Energy-saving Driving Strategy for High-Speed Trains Based on Unit and Multivariate Fusion Prediction Model

Abstract

The present invention discloses an energy-saving driving strategy for high-speed trains based on a unit and multi-element fusion prediction model. First, the original data needs to be classified according to the operation handle level. The undersampling method is used to randomly sample the data set to obtain a balanced data set. The redundant features in the balanced data set are removed through a wrapper feature selection algorithm based on the gradient boosting algorithm to obtain a training set. Finally, the training set after feature selection is input into the unit and multi-element fusion prediction model for training. A route containing all operation handle levels is selected as the test set. After processing the data in the same way as the training set, it is input into the trained model to predict the operation handle level. Through repeated training and optimization, an energy-saving strategy is finally obtained. The strategy in the present invention can be used in future train driving, which not only achieves the effect of energy saving, but also has low cost and meets the requirements of the country for high-speed train driving.

Description

Technical Field

[0001] The present invention belongs to the technical fields of high-speed train driving, neural networks, and machine learning, and specifically relates to an energy-saving driving strategy for high-speed trains based on a unit and multi-element fusion prediction model. Background Art

[0002] At present, high-speed trains have become an important means of transportation for the public due to their advantages such as comfort, punctuality, and less time consumption. As a means of transportation for medium and short-distance travel in China, they can save a large amount of time during the journey and have become the best choice for the vast majority of people.

[0003] In recent years, high-speed railway transportation has been developing rapidly. With the increase in the operating mileage and train operation density of high-speed railways, the power consumption of the traction power supply system has been increasing year by year, and the energy conservation and consumption reduction of high-speed railways have become a hot topic of concern. One of the most significant advantages of high-speed trains compared to other means of transportation is that they can arrive on time with very few delays, which requires speed planning and control during the train operation. According to statistics, the traction energy consumption of high-speed trains accounts for the largest proportion. Therefore, on the basis of ensuring the safe and punctual operation of the train, the traction energy consumption of high-speed trains can be effectively reduced by reducing it, achieving the effect of effectively reducing the train's energy consumption.

[0004] China's high-speed railway is an important type of transportation infrastructure in China. As of the end of 2021, the national railway operating mileage exceeded 1.5 million kilometers, and the operating mileage of high-speed railways exceeded 40,000 kilometers, ranking first in the world. The low-carbon and energy-saving functions of high-speed railway transportation are obvious, and it has become the backbone of China's modern transportation system. At the same time, high-speed railways also play a promoting role in the development of politics, economy, and culture, and play a positive driving role in the development of industry. While experiencing its fast and convenient nature, we are also clearly aware of its energy consumption problem. Therefore, the research on the energy-saving driving of high-speed trains has become a very important topic. Summary of the Invention

[0005] The purpose of the present invention is to provide an energy-saving driving strategy for high-speed trains based on a unit and multi-element fusion prediction model, which solves the problem of excessive energy consumption caused by improper operation during the operation of high-speed trains.

[0006] The technical solution adopted by the present invention is an energy-saving driving strategy for high-speed trains based on a unit and multi-element fusion prediction model, and the specific operation steps are as follows:

[0007] Step 1: Classify the runtime data according to the operating handle level into 12 categories according to different operating handle levels. The runtime data includes runtime, speed, and temperature. During the use of the data, the data of the passenger-activated emergency operation position will be removed, and the operating handle level labels in the remaining data are respectively recorded as 0-10. The undersampling method is used to perform undersampling operations on each category of data to obtain a balanced dataset with equal numbers of each category;

[0008] Step 2: Use the wrapper feature selection method based on the extreme gradient boosting algorithm to perform feature selection on the balanced dataset obtained in Step 2, and finally obtain 37 features related to the operating handle level. This balanced dataset will be used as the training set;

[0009] Step 3: Use the training set described in Step 2 as the input of the unit and multi-element fusion prediction model, and train the unit and multi-element fusion prediction model to obtain the classification results at different operating handle levels;

