Railway heavy haul train stable operation optimization method based on deep learning

Abstract

The invention discloses a railway heavy haul train stable operation optimization method based on deep learning. The method comprises the following steps that S1, real-time original data in train operation are collected in real time; s2, performing standardization processing and de-noising processing on the real-time original data according to a second level, a minute level and an hour level; s3, performing real-time continuous prediction by adopting a multi-scale Fourier neural operator model trained by historical data to obtain train longitudinal dynamic state prediction output; s4, constructing a local dynamic game decision model, and determining an optimal combination scheme of traction force and braking force in real time through reinforcement learning; s5, constructing a global dynamic game decision model, and calculating an optimal speed curve and an energy management strategy in real time through a multi-agent coevolution algorithm; and S6, the longitudinal control scheme and the overall operation scheme are fused in real time, and a final control instruction is determined. Longitudinal impact and energy consumption are effectively reduced, and the control precision and stability are improved.

Description

Technical Field

[0001] The present invention relates to the field of railway transport control technology, and in particular to a method for optimizing the smooth operation of heavy-load railway trains based on deep learning. Background Art

[0002] Heavy-haul rail transport is a vital component of the railway freight system. With the rapid growth of modern logistics demands, heavy-haul rail transport is experiencing a trend of increasing haulage loads, longer train lengths, and higher operating speeds. The smooth operation of heavy-haul trains is directly related to transport safety, efficiency, and energy consumption, and has become a core concern in the railway transportation sector.

[0003] Currently, methods for optimizing the operation of heavy-haul trains on railways generally rely on traditional manual experience-based models or classic automatic control technologies. These include fixed traction and braking control modes based on train operation plan curves, simple PID feedback control methods, and basic quantitative analysis methods. These methods primarily rely on human experience or limited mathematical models to control train speed, traction, and braking force. The technical principle is to provide feedback adjustments based on the difference between the real-time state and the target speed curve to achieve smooth train operation and energy consumption control.

[0004] However, with the advancement of railway transportation technology, heavy-haul train operating conditions have become increasingly complex and variable. The transportation process involves not only changes in static factors such as vehicle load, train structure, track gradient, and curve radius, but also complex dynamic factors such as track geometry errors, dynamic changes in real-time load distribution, and longitudinal impact fluctuations during traction and braking. These factors make it difficult for traditional operating modes based on fixed or single-scale feedback control to accurately and in real time adapt to these complex and changing operating conditions.

[0005] In recent years, with the rapid development of artificial intelligence technology, especially the initial application of deep learning methods in the field of rail transit, some studies have used reinforcement learning technology to perform single-scale state prediction and operation decision optimization, and have achieved initial results. However, the control accuracy and real-time performance of such methods are often difficult to meet the stringent requirements of the actual operation of railway transportation; some studies have also attempted to use traditional deep learning networks for single-state prediction. Although there have been some improvements, there are still limitations in the prediction and fine control of complex coupled states.

[0006] Although the existing railway heavy-load train operation optimization technology has made great progress, in actual application, there are still common problems such as significant longitudinal impact, insufficient operation accuracy and high overall energy consumption. In addition, the existing technology still has obvious shortcomings in real-time prediction accuracy, multi-scale state coupling analysis and intelligent control decision-making.

[0007] Therefore, how to provide a deep learning-based optimization method for smooth operation of heavy-load railway trains is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0008] One purpose of the present invention is to propose a method for optimizing the smooth operation of heavy-load railway trains based on deep learning, which has the effect of significantly reducing the longitudinal impact and energy consumption during train operation and effectively improving the control accuracy and stability.

[0009] A method for optimizing smooth operation of a heavy-load railway train based on deep learning according to an embodiment of the present invention includes the following steps:

[0010] S1. Real-time collection of high-frequency data on longitudinal velocity, longitudinal acceleration, traction, braking force, train load, load distribution changes, track slope, curve radius, and track geometry error during the operation of heavy-load railway trains as real-time raw data;

[0011] S2. Based on the real-time raw data, normalization and denoising are performed at the second, minute, and hour scales to obtain a multi-scale spatiotemporal coupling input feature dataset;

[0012] S3. Using a multi-scale Fourier neural operator model trained with historical data, perform real-time continuous prediction calculations on a multi-scale spatiotemporal coupled input feature dataset to obtain a continuous prediction output describing the longitudinal dynamics of a heavy-load railway train.

[0013] S4. Build a local dynamic game decision-making model, using continuous prediction output as real-time input. Use a reinforcement learning algorithm to determine the optimal combination of traction and braking forces in real time, thereby obtaining a real-time optimized train longitudinal control plan.

[0014] S5. Build a global dynamic game decision-making model, using continuous prediction output as real-time input. Use a multi-agent collaborative evolutionary algorithm to calculate the optimal speed curve and energy management strategy in real time, and obtain a real-time optimized overall train operation plan.

[0015] S6. Integrate the train longitudinal control plan and the overall train operation plan in real time to determine the final control instructions.

[0016] Optionally, the S2 specifically includes:

[0017] S21, dividing the real-time raw data into continuous data segments with a second-level scale as the unit, and calculating the statistical distribution characteristic parameters of each data in each data segment to obtain a second-level statistical feature set;

[0018] S22. Based on the second-level statistical feature set, the abnormal dynamic components are stripped off to obtain effective signal data at the second level;

[0019] S23, dividing the real-time raw data into continuous data segments with minute-level scale as the unit, and extracting frequency domain energy distribution features for each data segment to obtain a minute-level frequency domain energy feature set;

[0020] S24. Based on the minute-level frequency domain energy feature set, the stable low-frequency pattern implicit in the data is extracted to obtain effective signal data at the minute-level scale;

[0021] S25, dividing the real-time raw data into continuous data segments with hourly scale as the unit, and establishing an autocorrelation matrix for each data segment to obtain a set of hourly data autocorrelation matrices;

[0022] S26. Based on the hourly data autocorrelation matrix set, the matrix eigenvalue decomposition method is used to extract the data time series stability pattern to obtain the effective signal data at the hourly scale;

[0023] S27. Using a unified spatiotemporal synchronization mechanism, the effective signal data at the second, minute, and hour scales are aligned and coupled one by one to construct a multi-scale spatiotemporal coupling input feature dataset.

[0024] Optionally, the S3 specifically includes:

[0025] S31, dividing the multi-scale spatiotemporal coupling input feature data set into a plurality of data sub-segments with fixed time lengths, calculating the nonlinear dynamics entropy parameter of each data sub-segment, and obtaining an input feature entropy parameter sequence;

[0026] S32, extracting the dynamic coupling strength characteristics between the data at each scale based on the input characteristic entropy value parameter sequence to obtain multi-scale coupling characteristic parameters;

[0027] S33, using the multi-scale coupling characteristic parameters as initial constraints, constructing a real-time nonlinear mapping optimization objective function, and performing numerical iterative optimization on the real-time nonlinear mapping optimization objective function to obtain an optimal mapping weight parameter matrix;

[0028] S34. Using the optimal mapping weight parameter matrix, establish the nonlinear spectrum mapping relationship of the multi-scale Fourier neural operator model in real time, and calculate the continuous prediction solution of the longitudinal dynamic state of the train in the Fourier spectrum space;

[0029] S35. Convert the continuous prediction solution from the Fourier spectrum space to the actual time domain through a joint time-frequency transformation to obtain a sequence of predicted instantaneous values ​​of the longitudinal dynamic state of the train in each data subsegment;

[0030] S36. Based on the predicted instantaneous value sequence, perform multi-dimensional continuity constraint calculation of the predicted values ​​of adjacent data sub-segments, and correct the predicted state boundary deviation in real time to obtain a continuous prediction output of the train longitudinal dynamic state.

