A train speed prediction method, device, equipment, medium and product

Abstract

The application discloses a train speed prediction method, device, equipment, medium and product, relates to the field of train speed prediction, and the method comprises the following steps: constructing a train speed prediction model; the train speed prediction model comprises a nonlinear autoregressive module, a block feature extraction module, a dynamic feature extractor and an FNN module; the nonlinear autoregressive module is constructed based on high-speed train single-particle dynamics; the dynamic feature extractor is constructed based on high-speed train multi-particle coupled dynamics; the average speed of the high-speed train at the current time and each historical time and the unit features of each dynamic unit are input into the train speed prediction model to obtain the predicted average speed of the train at the next time, and the application provides a train dynamics model which simultaneously meets the explainability, lightness and multi-particle global modeling, thereby improving the speed prediction efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of train speed prediction, and in particular, to a train speed prediction method, device, equipment, medium and product. BACKGROUND

[0002] Currently, researches related to High Speed Train (HST) speed modeling can be generally divided into mechanism-based and data-driven. For a mechanism-based HST speed prediction model, it is usually necessary to be based on train longitudinal dynamics, to deeply analyze motor traction characteristics, air-electric combined braking characteristics, and car coupling characteristics under the air-liquid buffer device, etc. However, with the long-distance operation demand and the speed improvement of modern high-speed trains, higher requirements are put forward for the adaptability and precision of the train model. The universality and effectiveness of the traditional mechanism model continue to decline due to the complex parameter identification and structure adjustment, especially in the future virtual marshalling scene of multiple train types, which will further increase the design burden of related engineers.

[0003] Considering that a large number of sensors are deployed on the vehicle end and the excellent nonlinear modeling capability of neural networks, a data-driven modeling scheme based on machine learning or deep learning has become another main research direction. Among them, the input layer and output layer of the deep neural network are verified by customization, and the deep neural network has higher prediction accuracy than the traditional regression model, but its black box characteristics hinder its practical application. In order to enhance the explainability of the model, an adaptive neuro-fuzzy inference system is introduced into the establishment of the train model, and is extended to multi-particle coupled dynamics modeling. However, establishing a separate adaptive neuro-fuzzy inference system for each car will lead to a doubling of the model parameter quantity. Considering that HST dynamics has significant time delay characteristics, a more suitable echo state network (ESN) for time series tasks is used to reduce the model parameter size. In order to make the model have both explainability and lightweight characteristics, the LAG-LSTM model is obtained by improving another variant LSTM. However, due to the limitations of the recurrent neural network and its variants in processing different time dependencies between input features, this type of network may not accurately distinguish the different effects of internal dynamics and track information on the train. From the above research, it can be found that although data-driven modeling has great advantages, it is difficult to simultaneously consider explainability, lightweight, and multi-particle global modeling, which hinders its practical application. SUMMARY

[0004] The purpose of the present application is to provide a train speed prediction method, device, equipment, medium and product, to provide a train dynamics model that simultaneously satisfies explainability, lightweight, and multi-particle global modeling, and to improve the prediction speed efficiency.

[0005] To achieve the above object, the application provides the following scheme: in a first aspect, the application provides a train speed prediction method, comprising: constructing a train speed prediction model; the train speed prediction model comprises an input layer, a block feature extraction module, a nonlinear autoregressive module, a dynamic feature extractor, and an FNN module; the output end of the input layer is connected with the first input end of the nonlinear autoregressive module, the input end of the block feature extraction module, the input end of the dynamic feature extractor, and the first input end of the FNN module respectively; the output end of the block feature extraction module is connected with the second input end of the nonlinear autoregressive module; the output end of the dynamic feature extractor is connected with the second input end of the FNN module; the output end of the nonlinear autoregressive module is connected with the output end of the FNN module; the nonlinear autoregressive module is constructed based on single-point dynamics of a high-speed train; and the dynamic feature extractor is constructed based on multi-point coupled dynamics of the high-speed train.

[0006] The actual average speed of the high-speed train at the current time and each historical time and the unit features of each power unit are acquired.

[0007] The actual average speed of the high-speed train at the current time and each historical time and the unit features of each power unit are input into the train speed prediction model to obtain the predicted average speed of the train at the next time.

[0008] In a second aspect, the application provides a train speed prediction device, comprising: a model construction module, configured to construct a train speed prediction model; the train speed prediction model comprises an input layer, a block feature extraction module, a nonlinear autoregressive module, a dynamic feature extractor, and an FNN module; the output end of the input layer is connected with the first input end of the nonlinear autoregressive module, the input end of the block feature extraction module, the input end of the dynamic feature extractor, and the first input end of the FNN module respectively; the output end of the block feature extraction module is connected with the second input end of the nonlinear autoregressive module; the output end of the dynamic feature extractor is connected with the second input end of the FNN module; the output end of the nonlinear autoregressive module is connected with the output end of the FNN module; the nonlinear autoregressive module is constructed based on single-point dynamics of a high-speed train; and the dynamic feature extractor is constructed based on multi-point coupled dynamics of the high-speed train.

[0009] The acquisition module is configured to acquire the actual average speed of the high-speed train at the current time and each historical time and the unit features of each power unit.

[0010] The prediction module is configured to input the actual average speed of the high-speed train at the current time and each historical time and the unit features of each power unit into the train speed prediction model to obtain the predicted average speed of the train at the next time.

[0011] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the train speed prediction method described above.

[0012] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the train speed prediction method described above.

[0013] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the train speed prediction method described above.