[0010] Step 4: First, select a route for testing. The selected route requires that in the record of the initial runtime data, the original operating handle level contains eleven different categories, and calculate the traction power consumption of the original runtime data according to the energy consumption value recorded in the initial runtime data;

[0011] Step 5: After the test set undergoes feature selection, input it into the trained unit and multi-element fusion prediction model to obtain the initial driving operation strategy Y. Finally, according to the predicted initial operation strategy Y and the energy consumption increase value brought by each operating handle level per second calculated from the initial runtime data, calculate the traction power energy Q' under the strategy Y;

[0012] Step 6: Compare the traction power energies obtained in Steps 4 and 5, adjust the network model, and seek the optimal operation strategy;

[0013] The characteristics of the present invention also lie in that,

[0014] Step 1 is specifically as follows:

[0015] Step 1.1: Classify the runtime data according to the brake handle level. The values of the brake handle level are several discrete quantities: {EB, REL, 1A, 1B, 2, 3, 4, 5, 6, 7, 8, OC}. EB is the emergency brake position, represented by the label 0; REL is the running position, that is, the traction position, represented by the label 1; the normal brake positions 1A, 1B, 2, 3, 4, 5, 6, 7, 8 are respectively represented by 2, 3, 4, 5, 6, 7, 8, 9, 10, and OC is to ignore the passenger-activated emergency brake position, represented by the label 11. Thus, the brake handle level is mapped to: {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11}, a total of twelve categories;

[0016] Step 1.2: Exclude data representing passengers activating the emergency braking position category with a brake handle level of 11; using the undersampling method, randomly select 60 groups of data from the data with brake handle level labels from 1 to 10, obtaining a total of 660 pieces of data.

[0017] Step 2 is specifically as follows:

[0018] Step 2.1: Use the wrapper feature selection algorithm based on the extreme gradient boosting algorithm to perform feature selection operations on the balanced dataset, and filter out features with zero importance according to the importance measure of features;

[0019] Step 2.2: Utilize the constructed tree model structure to sort the features with non-zero importance, and finally retain thirty-seven features with non-zero importance in the data as the features in the training set.

[0020] The unit in Step 3 and the unit model in the multi-element fusion prediction model is the ARIMA model, and the multi-element model is the GRU model. Using the Boosting idea and the waterfall fusion method, the ARIMA model and the GRU model are connected in series to form the unit and multi-element fusion prediction model; take the features other than the operation handle level in the training set, namely speed, acceleration, and bearing temperature attributes as inputs, use the ARIMA model for initial prediction to obtain the prediction result of the operation handle level, and compare the obtained prediction result with the original operation handle level in the training set to obtain the residual. Take the residual as a new feature, and combine it with the training set as the input to use the GRU model to re-predict the operation handle level; specifically as follows:

[0021] Step 3.1: Using the Boosting idea and the waterfall fusion method, connect the ARIMA model and the GRU model in series to construct the unit and multi-element fusion prediction model;

[0022] Step 3.2: Input the features other than the operation handle level in the training set, namely time, speed, and bearing temperature, into the ARIMA module of the unit and multi-element fusion prediction model, and use the ARIMA module to make a preliminary prediction of the operation handle level; compare the obtained operation handle level with the operation handle level in the training set to obtain the residual, and re-combine the residual with the remaining features other than the operation handle level in the training set into a new input dataset;

[0023] Step 3.3: Input the new input dataset obtained in Step 3.2 into the GRU module of the unit and multi-element fusion prediction model to make a discrete value prediction, and obtain the classification result through the Softmax classifier, that is, the classification results of different operation handle levels.

[0024] The GRU model consists of three layers: The first layer is the gru layer, which uses a gating mechanism to control information such as input and memory, and makes predictions at the current time step; The second layer is the fully connected layer, which maps n real numbers in (-∞, +∞) to K real numbers in (-∞, +∞); The third layer is the softmax layer, which maps K real numbers in (-∞, +∞) to K real numbers in (0, 1), while ensuring that their sum is 1.