[0031] Optionally, the S33 specifically includes:

[0032] S331. Based on the multi-scale coupling characteristic parameters, define the real-time nonlinear mapping optimization objective function as a weighted square sum function of the multi-scale spectral characteristic prediction errors;

[0033] S332, defining the solution process of the real-time nonlinear mapping optimization objective function as a constrained nonlinear optimization problem, and defining the constraint conditions as boundary constraints of the multi-scale coupling characteristic parameters;

[0034] S333, finding a gradient expression for the real-time nonlinear mapping optimization objective function;

[0035] S334. Based on the real-time gradient calculation formula, the weight parameter matrix is ​​updated in real time using the gradient descent method with momentum term;

[0036] S335, calculating the real-time nonlinear mapping optimization objective function value in real time after each update of the weight parameter matrix, and judging whether the convergence condition of the optimization objective function value satisfies the convergence judgment threshold condition;

[0037] S336. When the real-time nonlinear mapping optimization objective function value reaches a convergence condition, the weight parameter matrix after the current iteration update is output as the optimal mapping weight parameter matrix.

[0038] Optionally, the S4 specifically includes:

[0039] S41. Divide the continuous prediction output of the train longitudinal dynamic state into multiple continuous short-term data units, and sort each data unit by the sensitivity of the real-time state of traction and braking force to obtain a sensitivity priority data sequence;

[0040] S42. Based on the sensitivity priority data sequence, a dynamic association network of the longitudinal state of the train is constructed, and the impact path of the traction force and the braking force on the speed and acceleration is explicitly represented as a directed network topology structure;

[0041] S43. Determine a set of state transition paths for the traction and braking force action combinations based on the dynamic association network, and calculate the topological connection strength of each state transition path in real time;

[0042] S44, defining the topological connection strength threshold of the state transition path in real time, and screening the action combination path that meets the longitudinal control state stability requirement of the train according to the threshold condition;

[0043] S45. Adopting an adaptive action selection method, the screened action combination paths are screened and sorted in real time, and quantitatively sorted according to the state continuity and stability after the action is implemented;

[0044] S46. Based on the quantitative ranking results of the action combination paths, determine in real time the optimal action combination sequence of traction and braking force;

[0045] S47. Output the optimal action combination sequence in real time as a train longitudinal control optimization plan.

[0046] Optionally, the S42 specifically includes:

[0047] S421. Based on the sensitivity priority data sequence, identify the dominant change trends of the train longitudinal speed and longitudinal acceleration within each short-term data unit, and determine the sensitive state characteristics corresponding to the speed and acceleration change trends;

[0048] S422. Establishing, in real time, a temporal causal constraint relationship between the influence of the traction state and the braking force state on the speed and acceleration states based on the sensitive state characteristics, wherein the temporal causal constraint relationship is a set of causal ordered pairs;

[0049] S423. Based on the causal ordered pair set, construct a multi-level causal association tree structure to clarify the step-by-step influence relationship of the traction and braking force states on the longitudinal speed and longitudinal acceleration states;

[0050] S424. For each branch path of the multi-level causal association tree structure, calculate in real time a time delay parameter for the transmission of the traction force and braking force states to the longitudinal speed and longitudinal acceleration states, where the time delay parameter is a state impact delay duration;

[0051] S425. Establish constraints on the impact of the state on the delay duration in real time based on the state impact on the delay duration, and select causal association tree structure paths that meet the delay duration constraints;

[0052] S426. Based on the selected causal association tree structure paths, construct a dynamic association network topology structure of the train longitudinal state in real time, and determine the impact direction and delay parameters of each path in the topology structure;

[0053] S427. Mark the delay duration of the path state impact as a weight parameter of the directed path in the topological structure to obtain a complete dynamic association network of the longitudinal state of the train.

[0054] Optionally, the S5 specifically includes:

[0055] S51. Based on the continuous prediction output of the train longitudinal dynamic state, identify the key state turning point in the overall train operation process, where the key state turning point is the moment when the speed and energy consumption change trends change significantly at the same time;

[0056] S52: Constructing an interactive influence model between the train speed state and the energy consumption state for the key state turning point, and establishing a time series sensitivity matrix of the speed state to the energy consumption state in real time;

[0057] S53. Based on the timing sensitivity matrix, define a dynamic game field between the speed control agent and the energy management agent, wherein the dynamic game field includes the timing sensitivity influence relationship of the interaction between the agents;

[0058] S54, defining the dynamic equilibrium condition of the dynamic game action field in real time, calculating the real-time difference between the real-time change trend of the interaction between the speed state and the energy consumption state and the preset expected value of the operating state, and obtaining a real-time state difference sequence;

[0059] S55. For the real-time state difference sequence, a dynamic game update mechanism of the multi-agent co-evolutionary algorithm is established in real time, with the real-time state difference as the driving parameter for real-time update;

[0060] S56, executing the dynamic game update mechanism in real time, iteratively updating the speed curve control strategy and energy management strategy of the intelligent agent until the real-time state difference sequence meets the dynamic equilibrium condition;

[0061] S57. When the real-time state difference sequence meets the dynamic equilibrium condition, the corresponding speed curve control strategy and energy management strategy are output in real time as the overall train operation plan.

[0062] Optionally, the S52 specifically includes:

[0063] S521. For the key state turning point, determine in real time the dominant influence direction between the speed state and the energy consumption state, wherein the dominant influence direction is a clear causal sequence in which a change in the speed state leads to a change in the energy consumption state, or vice versa.

[0064] S522. Based on the dominant influencing direction, calculate in real time the temporal nonlinear mutual information between the speed state and the energy consumption state, wherein the temporal nonlinear mutual information is the nonlinear correlation strength of the speed state to the energy consumption state;

[0065] S523: defining a nonlinear mutual information threshold condition in real time, marking a data interval exceeding the nonlinear mutual information threshold condition as a significant interaction influence interval, and determining a specific time series range of the significant interaction influence interval;

[0066] S524. For each significant interaction influence interval, extract a nonlinear dynamic pattern of the speed state and the energy consumption state in real time, wherein the nonlinear dynamic pattern is a nonlinear dynamic mapping relationship between the speed and energy consumption state change trends within the significant interaction influence interval;

[0067] S525. Based on the extracted nonlinear dynamic pattern, a nonlinear sensitivity measurement index of the speed state to the energy consumption state is established in real time, wherein the nonlinear sensitivity measurement index is the nonlinear response degree of the speed state change to the energy consumption state change;

[0068] S526. Based on the nonlinear sensitivity measurement indicators of multiple significant interaction influence intervals, a time series sensitivity matrix of the speed state to the energy consumption state is constructed in real time, and the matrix elements are defined as specific values ​​of the nonlinear sensitivity measurement indicators.

[0069] Optionally, the S53 specifically includes:

[0070] S531. Based on the time-series sensitivity matrix, identify in real time the sensitivity peak state point between the speed control agent and the energy management agent, where the sensitivity peak state point is the state moment when the sensitivity value in the time-series sensitivity matrix exceeds a preset threshold and lasts for more than a set time;

[0071] S532. For the sensitivity peak state point, a dynamic game action propagation network is constructed in real time between the speed control agent and the energy management agent, wherein the nodes of the dynamic game action propagation network are the speed state variables and the energy consumption state variables, and the network edges are the dynamic game action propagation paths between the agents;

[0072] S533. Calculate the propagation effect parameters in the dynamic game action propagation network in real time. Define the propagation effect parameters as the propagation intensity and propagation delay time of the change in the state variable of another agent caused by the change in the state variable of the agent through the game action propagation network.

