Energy-saving driving strategy for high-speed trains based on SA-GRU

Abstract

The present invention discloses an energy-saving driving strategy for high-speed trains based on SA‑GRU. First, the initial runtime data needs to be preprocessed, and then the processed data is classified according to the brake handle level. The data set is randomly sampled by undersampling method to obtain a balanced data set. The redundant features in the balanced data set are removed through multiple feature selection to obtain a training set. Finally, the training set after multiple feature selection is input into the SA‑GRU classification model to train the classification model. The test set selects a route containing all brake handle levels. After the data is processed in the same way as the training set, it is input into the trained model to predict its brake handle level. After repeated training and optimization, an energy-saving strategy is finally obtained. The strategy of 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 national requirements for energy-saving driving of high-speed trains.

Description

Technical Field

[0001] The invention belongs to the technical field of rail transportation and neural network, and specifically relates to a study on an energy-saving driving strategy for high-speed trains based on SA-GRU. Background Art

[0002] In recent years, my country's high-speed railway has developed rapidly. The opening and operation of high-speed railway has not only promoted the development of my country's social economy, but also, as a mode of transportation for domestic short-distance travel, it can save a lot of time on the journey and become the best choice for most people. While promoting the development of the national economy, high-speed railway has also produced a good demonstration effect in energy conservation and emission reduction, which is specifically reflected in the following aspects: first, it has improved the level of railway electrification, optimized the railway energy-saving structure, and achieved a large-scale "replacing oil with electricity", reducing dependence on oil; second, high-speed railway technology has effectively improved the energy conservation and emission reduction effect of the railway industry and formed a green and environmentally friendly transportation artery. However, with the increase in the mileage and traffic density of high-speed railways, the power consumption of traction power supply systems has increased year by year, and 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 railways over other modes of transportation is that they can arrive at the station on time with very few delays, which requires speed planning and control during the operation of trains. Therefore, under the premise of ensuring the safety, reliability and punctuality of the high-speed train system, it is of great practical significance to carry out research on key theoretical technologies and methods for reducing the energy consumption level of high-speed train operations. Existing energy-saving methods mainly focus on the optimization research of the vehicle body itself, track routes and train driving strategies. Except for the train driving strategy, the remaining two have high investment costs. Optimizing the driving strategy has become an efficient and low-cost method. Summary of the invention

[0003] The purpose of the present invention is to provide an energy-saving driving strategy for high-speed trains based on SA-GRU, which solves the problem of excessive energy consumption caused by improper operation during the operation of high-speed trains.

[0004] The technical solution adopted by the present invention is a high-speed train energy-saving driving strategy based on SA-GRU, step 1, using interpolation method and chi-square test method to pre-process the running data generated during the running process of the train from the starting station to the terminal station for a series of single features such as running time, speed, carriage electrical cabinet temperature, transformer temperature, drive side shaft temperature, gear temperature, shaft temperature, speed, duct temperature and brake handle level;

[0005] Step 2: Divide the data into 12 categories according to 12 different categories of brake handle levels, and map each category of brake handle levels to numbers from 0 to 11 as labels, remove the data of the emergency brake handle level with too few numbers, that is, the data with label 11, and for the remaining train running data, the labels are 0 to 10, and the undersampling method is used to construct a balanced data set operation to obtain a balanced data set with the same number of categories;

[0006] Step 3, multiple feature selection is performed on the balanced data set in step 2 to remove features that have no effect or little effect on the brake handle level operation. After multiple selection, the data in the balanced data set obtained only contains thirty-four features with a large correlation with the brake handle level. This balanced data set will be used as a training set;

[0007] Step 4: Use the data except the brake handle level in the training set after multiple feature selection in step 3 as the input of the SA-GRU classification model, and use the brake handle level as the output to train the SA-GRU classification model;

[0008] Step 5: Select a route as the test set from all the data after preprocessing in step 1 according to the speed change, calculate the difference of the traction power consumption from the starting station to the terminal station recorded in the data set to obtain the traction power consumption J under the original braking strategy in this section of the road, and record it; then calculate the percentage of acceleration, uniform speed, and deceleration under each label according to the operating conditions in the initial running data, recorded as p i，0 , p i，1 , p i，2 ;

[0009] Step 6: After multiple feature selection, the test set is input into the trained SA-GRU model to obtain the initial braking strategy B. Since electric energy is consumed only in the acceleration and constant speed stages, and no electric energy is consumed in the deceleration stage, the number of energy consumption labels L can be obtained through the percentage in step 5 and the number of each brake handle level label in strategy B obtained by the SA-GRU model; finally, the traction power consumption Q under strategy B is calculated based on the predicted initial braking strategy B and the average energy consumption increase per second brought by each brake handle level calculated from the initial running data;

[0010] Step 7: Compare and analyze the traction power consumption in step 5 and step 6, adjust the network model, and seek the best energy-saving strategy.