[0025] Step 4 is specifically as follows:

[0026] Select a route from the runtime data as the test route, which is required to include eleven different types of operating handle positions. Calculate the total traction power consumption Q on this route according to the traction power consumption difference between the start and end time points, and calculate the percentages of the three states of traction, cruise, and braking under each label respectively, denoted as a i,0 , a i,1 , a i,2 , where i represents the category of different labels, i represents the operating handle position label, and the values are 0, 1, 2,..., 10.

[0027] Step 5 is specifically as follows:

[0028] Step 5.1: Select the features in the data of the test route data through the wrapper feature selection algorithm based on the gradient boosting algorithm to obtain thirty-seven features, and input the data containing thirty-seven features into the trained unit and multi-variable fusion prediction model to obtain the predicted operation strategy;

[0029] Step 5.2: Use r 0 , r 1 , …, r 10 to represent the number of different operating conditions of each type of operating handle position, and obtain the number Z of operating handle positions that consume power among the 11 operating handle positions, as shown in Equation (8):

[0030]

[0031] When the operating handle position is 1, it only consumes power to accelerate, so r 1 does not need to be divided into situations of acceleration, constant speed, and deceleration;

[0032] Step 5.3: Calculate the instantaneous energy consumption of each type of label with the help of excel software, and divide the sum of the instantaneous energy consumption by the number of operating handle positions to obtain the average energy consumption increase value of each operating handle position, denoted as: q 0 , q 1 , …, q 10 , from which the energy consumption Q' required for the strategy obtained by the unit and multi-variable fusion prediction model can be expressed as Equation (9):

[0033]

[0034] Step 6 is specifically as follows:

[0035] Step 6.1: From Steps 4 and 5, the power consumption of the test set in the initial runtime data can be obtained, that is, Q, and the power consumption under the operation strategy formulated by the unit and multi - element fusion prediction model is Q';

[0036] Step 6.2: If the power consumption relationship is Q > Q', then the strategy is determined as the initial energy - saving strategy Y. If the train cannot reduce its speed to 0 before reaching the terminal according to the calculation, then in order to meet the punctuality of the train, the time of uniform motion is increased to ensure that the train speed reaches 0 when it reaches the terminal;

[0037] Step 6.3: If the power consumption relationship is Q ≤ Q', then adjust the number of training times, the number of samples grabbed in one training, and the learning rate to achieve the power consumption relationship of Q > Q', and thus obtain the energy - saving strategy.

[0038] The beneficial effects of the present invention are as follows:

[0039] 1. The present invention transforms the problem of high - speed train energy consumption into a classification problem of operation handle positions, and then transforms the classification problem into a decision - making problem;

[0040] 2. The present invention constructs a unit and multi - element fusion prediction model, which is different from the traditional methods for solving energy consumption problems and is a combination of linear and non - linear, and can obtain the best energy - saving strategy faster.

[0041] 3. When selecting features, the present invention adopts a wrapper - type feature selection algorithm based on the gradient boosting algorithm. This algorithm can not only help the classifier improve the classification accuracy, but also minimize redundant features as much as possible and avoid the limitations of a single important metric. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 is the overall flowchart of the energy - saving driving strategy of the high - speed train based on the unit and multi - element fusion prediction model of the present invention;

[0043] Figure 2 is the convergence graph of the LOSS value of the model training of the energy - saving driving strategy of the high - speed train based on the unit and multi - element fusion prediction model of the present invention;

[0044] Figure 3 is the accuracy graph of the model training of the energy - saving driving strategy of the high - speed train based on the unit and multi - element fusion prediction model of the present invention. DETAILED IMPLEMENTATION MANNER