[0014] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a train speed prediction method, device, equipment, medium and product. The train speed prediction model includes a nonlinear autoregressive module, a dynamic feature extractor and a feedforward neural network (FNN) module. The nonlinear autoregressive module is constructed based on the single-mass dynamics of a high-speed train, so as to characterize the basic dynamics of a single mass with only a few interpretable parameters, avoid the redundant parameters of the "black box model" and thus realize the lightweight nature of the model. The dynamic feature extractor is constructed based on the multi-mass coupled dynamics of a high-speed train, so that the model attention weights can be directly mapped to physical factors, thereby realizing the interpretability of the model and global modeling of multiple masses. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating a train speed prediction method provided in one embodiment of this application.

[0017] Figure 2 This is a structural diagram of a high-speed train.

[0018] Figure 3 This is a structural diagram of the train speed prediction model provided in an embodiment of this application.

[0019] Figure 4 This is a structural diagram of the time attention module provided in an embodiment of this application.

[0020] Figure 5This is a structural diagram of the spatial attention module provided in an embodiment of this application.

[0021] Figure 6 This is a structural diagram of the FNN module provided in an embodiment of this application.

[0022] Figure 7 This is a graph showing the HST control speed curve for a certain day.

[0023] Figure 8 This is a graph showing the HST control command curve for a certain day.

[0024] Figure 9 The graph shows the single-step prediction results for different models.

[0025] Figure 10 The graph shows the single-step prediction error results for different models.

[0026] Figure 11 A graph comparing the performance of different models based on quantitative metrics.

[0027] Figure 12 The graph shows the multi-step prediction performance results for different models. Detailed Implementation

[0028] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0029] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0030] In one exemplary embodiment, such as Figure 1 As shown, a train speed prediction method is provided, including: Step 201: Constructing a train speed prediction model, abbreviated as NARXTSA; Figure 3As shown, the train speed prediction model includes an input layer, a block feature extraction module, a nonlinear autoregressive module, a dynamic feature extractor, and an FNN module. The output of the input layer is connected to the first input of the nonlinear autoregressive module, the input of the block feature extraction module, the input of the dynamic feature extractor, and the first input of the FNN module, respectively. The output of the block feature extraction module is connected to the second input of the nonlinear autoregressive module. The output of the dynamic feature extractor is connected to the second input of the FNN module. The output of the nonlinear autoregressive module is connected to the output of the FNN module. The nonlinear autoregressive module is constructed based on the single-mass dynamics of a high-speed train. The dynamic feature extractor is constructed based on the multi-mass coupled dynamics of a high-speed train.

[0031] Step 202: Obtain the actual average speed of the high-speed train at the current moment and at each historical moment, as well as the unit characteristics of each power unit.

[0032] Step 203: Input the actual average speed of the high-speed train at the current moment and at each historical moment, as well as the unit characteristics of each power unit, into the train speed prediction model to obtain the predicted average speed of the train at the next moment.

[0033] In practical applications, the output of the FNN module is connected to the output of the nonlinear autoregressive module through a fusion module.

[0034] In practical applications, high-speed trains typically consist of multiple carriages coupled together by couplers. Taking the CRH380B EMU as an example, it adopts a four-powered, four-trailer formation (4 power locomotives + 4 trailer cars), and the entire train can be divided into four power units, such as... Figure 2 As shown. During operation, each power unit simultaneously bears the force of the coupler of the adjacent unit, the force of the external environment, and its own traction or braking force. Based on Newtonian mechanics, the longitudinal dynamic equation of the multi-power unit is expressed as: (1).

[0035] (2).

[0036] in, Indicates to The derivative, Let be the acceleration coefficient of the i-th power unit. To work together, The actual traction or braking force of the i-th power unit itself. The basic resistance of the i-th power unit is... The additional resistance of the i-th power unit, Let the coupler force of the i-th power unit be... , , Let be the corresponding coefficient of the basic resistance of the i-th power unit. For the first The drag coefficient corresponding to each stiffness coefficient and Let be the displacement and velocity of the i-th power unit, respectively. Let be the slope of the i-th power unit. Let be the radius of curvature of the i-th dynamic element. The length of the tunnel. This is the stiffness coefficient. is the damping coefficient.

[0037] It is worth noting that train operation includes four typical conditions: traction, cruising, coasting, and braking. Traction, cruising, and coasting are achieved by the traction motor system, while the braking system is classified according to the actuator into electric braking, air braking, and electro-pneumatic combined braking. Under different conditions, the ideal braking command for the i-th power unit... With actual traction or braking force The mapping relationships are fundamentally different. For traction and electric braking conditions, the traction motor system achieves drive through cascaded control of the current loop and speed loop. At this time... and The relationship can be represented as: (3).

[0038] in For wheel-rail adhesion, Let g be the mass of the i-th power unit, and g be the acceleration due to gravity. The velocity-dependent viscosity coefficient, The upper limit of electric braking and , Represents the i-th power unit t- Ideal traction at all times Indicates air braking force. , Time-varying gain characterizes the dynamic response of power electronic devices, while and is the hysteresis constant, which characterizes the communication transmission delay and the delay of power switching devices. min() represents the minimum value and max() represents the maximum value.

[0039] Unlike electric braking, air braking involves a more complex pressure regulation process and an ideal braking force command output by the control algorithm. The target braking pressure needs to be solved in reverse, and then the brake cylinder pressure is regulated by valve control. The dynamic mechanism of this process can be described as follows: (4).

