. The feedback data includes the operation data of the power supply network, standby trains and signals, as well as the execution status data of the dispatch instruction set;

[0123] Extract indicator features of feedback data through convolutional neural networks;

[0124] The indicator features of the feedback data are learned through the long short-term memory network to predict the changing trend of resource status and update the resource status. At the same time, the node attributes of the topological graph, the parameters of the Bayesian network and the state transition probability are updated.

[0125] During the execution of the dispatch instruction set, the resource status will change. Real-time monitoring feedback data can timely grasp the resource operation status, provide an accurate basis for resource status updates, and ensure the effectiveness of dispatch decisions. Various sensors and monitoring equipment are deployed in the train operation system to collect real-time operation data such as the voltage, current and load rate of the power supply network, the equipment status and location of the standby train, the working status of the signal light, as well as execution status data such as the progress and completion of the dispatch instruction execution. The data is transmitted to the dispatch center server in real time via the wireless network. This enables real-time monitoring of the dispatch instruction execution process and resource operation status, and timely acquisition of accurate data, providing reliable information for subsequent resource status analysis and updates, ensuring train operation safety and smooth dispatching. The feedback data obtained provides raw data for subsequent extraction of indicator features and prediction of resource status change trends. It is the basis for resource status updates and determines the accuracy and timeliness of updates.

[0126] The feedback data is large and complex, and it is difficult to process it directly. The convolutional neural network has a powerful feature extraction capability and can automatically extract key indicator features from the data to provide effective information for analyzing resource status changes. The obtained feedback data is preprocessed and converted into a format suitable for convolutional neural network input. The convolutional neural network is constructed, and a suitable convolution layer, pooling layer and fully connected layer structure are designed. The preprocessed data is input into the network, and the local features in the data are extracted through the convolution kernel of the convolution layer. The pooling layer performs dimensionality reduction, and the fully connected layer integrates the features and outputs the extracted indicator feature vectors, such as power supply network abnormality features and standby train failure features. In this way, key indicator features can be extracted from complex feedback data efficiently and accurately, reducing the amount of data processing, highlighting important information in the data, providing strong support for subsequent prediction of resource status change trends, and improving the efficiency of resource status analysis. The extracted indicator feature vectors are used as input to the long and short-term memory network, providing a basis for learning data features and predicting resource status change trends, affecting the accuracy of the prediction results.

[0127] Resource status changes have time series characteristics. LSTM networks can effectively process time series data, learn data features, and predict changing trends. Based on the predicted results, resource status and related model parameters are updated, ensuring real-time and accurate resource status information, providing a reliable basis for scheduling decisions. The indicator feature vectors extracted by the convolutional neural network are organized into a chronological sequence and input into the LSTM network. The LSTM network automatically learns the long-term and short-term dependencies in the data through forget gates, input gates, and output gates, predicting the changing trends of resource status over a period of time. Based on the predicted results, resource status information is updated. For example, the status of predicted faulty equipment is marked as abnormal. The corresponding node attributes in the resource topology are also updated, and the Bayesian network is retrained to adjust the state transition probabilities between nodes. This allows for accurate prediction of resource status changing trends and timely updates of resource status and related model parameters, enabling proactive response to resource status changes, optimizing scheduling decisions, and improving the reliability and stability of train operations. The updated resource status information influences subsequent train scheduling decisions, including the sub-strategy for determining resource status and the sub-strategy for constructing evaluation vectors, ensuring that the entire scheduling process is based on the latest and most accurate resource status information.

[0128] Resampling strategies for heterogeneous data include:

[0129] Analyze the associations between track circuit signals, platform passenger flow images, and meteorological environment data through association rule mining algorithms, detect anomalies in these associations based on statistical methods, and analyze the causes of detected abnormal heterogeneous data.

[0130] Based on the results of anomaly detection and cause analysis, determine the sampling frequency adjustment mechanism and select the sampling method based on the type of abnormal heterogeneous data;

[0131] The quality of the resampled heterogeneous data is evaluated to determine whether to resample the heterogeneous data, and the resampled heterogeneous data is fused with the train data matrix to update the train data matrix.