[0045] The energy - saving driving strategy of the high - speed train based on the unit and multi - element fusion prediction model of the present invention, as Figure 1As shown in the figure, first, the data generated during the operation of the train from the starting station to the terminal station, such as operation time, speed, driving side shaft temperature, gear temperature, and the position of the operation handle, etc., are used to construct a balanced data set using the undersampling method according to different types of the position of the operation handle; then, the wrapped feature selection algorithm based on the extreme gradient boosting algorithm is used to perform feature selection on the constructed balanced data set to obtain the features highly correlated with the position of the operation handle, that is, the input features of the final unit and multi - variable fusion classification model, including speed, acceleration, etc.; the model input features are input into the unit and multi - variable fusion prediction model for training, and finally, the trained unit and multi - variable fusion prediction model is used to predict the position of the operation handle for the test route data to obtain the predicted handle position, that is, the optimized control strategy; calculate the energy consumption according to the predicted strategy and compare it with the initial energy consumption in the test set route to see if it is energy - saving. If the energy consumption decreases, the energy - saving strategy is obtained; otherwise, adjust the model and re - perform the prediction; the specific steps are as follows.

[0046] Step 1 is to perform undersampling operation on the data according to different types of the position of the operation handle to obtain a data set with balanced position of the operation handle. This is a method to alleviate class imbalance and is achieved by discarding samples, that is, performing less sampling on the category with a larger number of samples in the training set. The specific operations are as follows:

[0047] Step 1.1: Classify the data according to the different positions of the operation handle in the operation data. During the actual operation of the high - speed train, the driver controls the train through the following positions of the operation handle: {EB, REL, 1A, 1B, 2, 3, 4, 5, 6, 7, 8, OC}.

[0048] Table 1 Representation of the position of the operation handle

[0049]

[0050] As shown in Table 1, EB is the emergency operation position, represented by the label 0, REL is the running position, that is, the traction position, represented by the label 1, and the common operation positions are: 1A, 1B, 2, 3, 4, 5, 6, 7, 8, which are respectively represented by 2, 3, 4, 5, 6, 7, 8, 9, 10 in the present invention, and OC is the position of ignoring the passenger - activated emergency operation, represented by the label 11. After the above label representation, the position of the operation handle can be mapped to: {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11}, so the operation data is divided into twelve categories.

[0051] Step 1.2: In the given initial operating data sample, it can be divided into 12 categories according to the different levels of the operating handle. According to statistics, when the operating position is OC, there are only 5 pieces of data in the entire data sample, and this type of data cannot play a role in the training of the model, so it is excluded. Therefore, only the remaining 11 categories of data need to be undersampled. Undersampling is a process of randomly deleting a part of the majority class (the type with a large number) of data. Based on the minority class in these 11 categories of data, that is, the number of data with label 0, random sampling is performed on the remaining 10 categories of data to obtain 660 balanced sample sets.

[0052] Step 2 is a feature selection process. The present invention uses a wrapper feature selection method based on the extreme gradient boosting algorithm for feature selection, and finally obtains thirty-seven features related to the operating handle level. This balanced data set will be used as the training set, and the data obtained after feature selection will be used as the training set to train the model. The specific operations are as follows:

[0053] Step 2.1: The wrapper feature selection method based on the extreme gradient boosting algorithm performs feature selection on the basis of the tree model. In the construction of the tree model, features with zero importance, that is, features with zero feature node splitting times, will not be used to split any nodes, and removing them will not affect the final model performance. Therefore, the present invention uses the wrapper feature selection algorithm based on the extreme gradient boosting algorithm to perform feature selection operations on the balanced data set, and filters out features with zero importance such as transformer temperature, motor stator temperature, and all operations within the unit are alleviated according to the importance measure of the features;

[0054] Step 2.2: Using the constructed tree model structure, sort the features with non-zero importance, and perform two-way search using different importance measures to obtain better results. Finally, thirty-seven features with non-zero importance including speed, acceleration, gear temperature, etc. are retained in the data, and this data will be used as the training set;

[0055] Step 3 uses the data obtained in Step 2 as the training set to train the unit and multi-element fusion prediction model constructed in the present invention;