[0011] The present invention is also characterized in that:

[0012] Step 1 is as follows:

[0013] Step 1.1: Remove the attributes that will not change during the operation process, that is, the data generated during the operation of the train from the starting station to the terminal station, that is, the data that is irrelevant to the train driving energy consumption, which is not only invalid for the classification model training, but also interferes with the model training;

[0014] Step 1.2: Perform a chi-square test based on the relationship between each attribute and the brake handle level. The chi-square test value represents the degree of association between each attribute and the brake handle level. The 47 features with the highest chi-square value are retained.

[0015] Step 1.3: Perform interpolation to fill missing values ​​in the acquired running data. Use the values ​​of the two timestamps above and below the missing value to calculate the average. The obtained value is used to fill in the missing value. At this time, the final complete data containing 47 features is obtained.

[0016] Step 2 is as follows:

[0017] Step 2.1: 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 label 0; REL is the operation position, that is, the traction position, represented by label 1; the common brake positions 1A, 1B, 2, 3, 4, 5, 6, 7, 8 are represented by 2, 3, 4, 5, 6, 7, 8, 9, 10 respectively, and OC is the emergency brake position that ignores passenger activation, represented by label 11. Therefore, the brake handle level is mapped to: {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11}, a total of twelve categories;

[0018] Step 2.2: Remove the data with brake handle level 11, which represents the category of passenger-activated emergency brake. Use the undersampling method to randomly extract 60 groups of data from the data with brake handle level labels 1-10, and obtain a total of 660 data.

[0019] Step 3 is as follows:

[0020] Step 3 uses multiple feature selection to select data feature attributes, using traditional feature selection methods and XGBOOST for feature selection. The traditional feature selection method uses variance method and mutual information method, as follows:

[0021] Step 3.1: Feature selection based on variance: Perform feature selection on the 47 features in step 1.3, leaving K features with large variance, where 1≤K≤47; calculate the f-value of the linear regression test of a single feature and the brake handle level label, adjust the number of features to be selected, and select features from high to low according to the scoring criteria, with a K value of 35;

[0022] Step 3.2: Feature selection based on mutual information: After arranging the mutual information values ​​of the single features in step 1 and the brake handle levels in descending order, select K features to continue the operation, 1≤K≤47, and the K value is 35;

[0023] Step 3.3: Merge the features selected by the two feature selection methods in step 3.1 and step 3.2 to obtain K1 feature attributes;

[0024] Step 3.4: Encode the K1 feature attributes obtained in step 3.3. Each code corresponds to a feature attribute. The feature attribute coding range is 1-K1. Code 0 is the brake handle level label. Use XGBOOST to perform feature selection on the data set, and rank the relative importance of each feature attribute and the brake handle level. Finally, select 34 features with high importance as the feature attributes of the final training set.

[0025] Step 4 is as follows:

[0026] The balanced data set obtained in step 3 is used as the training set to input the SA-GRU classification model for training. The SA-GRU classification model consists of three layers: the first layer is the gru layer, which uses a gating mechanism to control input, memory and other information, and makes predictions at the current time step; the second layer is the 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 (0, 1) real numbers, while ensuring that their sum is 1;

[0027] Step 4.1: Use the final data obtained in step 3 as the input data of the SA-GRU classification model. The parameters are set as follows: the input dimension input_size is 35, the output dimension output_size is 1, the hidden dimension hidden_size is 35, the number of data samples captured for one training batch_size is 30, the number of SA-GRU layers is 2, and training is performed in the GRU under the above structure;

[0028] GRU includes a reset gate, an update gate, a candidate hidden state, and a hidden state;

[0029] The relevant parameter calculation and update process is as shown in equations (2)-(5):

[0030] R t =σ(I t W ir +H t-1 W hr +b r ) (2)

[0031] Zt =σ(I t W iz +H t-1 W hz +b z ) (3)

[0032]

[0033]

[0034] Among them, the number of hidden units is h, and the input at time step t is I t ∈R I×d , I×d represents the input vector dimension, and the hidden state at the previous time step t-1 is H t-1 ∈R I×h , W ir , W iz , W ih ∈R d×h and W hr , W hz , W hh ∈R h×h represents the weight parameter matrix, b r , b z , b h ∈R I×h is the bias parameter, σ represents the sigmoid activation function, ⊙ represents element-wise multiplication, R t , Z t , H t Represent the outputs of the reset gate, update gate, candidate hidden state, and hidden state respectively;