[0073] S534. Based on the propagation effect parameters, a dynamic game action field potential energy function between the speed control agent and the energy management agent is constructed in real time;

[0074] S535. Calculate the spatial gradient distribution of the potential energy function of the dynamic game field in real time and determine the location of the local extreme value point of the field potential energy function, where the local extreme value point is the state where the interaction of the state variables of each agent in the game field produces the maximum or minimum effect;

[0075] S536. Based on the location of the local extreme point, define in real time the relative equilibrium condition of the agent state in the dynamic game field, where the relative equilibrium condition is that the changing trend of the state variables of each agent simultaneously tends to the local extreme point of the field potential energy function;

[0076] S537. Establish a collaborative approach mechanism for the state variables of the intelligent agents in real time. Based on the relative equilibrium conditions of the dynamic game field, drive the state variables of the speed control intelligent agent and the energy management intelligent agent to approach the local extreme point of the potential energy function of the dynamic game field in real time.

[0077] Optionally, the S6 specifically includes:

[0078] S61, a state vector based on the train longitudinal control plan and the state vector of the train overall operation plan, wherein the state vector includes real-time target state values ​​of longitudinal velocity, longitudinal acceleration, traction force, and braking force;

[0079] S62. Constructing a multi-scale adaptive fusion mapping model between the control state vector and the overall operation state vector in real time, wherein the input of the fusion mapping model is the state difference value of the two state vectors at the second, minute, and hour scales;

[0080] S63, calculating the scale-adaptive fusion weight factor in the fusion mapping model in real time, and defining the scale-adaptive fusion weight factor as an adaptive weighting coefficient of the state difference value changing with the time scale;

[0081] S64, based on the scale-adaptive fusion weight factor, defines the nonlinear coupling mapping function of cross-scale state adaptive fusion in real time;

[0082] S65, determining in real time a stable convergence interval of the output value of the nonlinear coupling mapping function, wherein the stable convergence interval is a time interval in which the amplitude of the change of the output value of the fused state mapping is continuously lower than a preset convergence threshold;

[0083] S66. Determine, in real time, state consistency conditions for the longitudinal velocity, longitudinal acceleration, traction, and braking force based on the state map output value within the stable convergence interval, wherein the state consistency condition requires that each state variable approaches the state map output value and maintains a stable coordinated state;

[0084] S67. Based on the state coordination consistency condition, output the fused final control instruction in real time, where the final control instruction includes the coordinated longitudinal speed instruction, longitudinal acceleration instruction, traction force instruction, and braking force instruction.

[0085] The beneficial effects of the present invention are:

[0086] (1) The present invention adopts a multi-scale Fourier neural operator continuous prediction model to perform real-time, continuous, and high-precision prediction of the longitudinal dynamic state of heavy-load railway trains, thereby achieving accurate prediction of microscopic instantaneous states, mesoscopic dynamic behaviors, and macroscopic overall trends, effectively improving prediction accuracy and real-time response capabilities, significantly reducing the probability of longitudinal impact phenomena, and improving the smoothness of train operation.

[0087] (2) The present invention constructs a local dynamic game decision model and a global dynamic game decision model, and adopts reinforcement learning and multi-agent collaborative evolution algorithm for optimization calculation respectively, which can realize the real-time optimal combination optimization of traction and braking force and the overall optimization of speed curve and energy management strategy, significantly improving the accuracy of longitudinal control of the train and the refinement level of energy consumption management, and showing better adaptability and robustness in the complex and dynamic railway heavy-load transportation environment.

[0088] (3) In terms of the fusion optimization of the final control instructions, the present invention effectively solves the technical problem that the traditional single-scale control method is difficult to accurately coordinate local control with the overall operation plan by constructing a cross-scale state adaptive fusion mapping model and a scale-adaptive fusion weight factor mechanism. It breaks through the technical bottleneck of insufficient control accuracy and difficulty in coordination in the existing technology, and realizes real-time coordination and consistency among the longitudinal speed, longitudinal acceleration, traction and braking force states. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0090] Figure 1 This is a schematic diagram of the overall process of the deep learning-based railway heavy-load train smooth operation optimization method proposed by the present invention;

[0091] Figure 2 This is a schematic diagram of the structure of the multi-scale neural operator continuous prediction model for the deep learning-based railway heavy-haul train smooth operation optimization method proposed in the present invention;

[0092] Figure 3 Schematic diagram of the double-layer dynamic game decision fusion mechanism of the deep learning-based railway heavy-load train smooth operation optimization method proposed in this invention. DETAILED DESCRIPTION

[0093] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0094] refer to Figure 1-Figure 3 A method for optimizing the smooth operation of heavy-load railway trains based on deep learning is provided, comprising the following steps:

[0095] S1. Real-time collection of high-frequency data on longitudinal velocity, longitudinal acceleration, traction, braking force, train load, load distribution changes, track slope, curve radius, and track geometry error during the operation of heavy-load railway trains as real-time raw data;

[0096] S2. Based on the real-time raw data, normalization and denoising are performed at the second, minute, and hour scales to obtain a multi-scale spatiotemporal coupling input feature dataset;

[0097] S3. Using a multi-scale Fourier neural operator model trained with historical data, perform real-time continuous prediction calculations on a multi-scale spatiotemporal coupled input feature dataset to obtain a continuous prediction output describing the longitudinal dynamics of a heavy-load railway train.

[0098] S4. Build a local dynamic game decision-making model, using continuous prediction output as real-time input. Use a reinforcement learning algorithm to determine the optimal combination of traction and braking forces in real time, thereby obtaining a real-time optimized train longitudinal control plan.

[0099] S5. Build a global dynamic game decision-making model, using continuous prediction output as real-time input. Use a multi-agent collaborative evolutionary algorithm to calculate the optimal speed curve and energy management strategy in real time, and obtain a real-time optimized overall train operation plan.

[0100] S6. Integrate the train longitudinal control plan and the overall train operation plan in real time to determine the final control instructions.

[0101] The present invention adopts a multi-scale Fourier neural operator continuous prediction model to accurately predict the longitudinal dynamic state of the train in real time, which can effectively improve the prediction accuracy and reduce the longitudinal impact; further combined with the local dynamic game decision model and the global dynamic game decision model, reinforcement learning and multi-agent collaborative evolution algorithm are used to optimize the combination of traction and braking force, speed curve and energy management strategy, respectively, to achieve significant improvement in longitudinal control accuracy and overall operating energy consumption; at the same time, by real-time fusion of local control plans and overall operation plans to determine the final control instructions, the operating stability and control accuracy of heavy-load railway trains are effectively improved, overcoming the problems of insufficient real-time response capability and limited control accuracy of traditional methods.

[0102] In this embodiment, S2 specifically includes:

[0103] S21, dividing the real-time raw data into continuous data segments with a second-level scale as the unit, and calculating the statistical distribution characteristic parameters of each data in each data segment to obtain a second-level statistical feature set;

[0104] S22. Based on the second-level statistical feature set, the abnormal dynamic components are stripped off to obtain effective signal data at the second level;

[0105] S23, dividing the real-time raw data into continuous data segments with minute-level scale as the unit, and extracting frequency domain energy distribution features for each data segment to obtain a minute-level frequency domain energy feature set;

[0106] S24. Based on the minute-level frequency domain energy feature set, the stable low-frequency pattern implicit in the data is extracted to obtain effective signal data at the minute-level scale;

[0107] S25, dividing the real-time raw data into continuous data segments with hourly scale as the unit, and establishing an autocorrelation matrix for each data segment to obtain a set of hourly data autocorrelation matrices;

[0108] S26. Based on the hourly data autocorrelation matrix set, the matrix eigenvalue decomposition method is used to extract the data time series stability pattern to obtain the effective signal data at the hourly scale;

[0109] S27. Using a unified spatiotemporal synchronization mechanism, the effective signal data at the second, minute, and hour scales are aligned and coupled one by one to construct a multi-scale spatiotemporal coupling input feature dataset.