[0056] Step 3.1: Construct a unit and multi-element fusion prediction model: The unit model is an ARIMA model, and the multi-element model is a GRU model. Using the Boosting idea and the waterfall fusion method, the ARIMA model and the GRU model are connected in series to construct a unit and multi-element fusion prediction model. The core idea of Boosting is to use another model to train the prediction error samples after each layer of training, that is, to combine several weak classifiers into a strong classifier. In the present invention, both the ARIMA model and the GRU model are weak classifiers, and the unit and multi-element fusion prediction model constructed by them is a strong classifier;

[0057] Step 3.2: Input the features other than the operating handle level in the training set, such as time, speed, bearing temperature, etc., into the first part of the multi-source fusion prediction model, the ARIMA part, that is, the autoregressive integrated moving average regression model, also known as ARIMA(p,d,q), which is a commonly used statistical model for time series prediction. It only uses its own change information to achieve future prediction. Essentially, it is used to capture linear relationships. Here, it is used as the first part of the fusion prediction model to make a preliminary prediction of the operating handle level. As shown in ARIMA(p,d,q), the ARIMA model has three parameters p, d, and q, and their meanings are as follows:

[0058] p represents the number of lags of the time series data used in the prediction model, also known as AR (autoregressive).

[0059] d represents the number of differences the time series data needs to go through to be stable, called the integration term.

[0060] q represents the lag term of the prediction error in the prediction model, also known as the MA (Moving Average) term.

[0061] According to the data set adopted by the present invention, the operating handle level is not very stable and needs to be differenced before it can become a stable series, which means that d cannot be set to 0. The values of p and q are determined by the Bayesian Information Criterion (BIC). After experimental verification, it is most appropriate to set d to 1, and p and q to 3 and 3 respectively.

[0062] Therefore, the mathematical expression of the ARIMA model is as follows:

[0063]

[0064] BIC = -2ln(S) + ln(n) * r (2)

[0065] Among them, μ represents the constant term, α 1 ,α 2 ,…,α p represents the weight coefficients of the lag time series data, β 1 ,β 2 ,…,β q are the coefficients of the error term; h t represents the predicted value of ARIMA at time t, S is the maximum likelihood number under this model, n is the size of the data volume, and r is the number of model variables.

[0066] The present invention uses the ARIMA module to make a preliminary prediction of the operating handle level. The obtained operating handle level H tThe residuals are obtained by comparing with the operating handle level positions in the training set, and the residuals and the remaining features in the training set except the operating handle level positions are recombined into a new input data set.

[0067] Step 3.3: Input the new input data set obtained in Step 3.2 into the second part of the unit and multi-source fusion prediction model, the GRU part, to obtain the final intelligent driving strategy. The GRU model consists of three layers: the first layer is the gru layer, which uses a gating mechanism to control information such as input and memory and makes predictions at the current time step; the second layer is the fully connected layer, which maps n real numbers in (-∞, +∞) to K real numbers in (-∞, +∞); the third layer is the softmax layer, which maps K real numbers in (-∞, +∞) to K real numbers in (0, 1) while ensuring that their sum is 1.

[0068] Train the GRU model with the new input data set obtained in Step 3.2, and its parameter settings are as follows: the input dimension input_size is 37, the output dimension output_size is 1, the hidden layer dimension hidden_size is 37, the number of data samples grabbed in one training batch_size is 30, and the number of GRU layers is 2. Train in the GRU under the above structure.

[0069] GRU includes parts such as a reset gate, an update gate, a candidate hidden state, and a hidden state. The present invention uses a bidirectional GRU to simultaneously obtain time series information in a positive-negative combination manner and merge the results to classify the braking handle level positions as a whole, and can also effectively solve the problems of gradient disappearance and gradient explosion. The residuals and the remaining features in the training set except the braking handle level positions are used as time series and input into the forward GRU and the backward GRU at time t respectively, and the forward output is obtained through Equation (2) and Equation (3). and the backward output and the forward and backward outputs are concatenated according to Equation (4) to obtain the output H of the bidirectional GRU. i , and the results are merged into v = (H 1 , H 2 , …, H i , …, H n-1 , H n ), generating a two-dimensional matrix as the final output of the hidden layer and entering the Softmax layer for classification.