[0040] Where K represents the overall braking system coefficient, The average friction radius of the brake disc. For the effective friction radius, The diameter of the brake cylinder piston. For brake cylinder pressure, For braking ratio, For transmission efficiency, The number of brake discs in a single brake caliper. Indicates the rate of change of brake cylinder pressure. The adiabatic index, Atmospheric pressure, For relative density, For the brake cylinder volume, This refers to the pressure on the high-pressure side. For low-pressure side pressure, To control the valve cross-sectional area, For control signals, by Generated by the pressure distribution algorithm, when The brake cylinder is pressurized. Maintain pressure, Reduce stress during the day.

[0041] Given the strong nonlinearity and pressure hysteresis effect of equation (4), in air braking mode and The mapping relationship can be simplified as follows: (5).

[0042] in, Characterizing linear dynamic gain, Represents the i-th power unit Ideal traction at all times This refers to the hysteresis time of the air pressure system. Represents the nonlinear disturbance term (including the friction coefficient) (The rapid change characteristics and pressure dynamics are not modeled).

[0043] Based on the above and The mapping relationship can be discretized into equation (1) using the difference operator: (6).

[0044] in, At the current sampling time, This represents the nonlinear state transition function corresponding to the i-th dynamic unit. This represents the predicted velocity of the i-th power unit at time k+1. , , , Let represent the ideal traction or braking force command of the i-th power unit at time k, the displacement of the (i-1)-th power unit at time k, the track gradient at the i-th power unit, and the radius of curvature at the i-th power unit, respectively.

[0045] In equation (6), the displacement term can be further replaced by the cumulative relationship between velocity and control force. Furthermore, in actual scenarios, all power units are in approximately the same track conditions. In this case, considering external environmental and other influences as position disturbances, the following vector form can be obtained. (7).

[0046] in, This represents the vector composed of the predicted operating velocities of all power units at time k+1. Represents a nonlinear time-varying function. This represents the vector composed of the actual traction or braking forces of all power units at time k. This represents the track condition information at time k. Represents the actual traction or braking force of all power units. Represents the speed of all power units. Corresponding to 3 types of track condition information, For the generalized bounded disturbance at time k, and All are positive integers, representing the order of the system input and output.

[0047] Besides equation (1), if the coupling relationship between the couplers of each car is ignored, a simpler longitudinal dynamics model of a high-speed train can be expressed as follows: (8).

[0048] in, The average speed of the train. This represents the average speed of the train at time k. Indicates the acceleration coefficient. The basic resistance is represented by a, b, and c, which are the coefficients of the basic resistance polynomial. Indicates the track gradient. This represents the actual total traction or braking force of the train at time k. This refers to the train's actual total traction or braking force, taking into account additional resistance. Since the magnitude is relatively small, it can be further simplified to: (9).

[0049] in, , , The coefficients to be identified are... This indicates unmodeled dynamics and external disturbances.

[0050] As two basic forms of longitudinal dynamics representation for high-speed trains, Equation (1) accurately characterizes the multi-mass coupled dynamics of high-speed trains, while Equation (8) provides a concise single-mass approximation model. The latter introduces unmodeled dynamics due to neglecting the coupling and differences between units. The nonlinear autoregressive module constructed in this application is based on the single-mass dynamics of high-speed trains, and the dynamic feature extractor is based on the multi-mass coupled dynamics of high-speed trains. This data-driven hybrid architecture, which combines the structural simplicity of Equation (8) with the coupling representation capability of Equation (1), becomes a key path to overcome the limitations of existing modeling.

[0051] In practical applications, to better capture dynamics based on measurable data, equation (9) is reconstructed into a linear regression form: (10) The nonlinear autoregressive module is obtained.

[0052] in, This represents the predicted average velocity at time k+1 in the regression form. For input features, The parameter vector to be identified, This is the ideal setpoint for the total traction or braking force of the train at time k. Indicates the nth v +1 linear coefficient of historical velocity term, Indicates the nth v +1 nonlinear coefficient of the historical velocity squared term, Indicates the nth u +1 coefficient of historical traction / braking force command items, with the superscript T indicating transpose.

[0053] In practical applications, considering that while the actuator types of each power unit are consistent under the same operating conditions, they differ due to factors such as workshop coupling, external disturbances, and the load of the carriage, the dynamic feature extractor adopts a lightweight spatiotemporal joint structure: including a temporal attention module and a spatial attention module. The temporal attention module includes two parallel temporal attention variants, which process the input features of all power units in parallel, efficiently extracting the similar nonlinear temporal dependencies among multiple power units. The spatial attention module includes two parallel spatial attention variants, which update the unit node features based on physical connectivity, implicitly extracting the unit coupling difference features, such as... Figure 3 As shown, in another exemplary embodiment of this application, the dynamic feature extractor TSA includes a first temporal attention variant, a second temporal attention variant, a first fusion module, a first spatial attention variant, a second spatial attention variant, and a second fusion module.

[0054] The output of the input layer is connected to the input of the first temporal attention variant, and the output of the block feature extraction module is connected to the input of the second temporal attention variant. The outputs of both the first and second temporal attention variants are connected to the input of the first fusion module. The output of the first fusion module is connected to the inputs of both the first and second spatial attention variants. The outputs of both the first and second spatial attention variants are connected to the input of the second fusion module. The output of the second fusion module is connected to the second input of the FNN module.

[0055] In practical applications, considering the short dynamic temporal dependencies and clear modeling features of high-speed trains, this application designs two variants of dot product attention, such as... Figure 4 As shown, when extracting the dependencies between train operation dynamics and different features at historical moments, a first-time attention variant is used to process all available physical information to obtain global dependencies. Based on this, to focus on the impact of short-term traction and braking switching and abrupt changes on train operation, a block feature extraction module is used to extract the TA... input The difference between force and velocity is used as input to the second temporal attention variant. Furthermore, it utilizes block-based linear layers to perform local feature fusion in the time dimension to achieve initial focusing. Further processing with a second time attention variant will increase the finer granularity of the recognition of local feature changes.