[0132] Heterogeneous data has inherent correlations. Analyzing these correlations helps uncover data patterns. Abnormal heterogeneous data can affect the accuracy of scheduling decisions. Through anomaly detection and cause analysis, problematic data can be promptly identified and their causes can be determined, ensuring data quality. Association rule mining algorithms such as Apriori are used to analyze track circuit signals, platform passenger flow images, and meteorological environment data. Frequent item sets and association rules between the data are mined. Statistical methods (such as the 3σ principle and boxplots) are then used to detect anomalies in the correlations and determine whether the data deviates from normal patterns. For detected abnormal heterogeneous data, cause analysis is conducted based on information such as data collection time, equipment status, and external environment to determine whether it is caused by data collection errors, equipment failures, or changes in actual conditions. This allows for a deeper understanding of the correlations between heterogeneous data, timely detection of abnormal data, and identification of its causes. This prevents misleading scheduling decisions from abnormal data, improves data reliability and effectiveness, and ensures accurate scheduling. Based on the results of anomaly detection and cause analysis, it is determined whether the sampling strategy needs to be adjusted. This provides a basis for the subsequent determination of the sampling frequency adjustment mechanism and the selection of the sampling method, ensuring the targeted and effective resampling operation.

[0133] Different abnormal situations and data types have different requirements for sampling frequency and method. Reasonable adjustment of sampling frequency and selection of sampling method can improve data collection efficiency and reduce data processing costs while ensuring data quality. If the abnormality is caused by large data fluctuations or important events, increase the sampling frequency. If it is a data collection error, adjust the sampling frequency according to the actual situation after correcting the error. For continuous data, such as track circuit signal strength, use equal interval sampling or stratified sampling. For discrete data, such as the number of passengers on the platform, use random sampling or systematic sampling. For image data, such as platform passenger flow images, use feature-based sampling. This achieves adaptive adjustment of the sampling strategy, selects the optimal sampling method and frequency for different situations, improves the accuracy and efficiency of data acquisition, obtains more valuable data, and meets the data needs of train dispatching. The quality of the resampled data directly affects the subsequent quality assessment and data fusion. The appropriate sampling method and frequency can ensure the reliability of the resampled data and provide high-quality data for updating the train data matrix.

[0134] The resampled data still has quality problems. Quality assessment can ensure that the data meets the requirements. The qualified resampled data is merged with the train data matrix, and the matrix information is updated so that the train data matrix always reflects the latest and accurate data, providing a reliable basis for scheduling decisions. A data quality assessment indicator system is established to evaluate the resampled data from the aspects of data integrity, accuracy, consistency and timeliness. The entropy weight method and other methods are used to determine the weight of each indicator, calculate the comprehensive data quality score, and set the quality threshold. If the comprehensive score is less than the quality threshold, the resampling operation is performed again. If it is greater than the quality threshold, the resampled data is merged with the train data matrix and the corresponding data in the matrix is updated. In this way, the quality of the resampled data is guaranteed, low-quality data is prevented from entering the train data matrix and affecting scheduling decisions, and the train data matrix is updated in a timely manner so that scheduling decisions are based on the latest and accurate data, improving the scientificity and accuracy of scheduling decisions. The updated train data matrix will be used for the construction of subsequent evaluation vectors and the generation of scheduling plans, affecting the accuracy and effectiveness of the entire train scheduling process.

[0135] Example 2

[0136] An embodiment of the present application provides a train dynamic resource allocation system, which includes a resource perception module, a flexible allocation module and a feedback control module.

[0137] The resource perception module is used to acquire heterogeneous data in real time. The heterogeneous data includes track circuit signals, platform passenger flow images, and meteorological environment data. The heterogeneous data is aligned in time and space through the on-board edge nodes and processed. At the same time, the load rate of the power supply network, the equipment health and location of the standby train, and the failure probability of the signal are obtained to determine the resource status and output the train data matrix.