[0110] The present invention divides the real-time raw data into second-level, minute-level and hour-level scales and extracts statistical distribution characteristics, frequency domain energy characteristics and autocorrelation matrix characteristics. It then adopts abnormal dynamic component stripping, spectrum sparse reconstruction and matrix eigenvalue decomposition methods to obtain effective signal data of different scales and perform unified spatiotemporal synchronous fusion, which significantly improves the signal validity and feature stability in the data preprocessing stage, effectively overcomes the problem that traditional single-scale data preprocessing methods are difficult to simultaneously take into account changes in microscopic, mesoscopic and macroscopic states, ensures the input data quality of subsequent prediction models, and thus improves the accuracy and robustness of the overall prediction.

[0111] In this embodiment, S3 specifically includes:

[0112] S31, dividing the multi-scale spatiotemporal coupling input feature data set into a plurality of data sub-segments with fixed time lengths, calculating the nonlinear dynamics entropy parameter of each data sub-segment, and obtaining an input feature entropy parameter sequence;

[0113] S32, extracting the dynamic coupling strength characteristics between the data at each scale based on the input characteristic entropy value parameter sequence to obtain multi-scale coupling characteristic parameters;

[0114] S33, using the multi-scale coupling characteristic parameters as initial constraints, constructing a real-time nonlinear mapping optimization objective function, and performing numerical iterative optimization on the real-time nonlinear mapping optimization objective function to obtain an optimal mapping weight parameter matrix;

[0115] S34. Using the optimal mapping weight parameter matrix, establish the nonlinear spectrum mapping relationship of the multi-scale Fourier neural operator model in real time, and calculate the continuous prediction solution of the longitudinal dynamic state of the train in the Fourier spectrum space;

[0116] S35. Convert the continuous prediction solution from the Fourier spectrum space to the actual time domain through a joint time-frequency transformation to obtain a sequence of predicted instantaneous values ​​of the longitudinal dynamic state of the train in each data subsegment;

[0117] S36. Based on the predicted instantaneous value sequence, perform multi-dimensional continuity constraint calculation of the predicted values ​​of adjacent data sub-segments, and correct the predicted state boundary deviation in real time to obtain a continuous prediction output of the train longitudinal dynamic state.

[0118] The present invention divides the multi-scale spatiotemporal coupling input data into sub-segments and calculates the nonlinear dynamic entropy parameters, further extracts the dynamic coupling strength characteristics and constructs a real-time nonlinear mapping optimization objective function, and determines the optimal mapping weight parameter matrix through numerical iterative optimization, thereby realizing the nonlinear spectrum mapping and continuous prediction solution of the multi-scale Fourier neural operator model, effectively improving the accuracy and real-time performance of the train longitudinal dynamic state prediction; at the same time, through the joint time-frequency transformation and multi-dimensional continuity constraint calculation, the stability and consistency of the predicted state are ensured, effectively overcoming the problems of difficulty in capturing nonlinear dynamic characteristics and insufficient prediction stability in existing single-scale prediction methods, and significantly enhancing the reliability of railway heavy-load train operation state prediction.

[0119] In this embodiment, the S33 specifically includes:

[0120] S331. Based on the multi-scale coupling characteristic parameters, define the real-time nonlinear mapping optimization objective function as a weighted square sum function of the multi-scale spectrum feature prediction error:

[0121]

[0122] Where J(W) represents the real-time nonlinear mapping optimization objective function, W represents the mapping weight parameter matrix to be optimized, N represents the total number of multi-scale spectral features, α i Represents the weight coefficient of the spectrum characteristics of each scale, F i represents the actual spectrum eigenvalue of the i-th scale, represents the predicted spectrum eigenvalue of the i-th scale calculated based on the weight parameter matrix W;

[0123] S332. Define the solution process of the real-time nonlinear mapping optimization objective function as a constrained nonlinear optimization problem, and define the constraint conditions as the boundary constraints of the multi-scale coupling characteristic parameters:

[0124] W min ≤W≤W max ;

[0125] Among them, W min Represents the lower limit constraint of the weight parameter matrix, W max represents the upper limit constraint of the weight parameter matrix;

[0126] S333. Calculate the gradient expression of the real-time nonlinear mapping optimization objective function:

[0127]

[0128] in, represents the gradient of the objective function of real-time nonlinear mapping optimization, Represents the partial derivative of the predicted spectrum eigenvalue with respect to the weight parameter matrix;

[0129] S334. Based on the real-time gradient calculation formula, the weight parameter matrix is ​​updated in real time using the gradient descent method with momentum term:

[0130]

[0131] Among them, W (k+1) Represents the weight parameter matrix after the k+1th iteration update, W (k) represents the weight parameter matrix of the kth iteration, η represents the learning rate parameter of the gradient descent, β represents the momentum term parameter, W (k-1) represents the weight parameter matrix of the k-1th iteration;

[0132] S335. After each update of the weight parameter matrix, the real-time nonlinear mapping optimization objective function value is calculated in real time, and it is determined whether the convergence condition of the optimization objective function value meets the convergence judgment threshold condition:

[0133] |J(W (k+1) )-J(W (k) )|≤ε;

[0134] Among them, ε represents the convergence judgment threshold;

[0135] S336. When the real-time nonlinear mapping optimization objective function value reaches a convergence condition, the weight parameter matrix after the current iteration update is output as the optimal mapping weight parameter matrix.

[0136] The present invention constructs a real-time nonlinear mapping optimization objective function, performs real-time numerical iterative optimization using the weighted square sum of multi-scale spectral feature prediction errors, obtains the optimal mapping weight parameter matrix, and establishes the nonlinear spectral mapping relationship of the multi-scale Fourier neural operator model in real time, which can effectively improve the nonlinear adaptability of the prediction model and the spectral domain prediction accuracy; at the same time, the gradient descent method with momentum term is used to update the optimization objective function in real time, making the optimization iteration faster and more stable, thereby solving the technical bottlenecks of traditional linear or single-scale prediction models that are difficult to accurately capture complex nonlinear spectral features and the insufficient real-time optimization convergence speed, and significantly improving the real-time performance and accuracy of the railway heavy-load train state prediction model.

[0137] In this embodiment, the S4 specifically includes:

[0138] S41. Divide the continuous prediction output of the train longitudinal dynamic state into multiple continuous short-term data units, and sort each data unit by the sensitivity of the real-time state of traction and braking force to obtain a sensitivity priority data sequence;

[0139] S42. Based on the sensitivity priority data sequence, a dynamic association network of the longitudinal state of the train is constructed, and the impact path of the traction force and the braking force on the speed and acceleration is explicitly represented as a directed network topology structure;

[0140] S43. Determine a set of state transition paths for the traction and braking force action combinations based on the dynamic association network, and calculate the topological connection strength of each state transition path in real time;

[0141] S44, defining the topological connection strength threshold of the state transition path in real time, and screening the action combination path that meets the longitudinal control state stability requirement of the train according to the threshold condition;

[0142] S45. Adopting an adaptive action selection method, the screened action combination paths are screened and sorted in real time, and quantitatively sorted according to the state continuity and stability after the action is implemented;

[0143] S46. Based on the quantitative ranking results of the action combination paths, determine in real time the optimal action combination sequence of traction and braking force;

[0144] S47. Output the optimal action combination sequence in real time as a train longitudinal control optimization plan.