[0056] Therefore, in another exemplary embodiment of this application, the first-time attention variant satisfies the formula: (11), Where H1 is the output of the first-time attention variant, representing the global temporal dependency features extracted by the first-time attention variant, and softmax() represents the softmax activation function. The input for the first-time attention variant is composed of the unit features of all dynamic units. This represents the first Q-weight matrix. This represents the feature dimension of the value space of the first-time attention variant. express The transpose of the matrix, Let V represent the first weight matrix. This serves as the input to the first-time attention variant. This represents the merging of the element characteristics of all dynamical units in the highest dimension, where... Let be the number of power units, and the th be... Unit characteristics of a power unit (train) Represented as: (12).

[0057] in, This represents the ideal traction force of the train at time t. The ideal electric braking force for the train at time t. The ideal air braking force of the train at time t is given, and the time series alignment satisfies the following conditions. (This operation is reasonable under the influence of the attention mechanism.) Let i represent the ideal speed force of the i-th train at time t. This represents the road conditions at time t.

[0058] In another exemplary embodiment of this application, the output of the block feature extraction module includes all block features of all dynamic units, and the second temporal attention variant satisfies the formula. (13) in, This represents the Nth local feature of a multi-dynamic unit, which is composed of the Nth block features of all dynamic units. N c This represents the total number of power units, and the time series alignment satisfies N. T =max(n v ,n u ), and All are positive integers, representing the order of the system input and output. express The transpose of is given, H2 is the output of the second temporal attention variant, representing local temporal features, and softmax() represents the softmax activation function. This represents the second Q-weight matrix. This represents the feature dimension of the value space of the second temporal attention variant. Let A represent the weight matrix. This represents the second V weight matrix. express The transpose of the matrix, This represents the local feature matrix, constructed to accommodate subsequent local attention calculations. This indicates that the data within [ ] will be concatenated along the first dimension. This indicates that the last two dimensions of the concatenated high-dimensional matrix are transposed. The output of the block linear layer represents Block features after linear processing , The number of segments in the block-based linear layer. The core objective is to adjust the dimensional order of the feature matrix to satisfy the linear operation requirement of "feature dimension - sample dimension". The output of the block feature extraction module consists of the block features of all dynamic units. Indicates and The weight matrix corresponding to the dimension Indicates and The bias term corresponding to the dimension. During initial focusing, the first... The first power unit Each block feature representation satisfies the formula: (14).

[0059] in, , , , , , , , They represent the first Differential information of ideal traction at any given moment, the first Differential information of ideal traction at time, the first Differential information of electric braking force at time, the first Differential information of electric braking force at time, the first Differential information of air braking force at time, the first Differential information of air braking force at time, the first The difference information of velocity at time and the first Differential information of velocity at different times, To represent the number of block-based linear layer segmentations, similarly, the local features of multi-dynamic units are represented as follows: .

[0060] In practical applications, the first fusion module satisfies the formula: (15). Among them, The output of the first fusion module includes the original characteristics of each power unit at the current moment and at each historical moment. Represents a high-dimensional matrix The last two dimensions are transposed. Represents the weight matrix. This indicates the bias term. This means concatenating H1 and H2 along the second dimension. For the feature dimension of the value space, , , , , , , , , , The superscript indicates the dimension, and the superscript indicates the specific dimension value.

[0061] Spatial attention module such as Figure 5 As shown, the output of the time attention module is regarded as node features, and an adjacency matrix is ​​defined based on the physical relationship between the power units of the high-speed train. Its core consists of an immediate adjacency fusion mechanism (the first spatial attention variant) and a delayed non-adjacency fusion mechanism (the second spatial attention variant): the immediate adjacency fusion mechanism updates its own node according to the information of directly coupled nodes to extract the coupling effect at the same time; the delayed non-adjacency fusion mechanism updates the current node according to the information of indirectly coupled nodes at historical time points to extract the indirect effect of delay.

[0062] In another exemplary embodiment of this application, the first spatial attention variant includes a first dynamic unit updating a feature unit and a first fusion unit; the first dynamic unit updating a feature unit satisfies the formula... (16), among which, This represents the concatenated feature vector of the i-th power unit and the j-th power unit at time k. This represents the original characteristics of the i-th power unit at time k. This represents the original characteristics of the j-th power unit at time k. express The transpose of the matrix, This indicates that the instantaneous adjacency fusion mechanism can learn weight vectors. This represents the leaky ReLU activation function, denoted as... ,in It is an exponential function. Used to generate attention scores , This represents the attention score between the i-th and j-th dynamic units at time k, quantifying the degree of feature correlation between them. The i-th and j-th dynamic units are adjacent. This represents the attention score constraint between the i-th dynamic unit and the j-th dynamic unit at time k. The attention score is calculated only when i and j are physically coupled; otherwise, the score is set to zero. After softmax, the weights are 0, ensuring that non-physically coupled units do not participate in feature fusion and guaranteeing physical interpretability. This represents the data located in the i-th row and j-th column of the adjacency matrix A. It is an adjacency matrix, constructed based on the physical structure of high-speed trains. This represents the direct physical coupling between the i-th power unit and the j-th power unit. This represents the degree of dependence between the i-th and j-th power units at time k, quantifying the degree of dependence between the i-th and j-th power units. The learnable factor is used to address the weight imbalance problem with a small number of adjacencies. exp() represents the exp function, and N c Indicates the total number of power units. The output of the feature unit is updated for the first power unit, representing the updated feature of the i-th power unit at time k. .