[0138] At a large railway hub, sensors deployed along the tracks and on-board equipment on trains are working continuously. Track circuit signal sensors collect signal data multiple times per second, high-definition cameras capture platform passenger flow images in real time, and meteorological monitoring equipment simultaneously collects meteorological environmental data such as wind speed, rainfall, and temperature. At the same time, current and voltage sensors installed at key nodes in the power supply network monitor the load rate in real time. Various sensors on standby trains collect equipment health data, the positioning system provides real-time feedback on the train's position, and the signal light's built-in detection module monitors the probability of failure.

[0139] After these data are transmitted to the onboard edge node, the spatiotemporal attention mechanism within the node begins to operate. Taking a train about to enter the station as an example, the onboard edge node receives the circuit signal of the train's current track, the real-time passenger flow image of the platform, and the meteorological data at that time. The spatiotemporal attention mechanism quickly analyzes the timestamp and spatial position of the data, accurately aligns different types of data in the spatiotemporal dimensions, and standardizes the data, laying the foundation for subsequent in-depth data analysis. At the same time, it outputs a train data matrix containing resource status.

[0140] The flexible allocation module is used to construct an evaluation vector for train scheduling based on the train data matrix and the scheduling knowledge graph, dynamically adjust the dimensional weights of the evaluation vector based on the fuzzy hierarchical analysis method, generate train priorities, generate scheduling plans through reinforcement learning, and dynamically allocate the power quota of each section in the power supply network according to the train priorities in the scheduling plan, and generate resource allocation instructions based on the resource status.

[0141] After the resource allocation system obtains the train data matrix output by the resource perception module, it first normalizes it to ensure the comparability of different types of data in the matrix. For example, the signal strength, frequency, and phase data in the signal feature vector, as well as the passenger flow density coefficient and meteorological correction parameters, are uniformly mapped to a specific interval. Then, using data association technology, the various data in the train data matrix are linked to the relevant entities and relationships in the scheduling knowledge graph. For example, according to the train operation route, the historical scheduling cases and equipment maintenance records of the route in the knowledge graph are associated.

[0142] Combining the signal feature vectors and resource status in the train data matrix, the resource allocation system conducts an in-depth analysis of the train's operating status. If the signal feature vector shows that a certain track circuit signal is abnormal, and the resource status indicates that the failure probability of the signal in the area has increased, the resource allocation system extracts similar historical scheduling data from the scheduling knowledge graph and evaluates the delay risk value of the train in the current state. At the same time, based on the passenger flow density coefficient and platform carrying capacity, the passenger impact value caused by the train delay is calculated. For example, when the passenger flow density coefficient of a platform shows that it is at a peak state and the platform carrying capacity is close to saturation, if the train is delayed, the passenger impact value will increase significantly. Finally, the delay risk value and the passenger impact value are weighted and summed according to the preset weights to obtain the train's urgency factor.

[0143] Based on energy consumption and meteorological correction parameters, combined with the empirical consumption relationship of energy consumption under different meteorological conditions in the scheduling knowledge graph, the resource allocation system estimates the energy consumption cost of the train. For example, in windy weather, the additional energy consumption cost is calculated based on the meteorological correction parameters and energy consumption. At the same time, the resource costs such as equipment maintenance and manpower are summarized to obtain the operating cost of the train. Then, based on the passenger flow density coefficient and the ticket sales patterns under different passenger flow conditions in the scheduling knowledge graph, the operating income value of the train is predicted. The operating income value is subtracted from the operating cost to obtain the economic value, which is then normalized to generate an economic value factor.

[0144] The resource allocation system combines the passenger flow density coefficient and resource status in the train data matrix to analyze the load situation of the line where the train is located. If the passenger flow density coefficient is high and the resource status shows that the equipment load on the line is large, the line load increases. At the same time, the historical load data in the scheduling knowledge graph is referred to to evaluate the current line load value. Then, based on the line importance indicators in the knowledge graph, such as the economic status and passenger flow proportion of the city connected by the line, the line load value and the line importance are weighted and combined to obtain the line impact factor.