[0070]

[0071]

[0072] where σ is the activation function, W t and W′ t$W_{t}$ is the input weight matrix of the forward / backward GRU at time $t$, $x$ is the input at time $t$, $U$ t and $U'$ t are the hidden layer weight matrices of the forward / backward GRU at time $t - 1$.

[0073]

[0074] According to the Softmax() function, the output result is activated according to Equation (4), and the final feature vector $u$ can be obtained. Finally, $u$ is output as the probabilities of different categories through the output layer, and the maximum probability is used as the classification result of the operation handle, and the final driving strategy is obtained therefrom.

[0075] $v = \text{Softmax}(u)\ (6)$

[0076] To prove the effectiveness of the unit and multi - element fusion prediction model, the Loss value and classification accuracy are added to the experimental results, such as Figure 2 、 Figure 3 . The gradually decreasing Loss value and continuously improving classification accuracy indicate that the unit and multi - element fusion prediction model is effective for the dataset used in the present invention.

[0077] Step 4: Select a complete route as the test route according to the speed and acceleration changes in the runtime data. This route is required to include eleven different categories of operation handle levels, and then calculate the traction power consumption of the original runtime data. The specific calculation process is as follows:

[0078] Assume that the value of the power consumption of the selected route at the starting position is $Q$ 1 , and the value at the end position is $Q$ 2 . Assume that the total power consumption on this route is $Q$, and the expression is as shown in Equation (7):

[0079] $Q = Q$ 2 - $Q$ 1 (7)

[0080] Among them, except for the traction handle, the same operation handle level can be divided into three operating conditions: acceleration, constant speed, and deceleration. The traction power consumption is the largest during the acceleration stage. During the constant - speed stage, the traction energy required for the high - speed train only needs to offset the air resistance and its own friction. During the braking stage, no electrical energy is consumed, and the traction power consumption is 0 at this time.

[0081] According to the positive and negative conditions of the acceleration, count the quantities of the three conditions of acceleration, constant speed, and deceleration for the same operation handle, and calculate the percentage of each operating condition. The percentages of the three conditions of acceleration, constant speed, and deceleration are respectively denoted as $a$ i,0 , $a$ i,1 , $a$ i,2 , where $i$ represents the operation handle level label, and the value range is 0, 1, 2, ……, 10.

[0082] In step 5, the test set selected in step 4 is subjected to feature selection in step 2, and then input into the trained unit and multi - variable fusion prediction model. After obtaining the predicted operation strategy, the following energy consumption calculation is performed:

[0083] Step 5.1: Select thirty - seven features from the test route data through the wrapper feature selection algorithm based on the gradient boosting algorithm. Input the data containing thirty - seven features into the trained unit and multi - variable fusion prediction model to obtain the predicted operation strategy;

[0084] Step 5.2: Use the Excel tool to count the number of different operation handle levels, denoted by r 0 , r 1 , …, r 10 respectively. According to the percentages described in step 4, obtain the number of different operating conditions for each category, and then obtain the number of energy - consuming operation handle levels Z, as shown in Equation (8):

[0085]

[0086] When the operation handle level is 1, only electric energy is consumed to accelerate, so r 1 does not need to be divided into acceleration, constant speed, and deceleration situations;

[0087] Step 5.3: Perform energy consumption calculation: First, use the excel software to calculate the instantaneous energy consumption of each category of labels. Divide the sum of the instantaneous energy consumption by the number of operation handle levels to obtain the average energy consumption increase value for each operation handle level, denoted as: q 0 , q 1 , …, q 10 . Then, the energy consumption required for the model strategy finally obtained by the unit and multi - variable fusion prediction model can be expressed as Equation (9):

[0088]

[0089] Step 6: According to the result feedback, adjust the network model and seek the most suitable operation strategy according to the calculation results. The specific steps are as follows:

[0090] Step 6.1: The electric energy consumption Q of the test set in the initial operation data can be obtained from steps 4 and 5, and the energy consumption is Q′ under the operation strategy Y formulated by the unit and multi - variable fusion prediction model;