[0063] The first fusion unit is used to obtain the adjacency fusion feature matrix based on the updated features of each power unit at each time step. .

[0064] In another exemplary embodiment of this application, the second spatial attention variant includes a second dynamic unit updating a feature unit and a second fusion unit; the second dynamic unit updating a feature unit... (17), among which, This represents the original characteristics of the j-th non-adjacent dynamic unit at time kq. The matrix dimension is , express The characteristics of the j-th power unit at time k. Let represent the time series characteristics of the multi-dynamic unit at time k with a delay of q steps, where time k with a delay of q steps is time kq. This represents the original characteristics of the i-th power unit at time k. express transpose, This represents the concatenated vector of the current features of the i-th power unit and the historical features of the j-th power unit at time k, with a delay of q steps. This represents the learnable weight vector of the hysteresis non-adjacent fusion mechanism. This represents the leaky ReLU activation function. This represents the attention score between the i-th and j-th power units when there is a delay of q steps at time k. This represents the data located in the i-th row and j-th column of the adjacency matrix A, which is constructed based on the physical structure of high-speed trains. This represents the attention score constraint between the i-th and j-th dynamic units at time k with a delay of q steps. The attention score is calculated only if i and j are non-adjacent units or themselves; otherwise, the score is set to zero. After softmax, the weights are 0. exp() represents the exp function, N c Indicates the total number of power units. This represents the learnable balance factor, and its immediate adjacency. Similar in function, it is used to alleviate the imbalance in weight distribution when there are few non-adjacent units, such as when a unit has only one non-adjacent unit, to avoid excessive concentration of weight. This represents the delayed non-adjacent attention weight between the i-th and j-th power units at time k with delay q steps. Let i represent the set of non-adjacent units of the i-th power unit. This represents the upper limit of the interaction hysteresis between non-adjacent dynamic units. The output of the feature unit is updated for the second dynamic unit, representing the hysteresis non-adjacent fusion feature of the i-th dynamic unit at time k. This represents the number of lag steps.

[0065] The second fusion unit is used to obtain the hysteresis non-adjacent fusion feature matrix based on the hysteresis non-adjacent fusion characteristics of each dynamic unit at each time step. .

[0066] In practical applications, learnable fusion factors are utilized. The second fusion module balances the contributions of immediate adjacency and delayed non-adjacency, and performs mean pooling on the power unit node features to extract the features of the entire EMU train. The fusion module satisfies the following formula: (18).

[0067] in, , The characteristic matrix representing the entire train set at all times. Let be the characteristic matrix of the entire EMU train at time k.

[0068] In practical applications, the FNN module has the following structure: Figure 6 As shown, the input to the model output layer is... With another feature vector Composition, in which, Indicates the kNth T The difference in the average speed of the entire vehicle at time +1 describes the dynamic rate of change of speed. Indicates the kNth T The difference between the total traction or braking force command at time +1 reflects the abrupt change characteristics of the force command, and the output is an unmodeled speed compensation term. Specifically, it is expressed as: (19).

[0069] in, For vectorization operations, For activation function, and The first Each hidden node has its own input weight and bias. For the first Each hidden node outputs a weight. This represents the total number of nodes in the hidden layer.

[0070] The train speed prediction model of this application consists of four core modules: a nonlinear autoregressive module (NARX) is used to identify the known part of the single-mass mechanism described by equation (9); a dynamic feature extractor composed of a time attention module and a spatial attention module is responsible for extracting the nonlinear temporal dependence of each power unit and its overall coupling features; and an FNN module further processes these features and generates corresponding unmodeled compensation terms.

[0071] First-time attention variant uses only W1 V The value matrix V underwent a dimension upscaling transformation using W1. Q The query matrix Q was subjected to an equal-dimensional transformation and combined with the original feature K to calculate the attention score. K, as the original feature, allows the attention score to directly reflect the actual impact of the physical variables, enhancing interpretability while reducing parameter size. Based on this, a transformation matrix W was introduced between the attention score S and the value V. A A second temporal attention variant, W, was obtained. A The introduction of this kernel adds a second learnable temporal alignment kernel, allowing the model to dynamically adjust the vertical alignment of the value matrix and enhancing its ability to capture local dynamic changes.

[0072] The node update mechanism of the spatial attention module draws on the core idea of ​​graph attention networks, namely, dynamically aggregating the features of neighboring nodes through attention weights. The instantaneous adjacency fusion in formula (16) is achieved through attention weights. The current moment adjacent to the power unit was quantified. For the target node The real-time impact intensity. The hysteresis non-adjacent fusion of formula (17) is for interval dynamic units, by introducing historical moment features. Capture the delayed coupling effect caused by vehicle body flexibility or airflow disturbance.

[0073] This application provides a specific embodiment that uses actual HST operation data to verify the effectiveness of the train speed prediction model and the edge-cloud collaborative online correction framework. Simultaneously, to further analyze the contribution of the core algorithms in the collaborative correction framework, an ablation experiment was conducted using a set of industrial gas turbine benchmark datasets.

[0074] All experiments were conducted on a computer equipped with an Intel(R) Core(TM) i5-13600KF CPU (base frequency 3.5 GHz) and an NVIDIA GeForce RTX 4060 GPU (16GB of memory), and were verified and analyzed using Python with 20 independent replicate experiments.