[0145] The calculated urgency factor, economic value factor and line impact factor are integrated to construct an evaluation vector for train scheduling. Then, the fuzzy hierarchical analysis method is used, with train priority as the target layer and the evaluation vector as the criterion layer. The importance of each factor in the criterion layer is compared pairwise based on experience and actual conditions. The comparison results are represented by fuzzy numbers, and the fuzzy judgment matrix is determined. After consistency check and correction of the fuzzy judgment matrix, the fuzzy characteristic equation is solved to obtain the fuzzy weight vector of the evaluation vector. Fuzzification processing is then performed to obtain the dimensional weight of the evaluation vector. Finally, the evaluation vector and the dimensional weight are weighted and summed to obtain the comprehensive evaluation value of the train. The train priority is generated based on the comprehensive evaluation value. For example, through calculation, the comprehensive evaluation value of a train is higher, and its priority is set to a higher level in the current scheduling task.

[0146] The generated train priority and resource status are used as the state space of reinforcement learning, and operations such as train speed adjustment, train route selection, and resource allocation are defined as the action space. On-time performance, energy consumption reduction rate, and conflict risk value are used as reward functions to construct reinforcement learning. For example, when a train has a higher priority and the resource status of the line ahead allows it, reinforcement learning can choose to adjust the train speed to improve the punctuality rate as an action, and give corresponding rewards based on the actual effect.

[0147] The deep Q network is trained using the constructed state space and action space. During the training process, the resource allocation system stores the state, action, reward and next state of each train scheduling decision in the experience pool, randomly extracts a batch of data from the experience pool at regular intervals, calculates the target Q value according to the Bellman equation, and then updates the parameters of the deep Q network based on the mean square error loss function. After multiple iterative training, the deep Q network can accurately evaluate the value of different actions under various states.

[0148] For the current train operation status, the deep Q network selects the action with the largest target Q value from the action space. If the action corresponding to the maximum Q value is to adjust the operation path of a certain train to avoid congested sections, a corresponding scheduling plan is generated based on the action, including train path adjustment and resource allocation instructions. In terms of resource allocation, the power quota of each section in the power supply network is dynamically allocated according to the train priority in the scheduling plan. For high-priority trains, the power supply of the line section where they are located is given priority. At the same time, resource allocation instructions such as standby train deployment and signal setting are generated according to the resource status to ensure the rational allocation of train operation resources.

[0149] The feedback control module is used to perform Monte Carlo simulation on each scheduling scheme in parallel through the GPU, verify the conflict probability, energy consumption deviation and delay margin of each scheduling scheme based on the threshold, output the scheduling instruction set after verification, execute the scheduling instruction set and monitor the execution effect of the scheduling instruction set in real time to determine whether to resample heterogeneous data and update resource status.

[0150] The resource allocation system uses a GPU parallel computing architecture to perform Monte Carlo simulation on each scheduling scheme generated by the elastic allocation module. It determines the initial thresholds for conflict probability, energy consumption deviation, and delay margin based on historical operating data, and uses a long-short-term memory network to perform time series analysis on the relevant indicators of each scheduling scheme. At the same time, it combines real-time meteorological and environmental data, passenger flow density coefficient, and resource status to dynamically adjust the thresholds through an attention mechanism.

[0151] During the simulation process, the fuzzy comprehensive evaluation method is used to fuzzy match the conflict probability, energy consumption deviation and delay margin of each scheduling scheme obtained by Monte Carlo simulation with the threshold, and a membership function is set for each indicator. The evaluation result of the scheduling scheme is determined by integrating the membership function through the fuzzy synthesis operator, and compared with the evaluation threshold to determine whether the scheduling scheme has passed the verification, and the scheduling instruction set after verification is output.

[0152] After executing the dispatch instruction set, the resource loss system monitors the feedback data during the execution process in real time, including the operating data of the power supply network, standby trains and signals, as well as the execution status data of the dispatch instruction set. It extracts the indicator features of the feedback data through the convolutional neural network, and then uses the long short-term memory network to learn these features, predict the changing trend of the resource status, and update the resource status according to the prediction results. At the same time, it updates the node attributes of the topology map, the parameters of the Bayesian network and the state transition probability.

[0153] The implementation principle of the above train dynamic resource allocation system can be found in the specific embodiments of the train dynamic scheduling method, which will not be described in detail here.