[0145] The present invention constructs a dynamic correlation network of the longitudinal state of the train in real time, and calculates the topological connection strength of the action combination state transfer path based on the influence path of traction and braking force on longitudinal speed and acceleration, and uses an adaptive action selection method to perform real-time screening and sorting, thereby determining the optimal traction and braking force combination sequence in real time, significantly improving the accuracy and real-time performance of the train's longitudinal control decisions; the method of the present invention effectively solves the problem that traditional control decision-making schemes are difficult to evaluate and select the optimal action combination in real time, significantly reduces the risk of longitudinal impact in the operation of heavy-loaded trains, and improves the overall control stability and safety.

[0146] In this embodiment, the S42 specifically includes:

[0147] S421. Based on the sensitivity priority data sequence, identify the dominant change trends of the train longitudinal speed and longitudinal acceleration within each short-term data unit, and determine the sensitive state characteristics corresponding to the speed and acceleration change trends;

[0148] S422. Establishing, in real time, a temporal causal constraint relationship between the influence of the traction state and the braking force state on the speed and acceleration states based on the sensitive state characteristics, wherein the temporal causal constraint relationship is a set of causal ordered pairs;

[0149] S423. Based on the causal ordered pair set, construct a multi-level causal association tree structure to clarify the step-by-step influence relationship of the traction and braking force states on the longitudinal speed and longitudinal acceleration states;

[0150] S424. For each branch path of the multi-level causal association tree structure, calculate in real time a time delay parameter for the transmission of the traction force and braking force states to the longitudinal speed and longitudinal acceleration states, where the time delay parameter is a state impact delay duration;

[0151] S425. Establish constraints on the impact of the state on the delay duration in real time based on the state impact on the delay duration, and select causal association tree structure paths that meet the delay duration constraints;

[0152] S426. Based on the selected causal association tree structure paths, construct a dynamic association network topology structure of the train longitudinal state in real time, and determine the impact direction and delay parameters of each path in the topology structure;

[0153] S427. Mark the delay duration of the path state impact as a weight parameter of the directed path in the topological structure to obtain a complete dynamic association network of the longitudinal state of the train.

[0154] The present invention constructs a multi-level causal association tree structure, clarifies the step-by-step influence relationship of traction and braking force states on longitudinal speed and longitudinal acceleration states, calculates the state impact delay time in real time, and establishes causal association path screening conditions, which can accurately identify and characterize the dynamic causal relationship and transmission characteristics between states; and then forms a complete longitudinal state dynamic association network, which solves the problem that traditional methods are difficult to effectively capture and express the complex dynamic causal relationship between train operation states, improves the accuracy and real-time performance of longitudinal state prediction and control, and thus enhances the reliability and safety of railway heavy-load train operation optimization.

[0155] In this embodiment, the S5 specifically includes:

[0156] S51. Based on the continuous prediction output of the train longitudinal dynamic state, identify the key state turning point in the overall train operation process, where the key state turning point is the moment when the speed and energy consumption change trends change significantly at the same time;

[0157] S52: Constructing an interactive influence model between the train speed state and the energy consumption state for the key state turning point, and establishing a time series sensitivity matrix of the speed state to the energy consumption state in real time;

[0158] S53. Based on the timing sensitivity matrix, define a dynamic game field between the speed control agent and the energy management agent, wherein the dynamic game field includes the timing sensitivity influence relationship of the interaction between the agents;

[0159] S54, defining the dynamic equilibrium condition of the dynamic game action field in real time, calculating the real-time difference between the real-time change trend of the interaction between the speed state and the energy consumption state and the preset expected value of the operating state, and obtaining a real-time state difference sequence;

[0160] S55. For the real-time state difference sequence, a dynamic game update mechanism of the multi-agent co-evolutionary algorithm is established in real time, with the real-time state difference as the driving parameter for real-time update;

[0161] S56, executing the dynamic game update mechanism in real time, iteratively updating the speed curve control strategy and energy management strategy of the intelligent agent until the real-time state difference sequence meets the dynamic equilibrium condition;

[0162] S57. When the real-time state difference sequence meets the dynamic equilibrium condition, the corresponding speed curve control strategy and energy management strategy are output in real time as the overall train operation plan.

[0163] The present invention identifies key state turning points and constructs the time-series sensitivity matrix of speed state and energy consumption state in real time, defines the dynamic game field between intelligent agents and calculates the real-time state difference, and drives state optimization in real time with the dynamic game update mechanism of multi-agent collaborative evolution algorithm, thereby realizing real-time coordination and precise optimization of the longitudinal speed curve and energy management strategy. It breaks through the technical limitations of traditional methods that are difficult to effectively capture and coordinate the complex dynamic relationship between speed and energy consumption state in real time, significantly improves the integrity, accuracy and adaptability of the control plan during the operation of heavy-load railway trains, and effectively improves the operating efficiency and energy-saving performance.

[0164] In this embodiment, the S52 specifically includes:

[0165] S521. For the key state turning point, determine in real time the dominant influence direction between the speed state and the energy consumption state, wherein the dominant influence direction is a clear causal sequence in which a change in the speed state leads to a change in the energy consumption state, or vice versa.

[0166] S522. Based on the dominant influencing direction, calculate in real time the temporal nonlinear mutual information between the speed state and the energy consumption state, wherein the temporal nonlinear mutual information is the nonlinear correlation strength of the speed state to the energy consumption state;

[0167] S523: defining a nonlinear mutual information threshold condition in real time, marking a data interval exceeding the nonlinear mutual information threshold condition as a significant interaction influence interval, and determining a specific time series range of the significant interaction influence interval;

[0168] S524. For each significant interaction influence interval, extract a nonlinear dynamic pattern of the speed state and the energy consumption state in real time, wherein the nonlinear dynamic pattern is a nonlinear dynamic mapping relationship between the speed and energy consumption state change trends within the significant interaction influence interval;

[0169] S525. Based on the extracted nonlinear dynamic pattern, a nonlinear sensitivity measurement index of the speed state to the energy consumption state is established in real time, wherein the nonlinear sensitivity measurement index is the nonlinear response degree of the speed state change to the energy consumption state change;

[0170] S526. Based on the nonlinear sensitivity measurement indicators of multiple significant interaction influence intervals, a time series sensitivity matrix of the speed state to the energy consumption state is constructed in real time, and the matrix elements are defined as specific values ​​of the nonlinear sensitivity measurement indicators.

[0171] The present invention determines the dominant influencing direction between the speed state and the energy consumption state in real time and calculates the time-series nonlinear mutual information, accurately identifies and marks the significant interactive influence intervals, and then extracts the nonlinear dynamic mode and constructs a nonlinear sensitivity measurement index to form a complete time-series sensitivity matrix of the speed state to the energy consumption state; thereby significantly improving the ability to accurately capture complex nonlinear dynamic relationships, effectively solving the problem of insufficient accuracy in the existing technology for analyzing the nonlinear interactive influence of speed and energy consumption states, and improving the real-time and accuracy of railway heavy-load train operation status prediction and control optimization.

[0172] In this embodiment, the S53 specifically includes:

[0173] S531. Based on the time-series sensitivity matrix, identify in real time the sensitivity peak state point between the speed control agent and the energy management agent, where the sensitivity peak state point is the state moment when the sensitivity value in the time-series sensitivity matrix exceeds a preset threshold and lasts for more than a set time;

[0174] S532. For the sensitivity peak state point, a dynamic game action propagation network is constructed in real time between the speed control agent and the energy management agent, wherein the nodes of the dynamic game action propagation network are the speed state variables and the energy consumption state variables, and the network edges are the dynamic game action propagation paths between the agents;

[0175] S533. Calculate the propagation effect parameters in the dynamic game action propagation network in real time. Define the propagation effect parameters as the propagation intensity and propagation delay time of the change in the state variable of another agent caused by the change in the state variable of the agent through the game action propagation network.