[0035] Step 4.2: The 34 features selected in step 3 are processed by bidirectional GRU as the input of the fully connected layer. The fully connected layer uses ReLu as the activation function and adds a dropout mechanism. During the training process, in order to prevent overfitting, the neural network units in the fully connected layer are temporarily discarded from the network with a probability of 0.5 to prevent the model from over-learning certain features. Specifically, as shown in formula (6):

[0036]

[0037] w j is regarded as the weight of the j-th feature, that is, the influence of the importance of each dimension feature on the final score, b j is the bias term, and the score z of each category is obtained by weighted summing of the features. j ; z j The value has a probability of 0.5 being discarded and will not be transmitted further;

[0038] Step 4.3: Finally, the probability distribution of different brake handle level labels is obtained through the Softmax classifier. Specifically as formula (7):

[0039]

[0040] Among them, x is the input of the fully connected layer, W n×K is the weight, b is the bias term, and the calculation method of Softmax is as shown in formula (8) and formula (9):

[0041]

[0042] The probability of each label category is as follows:

[0043]

[0044] The most likely The label category indicated is the braking operation to be performed next.

[0045] Step 5 is as follows:

[0046] First, a route is selected for testing. The selected route requires that the original brake handle level contains eleven different categories in the initial operation data record. Then, the traction power consumption of the original operation data is calculated. The specific calculation process is as follows:

[0047] The energy consumption of the selected route is j1 at the starting position and j2 at the end position, then the total energy consumption on this route is J, as shown in formula (10):

[0048] J=j2-j1 (10)

[0049] Among them, except for the traction handle, the same brake handle level can be divided into three operating conditions: acceleration, uniform speed, and deceleration; according to the changes in speed and acceleration, the number of acceleration, uniform speed, and deceleration of the same brake handle is counted, and the percentage of each operating condition is calculated based on the statistical number. The percentages of acceleration, uniform speed, and deceleration are recorded as p i，0 , p i，1 , p i，2 , i represents the brake handle level label, and its value is 0, 1, 2, ..., 10.

[0050] Step 6 is as follows:

[0051] The selected route is used as a test set for prediction. The test set needs to go through multiple feature selections and then input into the trained SA-GRU model to obtain the predicted braking strategy and then calculate the energy consumption. The specific process is as follows:

[0052] Step 6.1, the test route data is subjected to the variance method and mutual information method to obtain thirty-seven features, and then the feature selection is performed by XGBOOST to obtain thirty-four features. The data containing the thirty-four features is input into the trained SA-GRU model to obtain the predicted braking strategy;

[0053] Step 6.2: Count the number of each type of brake handle level, using b0, b1, ..., b 10 The number of different operating conditions of each type of brake handle level is calculated, and the number of brake handle levels that consume electrical energy among the 11 types of brake handle levels is obtained, as shown in formula (11):

[0054] L=(p 0，0 +p 0，1 )×b0+b1+(p 2，0 +p 2，1 )×b2+…+(p 10，0 +p 10，1 )×b 10 (11)

[0055] Since the handle level is 1, it is traction, and only consumes electric energy to accelerate, so there is no need to distinguish whether b1 is divided into acceleration, constant speed, and deceleration;

[0056] Step 6.3, calculate the energy consumption of each type of brake handle level obtained in step 6.2. First, use Excel software to calculate the instantaneous energy consumption of each label in the initial running data, then add the instantaneous energy consumption of each brake handle level, divide the sum by the number of each brake handle level, and get the average energy consumption increase of each brake handle level. The average energy consumption increase of each brake handle level is expressed as: q0, q1, ..., q 10 ; Then the energy consumption required by the model strategy predicted by the final SA-GRU classification model can be expressed as formula (12):

[0057] Q=q0(p 0，0 +p 0，1 )×b0+q1b1+q0(p 2，0 +p 2，1 )×b2+…+q 10 (p 10，0 +p 10，1 )×b 10 (12).

[0058] Step 7 is as follows:

[0059] Step 7.1: From step 5 and step 6, we can get the power consumption of the test set in the initial runtime data, that is, J. The energy consumption under the operation strategy B formulated by the SA-GRU classification model is Q. The relationship between the two is as follows:

[0060]

[0061] Step 7.2, if the energy consumption relationship is J>Q, the strategy is determined to be the initial energy-saving strategy B; in the actual operation process, the train needs to be accelerated to a speed close to the speed limit, and then the initial energy-saving strategy B is executed. If the train speed has not reached the terminal station when it becomes 0, the uniform speed running time of the train is increased so that the high-speed train is exactly equal to the test route length when the speed becomes 0. The purpose of doing this is to meet the punctuality of the train in actual conditions;

[0062] Step 7.3: If the energy consumption relationship is J≤Q, adjust the number of training times, batch_size and the learning rate using the simulated annealing algorithm to achieve an energy consumption relationship of J>Q, thereby obtaining an energy-saving strategy; the simulated annealing algorithm trains the model with a trend of reducing the learning rate by 0.1 every 50 training times.