[0091] Step 6.2: If the energy consumption relationship is Q > Q′, then the strategy is determined as the initial energy-saving strategy Y. During the driver's driving process, the energy-saving strategy is implemented according to the actual situation. If the train cannot reduce its speed to 0 before reaching the terminal according to the calculation, then in order to meet the punctuality of the train, the time of uniform motion is increased to ensure that the speed of the train reaches 0 when it arrives at the terminal;

[0092] Step 6.3: If the energy consumption relationship is Q ≤ Q′, the number of training times, the number of grabs in one training, and the learning rate need to be adjusted to achieve the energy consumption relationship of Q > Q′, so as to obtain the energy-saving strategy.

Claims

1. High-speed train energy-saving driving strategy based on unit and multi-element fusion prediction model, Characterized in that, The specific operation steps are as follows: Step 1: Classify the operation data according to the different levels of the operation handle during operation into 12 categories. The operation data includes operation time, speed, and temperature. During the use of the data, the data of the passenger-activated emergency operation position will be removed, and the operation handle level labels in the remaining data are respectively recorded as 0-10. The undersampling method is used to perform undersampling operations on each category of data to obtain a balanced data set with equal numbers in each category; Step 2: Use the wrapper feature selection method based on the extreme gradient boosting algorithm to perform feature selection on the balanced data set obtained in Step 2, and finally obtain thirty-seven features related to the operation handle level. This balanced data set will be used as the training set; Step 3: Use the training set described in Step 2 as the input of the unit and multi-element fusion prediction model to train the unit and multi-element fusion prediction model to obtain classification results at different operation handle levels. Specifically as follows: The unit model in the unit and multi-element fusion prediction model is an ARIMA model, and the multi-element model is a GRU model. Using the Boosting idea and the waterfall fusion method, the ARIMA model and the GRU model are connected in series to form a unit and multi-element fusion prediction model; the features other than the operation handle level in the training set, that is, speed, acceleration, and bearing temperature attributes, are used as inputs, and the ARIMA model is used for initial prediction to obtain the operation handle level prediction result, and the obtained prediction result is compared with the original operation handle level in the training set to obtain the residual error, and the residual error is used as a new feature, and combined with the training set as the input to use the GRU model to re-predict the operation handle level. Specifically as follows: Step 3.1: Using the Boosting idea and the waterfall fusion method, connect the ARIMA model and the GRU model in series to construct a unit and multi-element fusion prediction model; Step 3.2: Input the features other than the operation handle level in the training set, that is, time, speed, and bearing temperature, into the ARIMA module of the unit and multi-element fusion prediction model, and use the ARIMA module to make a preliminary prediction of the operation handle level; compare the obtained operation handle level with the operation handle level in the training set to obtain the residual error, and recombine the residual error with the remaining features other than the operation handle level in the training set into a new input data set; Step 3.3: Input the new input data set obtained in Step 3.2 into the GRU module of the unit and multi-variable fusion prediction model for discrete value prediction, and obtain the classification result through the classifier, that is, the classification results of different operating handle levels; Step 4: First select a route for testing. The selected route requires that in the record of the initial operation data, the original operation handle level includes eleven different categories, and calculate the traction power consumption of the original operation data according to the energy consumption value recorded in the initial operation data; Step 5: After feature selection of the test set, input it into the trained unit and multi - element fusion prediction model to obtain the initial driving operation strategy , finally, according to the initially predicted operation strategy and the energy consumption increase value brought by each operating handle level per second calculated from the initial runtime data, calculate the traction power consumption under the strategy ; Step 6: Compare the traction power obtained in Steps 4 and 5, adjust the network model, and seek the best operation strategy. Specifically as follows: Step 6.1: From Steps 4 and 5, the power consumption of the test set in the initial runtime data can be obtained, that is , and the energy consumption under the operation strategy formulated by the unit and multi-element fusion prediction model is ; Step 6.2: If the energy consumption relationship is , then the strategy is determined as the initial energy-saving strategy , if the train cannot reduce its speed to 0 before reaching the terminal according to the calculation, then in order to meet the punctuality of the train, the time of uniform motion is increased to ensure that the speed of the train reaches 0 when it arrives at the terminal; Step 6.3: If the energy consumption relationship is , then adjust the number of training times, the number of grabs per training, and the learning rate to achieve an energy consumption relationship of , thereby obtaining an energy-saving strategy.