[0176] S534. Based on the propagation effect parameters, the potential energy function of the dynamic game action field between the speed control agent and the energy management agent is constructed in real time:

[0177]

[0178] Among them, P se (t) represents the potential energy function of the dynamic game field, α ij (t) represents the propagation intensity of the agent state variable i to the state variable j at the tth moment, τ ij (t) represents the propagation delay time of the state variable i acting on the state variable j at time t, and β is the delay time adjustment parameter;

[0179] S535. Calculate the spatial gradient distribution of the potential energy function of the dynamic game field in real time and determine the location of the local extreme value point of the field potential energy function, where the local extreme value point is the state where the interaction of the state variables of each agent in the game field produces the maximum or minimum effect;

[0180] S536. Based on the location of the local extreme point, define in real time the relative equilibrium condition of the agent state in the dynamic game field, where the relative equilibrium condition is that the changing trend of the state variables of each agent simultaneously tends to the local extreme point of the field potential energy function;

[0181] S537. Establish a collaborative approach mechanism for the state variables of the intelligent agents in real time. Based on the relative equilibrium conditions of the dynamic game field, drive the state variables of the speed control intelligent agent and the energy management intelligent agent to approach the local extreme point of the potential energy function of the dynamic game field in real time.

[0182] The present invention identifies the sensitivity peak state points between the speed control agent and the energy management agent in real time, constructs a dynamic game action propagation network and accurately calculates the propagation effect parameters, establishes the dynamic game action field potential energy function and determines its spatial gradient distribution and local extreme point positions, thereby defining the relative balance conditions and collaborative convergence mechanism of the agent states, and driving the state variables of each agent to coordinately approach the optimal state point in real time; it effectively solves the technical problem in the existing technology that real-time state collaborative optimization is difficult to accurately capture the interactive effects of dynamic games, improves the coordination, real-time and accuracy of the overall operation control of heavy-load railway trains, and significantly improves the smoothness and operation efficiency of train control.

[0183] In this embodiment, S6 specifically includes:

[0184] S61, a state vector based on the train longitudinal control plan and the state vector of the train overall operation plan, wherein the state vector includes real-time target state values ​​of longitudinal velocity, longitudinal acceleration, traction force, and braking force;

[0185] S62. Constructing a multi-scale adaptive fusion mapping model between the control state vector and the overall operation state vector in real time, wherein the input of the fusion mapping model is the state difference value of the two state vectors at the second, minute, and hour scales;

[0186] S63. Calculate the scale-adaptive fusion weight factor in the fusion mapping model in real time, and define the scale-adaptive fusion weight factor as an adaptive weighting coefficient of the state difference value changing with the time scale:

[0187]

[0188] Among them, w s (t) represents the scale-adaptive fusion weight factor at time t, ΔX s (t) represents the state difference value of scale s at time t, δ s represents the sensitivity adjustment parameter with a scale of s, and the exponential function exp(·) is used to amplify the weight gap of the state difference value;

[0189] S64, based on the scale-adaptive fusion weight factor, defines the nonlinear coupling mapping function of cross-scale state adaptive fusion in real time:

[0190] Z fusion (t)=∑ s∈{ms,s,min} w s (t)·tanh[λ s ΔX s (t)];

[0191] Among them, Z fusion (t) represents the state mapping output value after fusion at time t, tanh(·) is the hyperbolic tangent function, which is used to achieve nonlinear smooth mapping of cross-scale state difference values, and λ s represents the nonlinear smoothing parameter with scale s;

[0192] S65, determining in real time a stable convergence interval of the output value of the nonlinear coupling mapping function, wherein the stable convergence interval is a time interval in which the amplitude of the change of the output value of the fused state mapping is continuously lower than a preset convergence threshold;

[0193] S66. Determine, in real time, state consistency conditions for the longitudinal velocity, longitudinal acceleration, traction, and braking force based on the state map output value within the stable convergence interval, wherein the state consistency condition requires that each state variable approaches the state map output value and maintains a stable coordinated state;

[0194] S67. Based on the state coordination consistency condition, output the fused final control instruction in real time, where the final control instruction includes the coordinated longitudinal speed instruction, longitudinal acceleration instruction, traction force instruction, and braking force instruction.

[0195] The present invention constructs a multi-scale adaptive fusion mapping model between state vectors in real time, defines scale-adaptive fusion weight factors and establishes a nonlinear coupling mapping function for cross-scale state difference values, determines the stable convergence interval and then determines the coordination consistency conditions between the longitudinal speed, longitudinal acceleration, traction and braking force states in real time, and finally outputs the coordinated and optimized control instructions in real time; it effectively overcomes the problem that traditional single-scale or static fusion methods are difficult to achieve real-time high-precision state coordination fusion, significantly improves the longitudinal control stability, real-time performance and accuracy of heavy-load railway trains, and significantly reduces the longitudinal impact and energy consumption during operation.

[0196] Example 1:

[0197] To verify the feasibility and effectiveness of this invention, the deep learning-based method for optimizing the smooth handling of heavy-haul trains was applied to a typical heavy-haul railway line of a domestic freight company. This line transports large quantities of important materials such as coal and ore. The line features long train formations, heavy loads, and a complex and changing operating environment. Longitudinal impact and high energy consumption are long-standing issues that require urgent resolution.

[0198] In its implementation, the method first deployed a real-time, high-speed data acquisition system on the line to comprehensively capture dynamic state data related to train operation. This includes longitudinal velocity, longitudinal acceleration, traction, braking force, train load and weight distribution changes, as well as key parameters such as line slope, curve radius, and track geometry error. The collected data undergoes multi-scale spatiotemporal preprocessing to generate standardized input feature datasets at the second, minute, and hour scales. Anomalous noise and dynamic interference components are further removed to ensure data quality for the prediction model.

[0199] Subsequently, this embodiment conducted in-depth training and optimization of a multi-scale Fourier neural operator continuous prediction model based on a large amount of actual railway operation data from over three years, resulting in a high-precision continuous prediction model suitable for this line. During actual operation, the model receives preprocessed input data sets in real time and outputs continuous, high-precision predictions of longitudinal velocity and acceleration states, providing precise data support for further decision-making optimization.

[0200] Based on the above continuous prediction output, this embodiment constructs local and global dynamic game decision-making models. The local game decision-making model uses reinforcement learning to optimize the combination of traction and braking forces in real time, significantly reducing longitudinal impact during train operation. The global game decision-making model uses a multi-agent co-evolutionary algorithm to determine the optimal speed curve and energy management strategy for the entire train in real time, effectively reducing overall operating energy consumption.

[0201] During actual operation, this embodiment determines and outputs the final control instructions in real time by integrating the results of the above two decision models, thereby significantly improving the control accuracy and stability during train operation.

[0202] In order to objectively verify the technical effect of the method of the present invention, this embodiment conducted a comparative analysis of the train operation status of the line before and after implementation for a period of 6 months (January 2021 to June 2021), and sorted the relevant data into the following Table 1:

[0203] Table 1: Comparative data table of railway heavy-load train operation status before and after implementation of the present invention

[0204]

[0205]

[0206] The data in Table 1 clearly shows that before implementing the method of the present invention, the peak longitudinal impact force during train operation reached 950 kN, with an average of 28 significant longitudinal impacts per month, posing a significant safety hazard to train operation. Furthermore, the unit traction power consumption reached 17.2 kWh / 10,000 tons·km, resulting in high overall energy consumption. After implementing the method, the peak longitudinal impact force was significantly reduced to 270 kN, and the average number of impacts per month was reduced to 4, significantly improving train operation smoothness and significantly reducing safety risks. Simultaneously, the unit traction power consumption was significantly reduced to 11.9 kWh / 10,000 tons·km, saving approximately 30.8% of energy costs compared to before implementation. Furthermore, the speed tracking error was significantly reduced from ±4.0 km / h to ±0.8 km / h, and the control accuracy rate increased from 81.5% to 97.8%, demonstrating that the method of the present invention significantly improves the accuracy of train control. Furthermore, the dynamic response delay after implementation was reduced from 550 seconds to 65 seconds, significantly improving the efficiency of real-time control response.