[0075] In simulating a collaborative edge-cloud environment, independent storage and computing resource units are partitioned on the computer to simulate the device, edge nodes, and cloud platform. To realistically reflect the characteristics of a distributed correction system, the impact of communication and processing latency is given special attention. The experiment employs a multi-process architecture to simulate the parallel operation of multi-device algorithms, and a virtual latency mechanism is introduced in the inter-process communication stage to simulate the time overhead caused by actual network transmission.

[0076] This embodiment will verify the effectiveness of the proposed train speed prediction model NARXTSA under different task requirements from both single-step prediction and long-term rolling prediction, and compare its performance with models: LAG-LSTM, RSCN, LSTM, and ablation versions of NARX and TSA. Experimental data comes from the onboard intelligent data acquisition system of the CRH380B EMU on a designated section from October 25th to October 31st, 2021. Train operation data for a certain day is as follows: Figure 7 as well as Figure 8 As shown, Figure 8 Parts (a), (b), (c), and (d) represent the train operation command curves for the first, second, third, and fourth power units. The operational data from the first two stations will be used as offline training and validation samples, and a third station containing multiple operating condition transitions on a given day will be selected as the test dataset. The index shown in Equation (20) will be used as the evaluation standard for the train model's performance.

[0077] (20)

[0078] in, To assess sample size, The average speed of the train evaluation sample across the entire time domain. Let k be the actual average speed of the train at time k. Let be the predicted average velocity of the model at time k, MAE be the mean absolute error, RMSE be the root mean square error, and RRSE be the relative root mean square error.

[0079] The key parameters are set as follows: sampling time is 0.5s, and the step size of all model input sequences is... Multi-mass power unit TSA and NARXTSA share hyperparameters , , NARXTSA-specific parameters , The LSTM has 32 hidden neurons and 2 hidden layers; the LAG-LSTM has 16 hidden neurons and 1 hidden layer, and the RSCN parameters are set the same as in Algorithm 1; LSTM, LAG-LSTM, TSA and NARXTSA are learned offline using the gradient descent algorithm, with a batch size of 128.

[0080] The specific steps include: Step 3.1.1: Model single-step prediction performance.

[0081] First, we conduct a preliminary investigation into the single-step prediction performance of the models. The prediction results of all models on the test set are as follows: Figure 9 As shown, the corresponding prediction error is as follows: Figure 10 As shown, the quantization performance obtained according to equation (20) is presented in Figure 11 In order to comprehensively evaluate the overall performance of the model, Figure 11 Part (a) represents the quantitative performance obtained by taking the average of the evaluation index from multiple repeated experiments based on equation (20). Figure 11 Part (b) is the quantitative performance obtained by taking the standard deviation of the evaluation index according to the formula (20) in multiple repeated experiments. Based on the evaluation index of formula (20), three additional indicators were added: batch training time, total training time, and model parameter size. The corresponding results are shown in Table 1.

[0082] Compared to other models, combining Figure 9 and Figure 10 It can be seen that the train speed prediction model proposed in this application has a better fitting effect, which verifies the advanced nature of the proposed model in multi-condition identification of high-speed trains and its effectiveness in adapting to control tasks. According to Figure 11 The experimental results show that the TSA-based model outperforms other models on both the training and test sets, and its performance on the test set is consistently no worse than that on the training set. This demonstrates the positive role of TSA in promoting strong generalization of the model.

[0083] Based on the operational characteristics of each model in Table 1, the low model complexity of NARX, limited by its structure and recognition algorithm, is the fundamental reason for its poor recognition accuracy. Compared to LSTM-type models based on gating mechanisms, TSA-based models significantly reduce the parameter scale while improving modeling accuracy. Furthermore, comparing the overall training time of TSA and NARX shows that the addition of NARX accelerates the model's convergence speed. It is worth noting that although... Figure 11 Experimental results show that the RSCN network has slightly poorer generalization performance, but it is significantly faster than gradient-based training methods in terms of modeling speed, which highlights the natural advantage of randomized allocation algorithms in the field of real-time correction.

[0084] Table 1 Comparison of Operating Characteristics of Different Models

[0085]

[0086] Step 3.1.2 Model multi-step prediction performance.

[0087] Based on the offline model, this embodiment further explores the multi-step prediction performance of the model. This function will be used for iterative optimization of the position-velocity target curve and train state perception in the future. This requires the model to establish a complete interaction mechanism between traction or braking force and train state. Therefore, multi-step prediction can more accurately judge the model's generalization ability under different exogenous inputs and system states. Figure 12 The multi-step prediction results of each model are shown. Figure 12 Part (a) shows the multi-step prediction results of the NARX model, part (b) shows the multi-step prediction results of the TSA model, part (c) shows the multi-step prediction results of the NARXTSA model, part (d) shows the multi-step prediction results of the RSCN model, part (e) shows the multi-step prediction results of the LSTM model, and part (f) shows the multi-step prediction results of the LAG-LSTM model. More detailed quantification results are shown in Table 2.

[0088] For multi-step prediction tasks, the challenge lies in the variation of input time delay during operational condition transitions (such as a train switching from traction to coasting, or from coasting to braking). Figure 12 It can be seen that, compared with the TSA and NARXTSA models, other models showed more obvious deviations at some switching points. Combined with the quantization results in Table 2, the NARXTSA model has a more prominent advantage in multi-step prediction tasks. This verifies the effectiveness of the focused processing of differential features in the designed model and the temporal alignment kernel added in the second temporal attention variant.