2. The high-speed train energy-saving driving strategy based on the unit and multi-element fusion prediction model according to claim 1, Characterized in that, Step 1 is specifically as follows: Step 1.1: Classify the runtime data according to the brake handle position. The values of the brake handle position are several discrete quantities: , EB is the emergency brake position, represented by label 0; REL is the running position, i.e., the traction position, represented by label 1; the normal brake positions 1A, 1B, 2, 3, 4, 5, 6, 7, 8 are respectively represented by 2, 3, 4, 5, 6, 7, 8, 9, 10, and OC is the position of ignoring the passenger-activated emergency brake, represented by label 11. Thus, the brake handle position is mapped to: , a total of twelve categories; Step 1.2: Eliminate the data representing passengers activating the emergency braking position category with the brake handle level at 11; Using the undersampling method, randomly select 60 groups of data from the data with the brake handle level labels from 1 to 10, and a total of 660 pieces of data are obtained.

3. The energy-saving driving strategy for high-speed trains based on the unit and multi-element fusion prediction model according to claim 2, characterized in that, Step 2 is specifically as follows: Step 2.1: Use the wrapper feature selection algorithm based on the extreme gradient boosting algorithm to perform feature selection operations on the balanced data set, and filter out the features with zero importance according to the importance measure of the features; Step 2.2: Utilize the constructed tree model structure to sort the features with non-zero importance, and finally retain thirty-seven features with non-zero importance in the data as the features in the training set.

4. The energy-saving driving strategy for high-speed trains based on the unit and multi-element fusion prediction model according to claim 3, characterized in that, The GRU model consists of three layers: The first layer is the gru layer, which uses a gating mechanism to control information such as input and memory, and makes predictions at the current time step; The second layer is a fully connected layer, which maps n real numbers to K real numbers; The third layer is the softmax layer, which maps K real numbers to K real numbers in the range (0, 1), while ensuring that their sum is 1.

5. The energy-saving driving strategy for high-speed trains based on the unit and multi-element fusion prediction model according to claim 3, characterized in that, Step 4 is specifically as follows: Select a route from the runtime data as the test route, which is required to include eleven different categories of operation handle positions; calculate the total traction power consumption on this route based on the traction power consumption difference between the start and end time points. , and calculate the percentages of the three states of traction, cruise, and braking under each label respectively, denoted as , represents the category of different labels, represents the operation handle position label, with values of 0, 1, 2,..., 10.

6. The energy-saving driving strategy for high-speed trains based on the unit and multi-element fusion prediction model according to claim 5, characterized in that, Step 5 is specifically as follows: Step 5.1: Select the features in the test route data through the wrapper feature selection algorithm based on the gradient boosting algorithm to obtain thirty-seven features, and input the data containing thirty-seven features into the trained unit and multi-element fusion prediction model to obtain the prediction operation strategy; Step 5.2: Respectively use , , …, to represent the quantity of different operating conditions of each type of operating handle level, and obtain the quantity of the operating handle levels consuming electric energy among 11 operating handle levels, as shown in Equation (8): (8) When the operating handle level is 1, only electric energy is consumed to accelerate, so There is no need to distinguish between acceleration, constant speed, and deceleration; Step 5.3: Use excel software to calculate the instantaneous energy consumption of each type of tag. Divide the instantaneous energy consumption by the number of operating handle levels to obtain the average energy consumption increase value for each operating handle level, which is expressed as: , , …, , thus obtaining the energy consumption required for the strategy obtained by the unit-multivariate fusion prediction model , which can be expressed as Equation (9): (9)。

Images