Claims

1. A train dynamic dispatching method, characterized in that: include: Real-time acquisition of heterogeneous data, including track circuit signals, platform passenger flow images, and meteorological environment data. This data is then spatiotemporally aligned and processed at the onboard edge node using a spatiotemporal attention mechanism. The system also monitors the power supply network, backup trains, and signal machines to determine resource status, outputting a train data matrix that includes signal feature vectors, passenger flow density coefficients, meteorological correction parameters, and resource status. The output strategy of the train data matrix includes: Real-time acquisition of heterogeneous data, including track circuit signals, platform passenger flow images, and meteorological environment data. This data is then spatially and temporally aligned at the onboard edge node using a spatiotemporal attention mechanism, and then standardized. The track circuit signal is extracted through a residual convolutional neural network, and the extracted features are mapped into a signal feature vector, which includes signal strength, signal frequency, and signal phase. Perform semantic segmentation on platform passenger flow images and calculate the ratio of passenger pixel area to total platform area to obtain the initial passenger flow density. Combined with historical passenger flow data, the long short-term memory network is used to predict passenger flow trends, dynamically weighting the initial passenger flow density to generate a passenger flow density coefficient. The random forest algorithm is used to analyze the impact of different meteorological environmental data on train operation data, braking distance, and energy consumption. Based on the analysis results, an impact weight is assigned to each meteorological environmental data. The standardized meteorological environmental data and the impact weight are weighted and summed to generate meteorological correction parameters. Monitor the power supply network, standby trains, and signals to determine resource status, integrate signal feature vectors, passenger density coefficients, meteorological correction parameters, and resource status, and output a train data matrix; An evaluation vector for train scheduling is constructed based on the train data matrix and the scheduling knowledge graph. The evaluation vector includes an urgency factor, an economic value factor, and a line impact factor. The dimensional weights of the evaluation vector are dynamically adjusted based on the fuzzy analytic hierarchy process to generate train priorities. The sub-strategies for constructing the evaluation vector include: Normalize the train data matrix and associate it with the scheduling knowledge graph; The train's operating status is analyzed by combining the signal feature vectors and resource status in the train data matrix. Historical scheduling data from the scheduling knowledge graph is extracted to assess the delay risk value under the train's operating status. The passenger impact value of train delays is assessed based on the passenger flow density coefficient and platform carrying capacity. The delay risk value and passenger impact value are weighted and summed to obtain the train's urgency factor. Based on the energy consumption and meteorological correction parameters in the train data matrix, the energy consumption cost of the train is determined in combination with the dispatching knowledge graph. The operating cost of the train is obtained by combining the resource cost. The operating profit value is determined based on the passenger flow density coefficient and the dispatching knowledge graph. The economic value is determined by combining the operating profit value and the operating cost. The economic value factor of the train is obtained after normalizing the economic value. The passenger flow density coefficient and resource status in the train data matrix are combined to analyze the train line load. The line load value is evaluated by combining the historical load data in the scheduling knowledge graph. The line load value is weighted and combined with the line importance in the scheduling knowledge graph to obtain the line impact factor. Integrate the urgency factor, economic value factor and line impact factor to construct the evaluation vector for train scheduling; The generation strategy of the train priority includes: The fuzzy analytic hierarchy process is used to take the train priority as the target layer and the evaluation vector as the criterion layer. The importance between the criterion layers is characterized by fuzzy numbers and the fuzzy judgment matrix is determined. The consistency check and correction of the fuzzy judgment matrix are performed, and the fuzzy weight vector of the evaluation vector is obtained by solving the fuzzy characteristic equation. The fuzzy weight vector is fuzzified to obtain the dimension weight of the evaluation vector. The evaluation vector and the dimension weight are weighted and summed to obtain the comprehensive evaluation value of the train, and the train priority is generated according to the comprehensive evaluation value of the train; The scheduling plan is generated through reinforcement learning. The state space of reinforcement learning represents train priority and resource status, and the reward function of reinforcement learning represents time punctuality, energy consumption reduction rate and conflict risk value. Monte Carlo simulation is performed on each scheduling scheme in parallel using GPUs. The conflict probability, energy consumption deviation, and delay margin of each scheduling scheme are verified based on thresholds to output a verified scheduling instruction set. The scheduling instruction set includes train path adjustment and resource allocation instructions. The scheduling instruction set is executed and feedback data during the execution of the scheduling instruction set is monitored in real time to determine whether to resample heterogeneous data and update resource status.