[0089] Table 2 Comparison of Quantitative Indicator Performance of Different Models

[0090]

[0091] The train speed operation model constructed in this application combines mechanistic interpretability with data-driven flexibility. By integrating mechanistic priors and spatiotemporal attention mechanisms, it achieves accurate global dynamic modeling of multiple mass points while ensuring lightweight design.

[0092] Based on the same inventive concept, this application also provides a train speed prediction device for implementing the train speed prediction method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations of one or more train speed prediction device embodiments provided below can be found in the limitations of the train speed prediction method described above, and will not be repeated here.

[0093] In an exemplary embodiment, a train speed prediction device is provided, comprising: a model building module for building a train speed prediction model; the train speed prediction model includes an input layer, a block feature extraction module, a nonlinear autoregressive module, a dynamic feature extractor, and an FNN module; the output of the input layer is connected to the first input of the nonlinear autoregressive module, the input of the block feature extraction module, the input of the dynamic feature extractor, and the first input of the FNN module, respectively; the output of the block feature extraction module is connected to the second input of the nonlinear autoregressive module; the output of the dynamic feature extractor is connected to the second input of the FNN module; the output of the nonlinear autoregressive module is connected to the output of the FNN module; the nonlinear autoregressive module is built based on the single-mass dynamics of a high-speed train; and the dynamic feature extractor is built based on the multi-mass coupled dynamics of a high-speed train.

[0094] The acquisition module is used to acquire the actual average speed of the high-speed train at the current moment and at each historical moment, as well as the unit characteristics of each power unit.

[0095] The prediction module is used to input the actual average speed of the high-speed train at the current moment and at each historical moment, as well as the unit characteristics of each power unit, into the train speed prediction model to obtain the predicted average speed of the train at the next moment.

[0096] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal. The computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is connected to the system bus via the I / O interfaces. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the computer device stores train speed prediction data. The I / O interfaces of the computer device are used for exchanging information between the processor and external devices. The communication interface of the computer device is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a train speed prediction method.

[0097] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described method embodiments.

[0098] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the above-described method embodiments.

[0099] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the above-described method embodiments.

[0100] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0101] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0102] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0103] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0104] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A train speed prediction method, characterized in that, The train speed prediction method includes: A train speed prediction model is constructed. This model includes an input layer, a block feature extraction module, a nonlinear autoregressive module, a dynamic feature extractor, and an FNN module. The output of the input layer is connected to the first input of the nonlinear autoregressive module, the input of the block feature extraction module, the input of the dynamic feature extractor, and the first input of the FNN module. The output of the block feature extraction module is connected to the second input of the nonlinear autoregressive module. The output of the dynamic feature extractor is connected to the second input of the FNN module. The output of the nonlinear autoregressive module is connected to the output of the FNN module. The nonlinear autoregressive module is based on the single-mass dynamics of a high-speed train. The dynamic feature extractor is based on the multi-mass coupled dynamics of a high-speed train. The dynamic feature extractor includes a first time... The system comprises an attention variant, a second temporal attention variant, a first fusion module, a first spatial attention variant, a second spatial attention variant, and a second fusion module. The output of the input layer is connected to the input of the first temporal attention variant, and the output of the block feature extraction module is connected to the input of the second temporal attention variant. The outputs of both the first and second temporal attention variants are connected to the input of the first fusion module, and the output of the first fusion module is connected to the inputs of both the first and second spatial attention variants. The outputs of both the first and second spatial attention variants are connected to the input of the second fusion module, and the output of the second fusion module is connected to the second input of the FNN module. The first temporal attention variant satisfies the following formula: Where H1 is the output of the first-time attention variant, representing the global temporal features, and softmax() represents the softmax activation function. The input for the first-time attention variant is composed of the unit features of all dynamic units. This represents the first Q-weight matrix. This represents the feature dimension of the value space of the first-time attention variant. express The transpose of the matrix, The first V weight matrix is ​​represented; the output of the block feature extraction module includes all block features of all dynamic units, and the second temporal attention variant satisfies the formula: ,in, This represents the Nth local feature of a multi-dynamic unit, which is composed of the Nth block features of all dynamic units. express The transpose of is given, H2 is the output of the second temporal attention variant, representing local temporal features, and softmax() represents the softmax activation function. This represents the second Q-weight matrix. This represents the feature dimension of the value space of the second temporal attention variant. Let A represent the weight matrix. This represents the second V weight matrix. express The transpose of the matrix, Represents the local characteristic matrix. This indicates that the data within the brackets [ ] will be concatenated along the first dimension. This indicates that the last two dimensions of the matrix within the parentheses are transposed. express Block features after linear processing Indicates and The weight matrix corresponding to the dimension Indicates and The bias term corresponding to the dimension; Obtain the actual average speed of the high-speed train at the current moment and at each historical moment, as well as the unit characteristics of each power unit; The train speed prediction model is obtained by inputting the actual average speed of the high-speed train at the current moment and at each historical moment, as well as the unit characteristics of each power unit, into the train speed prediction model.