[0207] For example, during implementation, a significant rapid change in line grade occurred. Traditional methods typically result in severe impacts and even emergency braking. However, through precise prediction and rapid decision-making and response, the proposed system successfully avoided any significant impact and smoothly navigated the complex section, demonstrating excellent real-time control performance and decision-making responsiveness. Furthermore, over the six months following implementation, the system achieved average monthly energy savings of 285,000 kWh, generating significant economic benefits for railway transportation companies.

[0208] The above detailed data show that the railway heavy-load train smooth operation optimization method proposed in the present invention can effectively solve the key technical problems of large longitudinal impact, high energy consumption and insufficient real-time control accuracy in railway heavy-load transportation. It has significant practical application value and can be widely promoted and applied to various complex railway transportation routes.

[0209] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for optimizing the smooth operation of heavy-load railway trains based on deep learning, characterized in that: The steps include: S1. Real-time collection of high-frequency data on longitudinal velocity, longitudinal acceleration, traction, braking force, train load, load distribution changes, track slope, curve radius, and track geometry error during the operation of heavy-load railway trains as real-time raw data; S2. Based on the real-time raw data, normalization and denoising are performed at the second, minute, and hour scales to obtain a multi-scale spatiotemporal coupling input feature dataset; S3. Using a multi-scale Fourier neural operator model trained with historical data, perform real-time continuous prediction calculations on a multi-scale spatiotemporal coupled input feature dataset to obtain a continuous prediction output describing the longitudinal dynamics of a heavy-load railway train. S4. Build a local dynamic game decision-making model, using continuous prediction output as real-time input. Use a reinforcement learning algorithm to determine the optimal combination of traction and braking forces in real time, thereby obtaining a real-time optimized train longitudinal control plan. S5. Build a global dynamic game decision-making model, using continuous prediction output as real-time input. Use a multi-agent collaborative evolutionary algorithm to calculate the optimal speed curve and energy management strategy in real time, and obtain a real-time optimized overall train operation plan. S6. Integrate the train longitudinal control plan and the overall train operation plan in real time to determine the final control instructions.

2. The method for optimizing the smooth operation of heavy-load railway trains based on deep learning according to claim 1, characterized in that: The S2 specifically includes: S21, dividing the real-time raw data into continuous data segments with a second-level scale as the unit, and calculating the statistical distribution characteristic parameters of each data in each data segment to obtain a second-level statistical feature set; S22. Based on the second-level statistical feature set, the abnormal dynamic components are stripped off to obtain effective signal data at the second level; S23, dividing the real-time raw data into continuous data segments with minute-level scale as the unit, and extracting frequency domain energy distribution features for each data segment to obtain a minute-level frequency domain energy feature set; S24. Based on the minute-level frequency domain energy feature set, the stable low-frequency pattern implicit in the data is extracted to obtain effective signal data at the minute-level scale; S25, dividing the real-time raw data into continuous data segments with hourly scale as the unit, and establishing an autocorrelation matrix for each data segment to obtain a set of hourly data autocorrelation matrices; S26. Based on the hourly data autocorrelation matrix set, the matrix eigenvalue decomposition method is used to extract the data time series stability pattern to obtain the effective signal data at the hourly scale; S27. Using a unified spatiotemporal synchronization mechanism, the effective signal data at the second, minute, and hour scales are aligned and coupled one by one to construct a multi-scale spatiotemporal coupling input feature dataset.

3. The method for optimizing the smooth operation of heavy-load railway trains based on deep learning according to claim 1, characterized in that: The S3 specifically includes: S31, dividing the multi-scale spatiotemporal coupling input feature data set into a plurality of data sub-segments with fixed time lengths, calculating the nonlinear dynamics entropy parameter of each data sub-segment, and obtaining an input feature entropy parameter sequence; S32, extracting the dynamic coupling strength characteristics between the data at each scale based on the input characteristic entropy value parameter sequence to obtain multi-scale coupling characteristic parameters; S33, using the multi-scale coupling characteristic parameters as initial constraints, constructing a real-time nonlinear mapping optimization objective function, and performing numerical iterative optimization on the real-time nonlinear mapping optimization objective function to obtain an optimal mapping weight parameter matrix; S34. Using the optimal mapping weight parameter matrix, establish the nonlinear spectrum mapping relationship of the multi-scale Fourier neural operator model in real time, and calculate the continuous prediction solution of the longitudinal dynamic state of the train in the Fourier spectrum space; S35. Convert the continuous prediction solution from the Fourier spectrum space to the actual time domain through a joint time-frequency transformation to obtain a sequence of predicted instantaneous values ​​of the longitudinal dynamic state of the train in each data subsegment; S36. Based on the predicted instantaneous value sequence, perform multi-dimensional continuity constraint calculation of the predicted values ​​of adjacent data sub-segments, and correct the predicted state boundary deviation in real time to obtain a continuous prediction output of the train longitudinal dynamic state.

4. The method for optimizing the smooth operation of heavy-load railway trains based on deep learning according to claim 3, characterized in that: The S33 specifically includes: S331. Based on the multi-scale coupling characteristic parameters, define the real-time nonlinear mapping optimization objective function as a weighted square sum function of the multi-scale spectral characteristic prediction errors; S332, defining the solution process of the real-time nonlinear mapping optimization objective function as a constrained nonlinear optimization problem, and defining the constraint conditions as boundary constraints of the multi-scale coupling characteristic parameters; S333, finding a gradient expression for the real-time nonlinear mapping optimization objective function; S334. Based on the real-time gradient calculation formula, the weight parameter matrix is ​​updated in real time using the gradient descent method with momentum term; S335, calculating the real-time nonlinear mapping optimization objective function value in real time after each update of the weight parameter matrix, and judging whether the convergence condition of the optimization objective function value satisfies the convergence judgment threshold condition; S336. When the real-time nonlinear mapping optimization objective function value reaches a convergence condition, the weight parameter matrix after the current iteration update is output as the optimal mapping weight parameter matrix.

5. The method for optimizing the smooth operation of heavy-load railway trains based on deep learning according to claim 1, characterized in that: The S4 specifically includes: S41. Divide the continuous prediction output of the train longitudinal dynamic state into multiple continuous short-term data units, and sort each data unit by the sensitivity of the real-time state of traction and braking force to obtain a sensitivity priority data sequence; S42. Based on the sensitivity priority data sequence, a dynamic association network of the longitudinal state of the train is constructed, and the impact path of the traction force and the braking force on the speed and acceleration is explicitly represented as a directed network topology structure; S43. Determine a set of state transition paths for the traction and braking force action combinations based on the dynamic association network, and calculate the topological connection strength of each state transition path in real time; S44, defining the topological connection strength threshold of the state transition path in real time, and screening the action combination path that meets the longitudinal control state stability requirement of the train according to the threshold condition; S45. Adopting an adaptive action selection method, the screened action combination paths are screened and sorted in real time, and quantitatively sorted according to the state continuity and stability after the action is implemented; S46. Based on the quantitative ranking results of the action combination paths, determine in real time the optimal action combination sequence of traction and braking force; S47. Output the optimal action combination sequence in real time as a train longitudinal control optimization plan.