2. The train dynamic scheduling method according to claim 1, characterized in that: The resource status determination sub-strategy includes: Deploy sensors to monitor the power supply network, backup trains, and signals in real time to obtain resource data, including the load rate of the power supply network, the equipment health and location of backup trains, and the failure probability of signals. A topology graph is established using the power supply network, standby trains, and signal machines as nodes. Node attributes represent resource data, and edges represent the coupling relationship between resource data. Node feature vectors of the topology graph are extracted using a graph convolutional network. A Bayesian network is embedded in the topology graph, the node feature vector is input into the Bayesian network, and the Bayesian network is trained with historical operation data to determine the state transition probability between nodes, and the state transition probability between nodes is integrated to determine the resource status.

3. The train dynamic dispatching method according to claim 2, wherein: The generation strategy of the scheduling scheme includes: Build a reinforcement learning model. The state space of reinforcement learning represents train priority and resource status. The action space of reinforcement learning represents train speed adjustment, train route selection, and resource allocation. The reward function of reinforcement learning represents on-time performance, energy consumption reduction rate, and conflict risk value. Train the deep Q network through state space and action space, calculate the target Q value, and update the parameters of the deep Q network based on the mean square error loss function; The action with the largest target Q value is selected through the deep Q network, and a scheduling plan is generated based on the action. The generated scheduling plan is evaluated by the time punctuality rate, energy consumption reduction rate and conflict risk value in the reward function to determine whether to adjust the reinforcement learning.

4. The train dynamic dispatching method according to claim 3, characterized in that: The verification sub-strategy of the scheduling instruction set includes: Build a GPU parallel computing architecture and decompose the Monte Carlo simulation task into three layers of subtasks; Based on historical operating data, the threshold is determined and the conflict probability, energy consumption deviation and delay margin of each scheduling scheme are analyzed in time series using a long short-term memory network. Combined with meteorological environment data, passenger flow density coefficient and resource status, the threshold is dynamically adjusted through the attention mechanism; The conflict probability, energy consumption deviation and delay margin of each scheduling scheme obtained by Monte Carlo simulation are fuzzy matched with the threshold through fuzzy comprehensive evaluation method; Membership functions are set for conflict probability, energy consumption deviation and delay margin. The evaluation result of the scheduling scheme is determined by integrating the membership functions through the fuzzy synthesis operator. The evaluation result is compared with the evaluation threshold to determine whether the scheduling scheme has passed the verification, and the scheduling instruction set after verification is output.

5. The train dynamic dispatching method according to claim 4, characterized in that: The resource status update strategy includes: Execute the dispatch instruction set and monitor the feedback data during the execution of the dispatch instruction set in real time. The feedback data includes the operation data of the power supply network, standby trains and signals, as well as the execution status data of the dispatch instruction set; Extract indicator features of feedback data through convolutional neural network; The indicator features of the feedback data are learned through the long short-term memory network to predict the changing trend of resource status and update the resource status. At the same time, the node attributes of the topological graph, the parameters of the Bayesian network and the state transition probability are updated.

6. The train dynamic dispatching method according to claim 5, characterized in that: The resampling strategy for heterogeneous data includes: Analyze the associations between track circuit signals, platform passenger flow images, and meteorological environment data through association rule mining algorithms, detect anomalies in these associations based on statistical methods, and analyze the causes of detected abnormal heterogeneous data. Based on the results of anomaly detection and cause analysis, determine the sampling frequency adjustment mechanism and select the sampling method based on the type of abnormal heterogeneous data; The quality of the resampled heterogeneous data is evaluated to determine whether to resample the heterogeneous data, and the resampled heterogeneous data is fused with the train data matrix to update the train data matrix.

Images