2. The train speed prediction method according to claim 1, characterized in that, The output of the first fusion module includes the original features of each dynamic unit at the current time and at each historical time. The first spatial attention variant includes the first dynamic unit update feature unit and the first fusion unit. The first power unit updates the feature unit to satisfy the formula. ,in, This represents the concatenated feature vector of the i-th power unit and the j-th power unit at time k. This represents the original characteristics of the i-th power unit at time k. This represents the original characteristics of the j-th power unit at time k. express The transpose of the matrix, This indicates that the instantaneous adjacency fusion mechanism can learn weight vectors. This represents the leaky ReLU activation function. This represents the attention score between the i-th and j-th power units at time k. This represents the attention score constraint between the i-th dynamic unit and the j-th dynamic unit at time k. This represents the attention score constraint between the i-th dynamic unit and the J-th dynamic unit at time k. This represents the data located in the i-th row and j-th column of the adjacency matrix A. The adjacency matrix A is constructed based on the physical structure of the high-speed train and represents the coupling relationship between power units. This represents the degree of dependence between the i-th power unit and the j-th power unit at time k. N is a learnable factor, exp() represents the exp function, and N c Indicates the total number of power units. The output of the feature unit is updated for the first power unit, representing the updated feature of the i-th power unit at time k. The first fusion unit is used to obtain the adjacency fusion feature matrix based on the updated features of each power unit at the current time and at each historical time.

3. The train speed prediction method according to claim 1, characterized in that, The output of the first fusion module includes the original features of each dynamic unit at the current time and at each historical time. The second spatial attention variant includes the second dynamic unit update feature unit and the second fusion unit. The second power unit updates the feature unit to satisfy the formula. ,in, This represents the original characteristics of the j-th dynamic unit at time kq. The matrix dimension is , express The characteristics of the j-th power unit at time k. This represents the time series characteristics of multiple dynamic units at time k with a lag of q steps. This represents the original characteristics of the i-th power unit at time k. express transpose, Let represent the concatenated vector of the current features of the i-th power unit and the historical features of the j-th power unit at time k, with a delay of q steps. This represents the learnable weight vector of the hysteresis non-adjacent fusion mechanism. This represents the leaky ReLU activation function. This represents the attention score between the i-th and j-th power units when there is a delay of q steps at time k. This represents the data located in the i-th row and j-th column of the adjacency matrix A, which is constructed based on the physical structure of high-speed trains. This represents the attention score constraint between the i-th and j-th dynamic units at time k with a delay of q steps, where exp() represents the exp function, and N c Indicates the total number of power units. This represents a learnable balance factor. This represents the delayed non-adjacent attention weight between the i-th and j-th power units at time k with delay q steps. Let i represent the set of non-adjacent units of the i-th power unit. This represents the upper limit of the interaction hysteresis between non-adjacent dynamic units. The output of the feature unit is updated for the second dynamic unit, representing the hysteresis non-adjacent fusion feature of the i-th dynamic unit at time k; The second fusion unit is used to obtain the hysteresis non-adjacent fusion feature matrix based on the hysteresis non-adjacent fusion features of each power unit at the current time and at each historical time.

4. A train speed prediction device, characterized in that, The train speed prediction device includes: The model building module is used to construct a train speed prediction model. The train speed prediction model includes an input layer, a block feature extraction module, a nonlinear autoregressive module, a dynamic feature extractor, and an FNN module. The output of the input layer is connected to the first input of the nonlinear autoregressive module, the input of the block feature extraction module, the input of the dynamic feature extractor, and the first input of the FNN module. The output of the block feature extraction module is connected to the second input of the nonlinear autoregressive module. The output of the dynamic feature extractor is connected to the second input of the FNN module. The output of the nonlinear autoregressive module is connected to the output of the FNN module. The nonlinear autoregressive module is based on the single-mass dynamics of a high-speed train. The dynamic feature extractor is based on the multi-mass coupled dynamics of a high-speed train. The dynamic feature extractor includes... The system comprises a first temporal attention variant, a second temporal attention variant, a first fusion module, a first spatial attention variant, a second spatial attention variant, and a second fusion module. The output of the input layer is connected to the input of the first temporal attention variant, and the output of the block feature extraction module is connected to the input of the second temporal attention variant. The outputs of both the first and second temporal attention variants are connected to the input of the first fusion module, and the output of the first fusion module is connected to the inputs of both the first and second spatial attention variants. The outputs of both the first and second spatial attention variants are connected to the input of the second fusion module, and the output of the second fusion module is connected to the second input of the FNN module. The first temporal attention variant satisfies the following formula: Where H1 is the output of the first-time attention variant, representing the global temporal features, and softmax() represents the softmax activation function. The input for the first-time attention variant is composed of the unit features of all dynamic units. This represents the first Q-weight matrix. This represents the feature dimension of the value space of the first-time attention variant. express The transpose of the matrix, The first V weight matrix is ​​represented; the output of the block feature extraction module includes all block features of all dynamic units, and the second temporal attention variant satisfies the formula: ,in, This represents the Nth local feature of a multi-dynamic unit, which is composed of the Nth block features of all dynamic units. express The transpose of is given, H2 is the output of the second temporal attention variant, representing local temporal features, and softmax() represents the softmax activation function. This represents the second Q-weight matrix. This represents the feature dimension of the value space of the second temporal attention variant. Let A represent the weight matrix. This represents the second V weight matrix. express The transpose of the matrix, Represents the local characteristic matrix. This indicates that the data within the brackets [ ] will be concatenated along the first dimension. This indicates that the last two dimensions of the matrix within the parentheses are transposed. express Block features after linear processing Indicates and The weight matrix corresponding to the dimension Indicates and The bias term corresponding to the dimension; The acquisition module is used to acquire the actual average speed of the high-speed train at the current moment and at each historical moment, as well as the unit characteristics of each power unit. The prediction module is used to input the actual average speed of the high-speed train at the current moment and at each historical moment, as well as the unit characteristics of each power unit, into the train speed prediction model to obtain the predicted average speed of the train at the next moment.

5. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the train speed prediction method according to any one of claims 1-3.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the train speed prediction method according to any one of claims 1-3.

7. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the train speed prediction method according to any one of claims 1-3.

Images