6. The method for optimizing the smooth operation of heavy-load railway trains based on deep learning according to claim 5, characterized in that: The S42 specifically includes: S421. Based on the sensitivity priority data sequence, identify the dominant change trends of the train longitudinal speed and longitudinal acceleration within each short-term data unit, and determine the sensitive state characteristics corresponding to the speed and acceleration change trends; S422. Establishing, in real time, a temporal causal constraint relationship between the influence of the traction state and the braking force state on the speed and acceleration states based on the sensitive state characteristics, wherein the temporal causal constraint relationship is a set of causal ordered pairs; S423. Based on the causal ordered pair set, construct a multi-level causal association tree structure to clarify the step-by-step influence relationship of the traction and braking force states on the longitudinal speed and longitudinal acceleration states; S424. For each branch path of the multi-level causal association tree structure, calculate in real time a time delay parameter for the transmission of the traction force and braking force states to the longitudinal speed and longitudinal acceleration states, where the time delay parameter is a state impact delay duration; S425. Establish constraints on the impact of the state on the delay duration in real time based on the state impact on the delay duration, and select causal association tree structure paths that meet the delay duration constraints; S426. Based on the selected causal association tree structure paths, construct a dynamic association network topology structure of the train longitudinal state in real time, and determine the impact direction and delay parameters of each path in the topology structure; S427. Mark the delay duration of the path state impact as a weight parameter of the directed path in the topological structure to obtain a complete dynamic association network of the longitudinal state of the train.

7. The method for optimizing the smooth operation of heavy-load railway trains based on deep learning according to claim 1, characterized in that: The S5 specifically includes: S51. Based on the continuous prediction output of the train longitudinal dynamic state, identify the key state turning point in the overall train operation process, where the key state turning point is the moment when the speed and energy consumption change trends change significantly at the same time; S52: Constructing an interactive influence model between the train speed state and the energy consumption state for the key state turning point, and establishing a time series sensitivity matrix of the speed state to the energy consumption state in real time; S53. Based on the timing sensitivity matrix, define a dynamic game field between the speed control agent and the energy management agent, wherein the dynamic game field includes the timing sensitivity influence relationship of the interaction between the agents; S54, defining the dynamic equilibrium condition of the dynamic game action field in real time, calculating the real-time difference between the real-time change trend of the interaction between the speed state and the energy consumption state and the preset expected value of the operating state, and obtaining a real-time state difference sequence; S55. For the real-time state difference sequence, a dynamic game update mechanism of the multi-agent co-evolutionary algorithm is established in real time, with the real-time state difference as the driving parameter for real-time update; S56, executing the dynamic game update mechanism in real time, iteratively updating the speed curve control strategy and energy management strategy of the intelligent agent until the real-time state difference sequence meets the dynamic equilibrium condition; S57. When the real-time state difference sequence meets the dynamic equilibrium condition, the corresponding speed curve control strategy and energy management strategy are output in real time as the overall train operation plan.

8. The method for optimizing the smooth operation of heavy-load railway trains based on deep learning according to claim 7, characterized in that: The S52 specifically includes: S521. For the key state turning point, determine in real time the dominant influence direction between the speed state and the energy consumption state, wherein the dominant influence direction is a clear causal sequence in which a change in the speed state leads to a change in the energy consumption state, or vice versa. S522. Based on the dominant influencing direction, calculate in real time the temporal nonlinear mutual information between the speed state and the energy consumption state, wherein the temporal nonlinear mutual information is the nonlinear correlation strength of the speed state to the energy consumption state; S523: defining a nonlinear mutual information threshold condition in real time, marking a data interval exceeding the nonlinear mutual information threshold condition as a significant interaction influence interval, and determining a specific time series range of the significant interaction influence interval; S524. For each significant interaction influence interval, extract a nonlinear dynamic pattern of the speed state and the energy consumption state in real time, wherein the nonlinear dynamic pattern is a nonlinear dynamic mapping relationship between the speed and energy consumption state change trends within the significant interaction influence interval; S525. Based on the extracted nonlinear dynamic pattern, a nonlinear sensitivity measurement index of the speed state to the energy consumption state is established in real time, wherein the nonlinear sensitivity measurement index is the nonlinear response degree of the speed state change to the energy consumption state change; S526. Based on the nonlinear sensitivity measurement indicators of multiple significant interaction influence intervals, a time series sensitivity matrix of the speed state to the energy consumption state is constructed in real time, and the matrix elements are defined as specific values ​​of the nonlinear sensitivity measurement indicators.

9. The method for optimizing the smooth operation of heavy-load railway trains based on deep learning according to claim 7, characterized in that: The S53 specifically includes: S531. Based on the time-series sensitivity matrix, identify in real time the sensitivity peak state point between the speed control agent and the energy management agent, where the sensitivity peak state point is the state moment when the sensitivity value in the time-series sensitivity matrix exceeds a preset threshold and lasts for more than a set time; S532. For the sensitivity peak state point, a dynamic game action propagation network is constructed in real time between the speed control agent and the energy management agent, wherein the nodes of the dynamic game action propagation network are the speed state variables and the energy consumption state variables, and the network edges are the dynamic game action propagation paths between the agents; S533. Calculate the propagation effect parameters in the dynamic game action propagation network in real time. Define the propagation effect parameters as the propagation intensity and propagation delay time of the change in the state variable of another agent caused by the change in the state variable of the agent through the game action propagation network. S534. Based on the propagation effect parameters, a dynamic game action field potential energy function between the speed control agent and the energy management agent is constructed in real time; S535. Calculate the spatial gradient distribution of the potential energy function of the dynamic game field in real time and determine the location of the local extreme value point of the field potential energy function, where the local extreme value point is the state where the interaction of the state variables of each agent in the game field produces the maximum or minimum effect; S536. Based on the location of the local extreme point, define in real time the relative equilibrium condition of the agent state in the dynamic game field, where the relative equilibrium condition is that the changing trend of the state variables of each agent simultaneously tends to the local extreme point of the field potential energy function; S537. Establish a collaborative approach mechanism for the state variables of the intelligent agents in real time. Based on the relative equilibrium conditions of the dynamic game field, drive the state variables of the speed control intelligent agent and the energy management intelligent agent to approach the local extreme point of the potential energy function of the dynamic game field in real time.

10. The method for optimizing the smooth operation of heavy-load railway trains based on deep learning according to claim 1, characterized in that: The S6 specifically includes: S61, a state vector based on the train longitudinal control plan and the state vector of the train overall operation plan, wherein the state vector includes real-time target state values ​​of longitudinal velocity, longitudinal acceleration, traction force, and braking force; S62. Constructing a multi-scale adaptive fusion mapping model between the control state vector and the overall operation state vector in real time, wherein the input of the fusion mapping model is the state difference value of the two state vectors at the second, minute, and hour scales; S63, calculating the scale-adaptive fusion weight factor in the fusion mapping model in real time, and defining the scale-adaptive fusion weight factor as an adaptive weighting coefficient of the state difference value changing with the time scale; S64, based on the scale-adaptive fusion weight factor, defines the nonlinear coupling mapping function of cross-scale state adaptive fusion in real time; S65, determining in real time a stable convergence interval of the output value of the nonlinear coupling mapping function, wherein the stable convergence interval is a time interval in which the amplitude of the change of the output value of the fused state mapping is continuously lower than a preset convergence threshold; S66. Determine, in real time, state consistency conditions for the longitudinal velocity, longitudinal acceleration, traction, and braking force based on the state map output value within the stable convergence interval, wherein the state consistency condition requires that each state variable approaches the state map output value and maintains a stable coordinated state; S67. Based on the state coordination consistency condition, output the fused final control instruction in real time, where the final control instruction includes the coordinated longitudinal speed instruction, longitudinal acceleration instruction, traction force instruction, and braking force instruction.

Images