[0063] XGBoost in the present invention represents an extreme gradient boosting algorithm.

[0064] The beneficial effects of the present invention are:

[0065] 1: The present invention transforms the energy-saving problem of high-speed trains into a classification operation of the brake handle levels, that is, transforms the numerical calculation problem into a multi-classification problem study.

[0066] 2: The present invention applies the SA-GRU model to the prediction of the energy-saving driving strategy of high-speed trains. Since GRU has one less gate function than LSTM, the number of parameters is less than LSTM, so the training speed of GRU is faster. When there is not much data, GRU performs better than LSTM and can obtain the best energy-saving strategy more quickly.

[0067] 3: The present invention adds the idea of ​​simulated annealing algorithm to the GRU model to form a GRU variant SA-GRU, and uses Adadelta and Adam optimizers to optimize the parameters of the model. The idea of ​​simulated annealing algorithm is used to change the learning rate in the two optimizers with the number of training times to pursue the optimal energy-saving driving strategy. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 It is the overall flow chart of the high-speed train energy-saving driving strategy based on SA-GRU of the present invention;

[0069] Figure 2It is a data processing flow chart of the high-speed train energy-saving driving strategy based on SA-GRU of the present invention;

[0070] Figure 3 It is a schematic diagram of the classification model of the high-speed train energy-saving driving strategy based on SA-GRU of the present invention;

[0071] Specific implementation process

[0072] The SA-GRU-based high-speed train energy-saving driving strategy of the present invention is as follows: Figure 1 As shown in the figure, the specific method is to first use the interpolation method, chi-square test and other methods to preprocess the data generated during the operation of the train from the starting station to the terminal station, such as the running time, speed, car electrical cabinet temperature, transformer temperature, drive side shaft temperature, gear temperature, shaft temperature, speed, duct temperature and brake handle level and other attributes to obtain the train running data set; then use the undersampling method to construct a balanced data set using the train running data set; then use multiple feature selection to select features of the constructed balanced data set to obtain features with a high correlation with the brake handle level, that is, the final SA-GRU model input features, including speed, acceleration, and distance to the station; input the model input features into the SA-GRU model for training, and finally use the trained SA-GRU model to predict the brake handle level of the test route data to obtain the predicted handle level, that is, the operation strategy at run time; calculate the energy consumption according to the predicted strategy, and compare it with the initial energy consumption in the test set route to see whether it is energy-saving. If the energy consumption is reduced, an energy-saving strategy is obtained, otherwise, the model is adjusted to predict again; combined with Figure 2-3 The details are as follows:

[0073] Step 1 is to process the initial data recorded from the starting station to the terminal station, including removing irrelevant attributes, chi-square test and interpolation filling. The specific operations are as follows:

[0074] Step 1.1: First, remove irrelevant attributes from the running data. Since the original data records all the attributes of the high-speed train, it contains many useless attributes, such as train number, carriage number, etc. The useless attributes will not change before and after the high-speed train runs, and have little significance for the calculation of train running energy consumption, so they need to be removed;

[0075] Step 1.2, after removing irrelevant attributes, perform a chi-square test. The chi-square test is a widely used hypothesis testing method. Statistical inference is performed on categorical data based on the relationship between two qualitative variables, which refers to the relationship between each attribute and the brake handle level. Use a cross-tabulation table in the SPSS software to perform statistics on the data after step 1.1, and obtain the chi-square value of each attribute relative to the brake handle level. The 47 characteristic attributes with high chi-square values ​​are used as characteristic data for subsequent use;

[0076] Step 1.3: Since there are some missing values ​​in the data, the data mainly includes missing records, which will have a great impact on data analysis and make the uncertainty of the results more significant, so the interpolation method is selected to fill the missing data. In the present invention, the mean interpolation method is selected to fill the missing data, and the mean of the same attribute value at the upper and lower time points of the missing data is calculated, and the obtained value is the missing value; as shown in formula (1):

[0077]

[0078] where d i It means that the vacancies in the i-th row of the d-th attribute of the high-speed train need to be filled, which is represented by the average of the two time series above and below, that is, d i-1 , d i+1 .

[0079] Step 2 is to perform undersampling on the data after preprocessing it, to obtain a data set with balanced brake handle levels, to prevent the problem of increased classification error due to data imbalance. The specific operations are as follows:

[0080] Step 2.1, classify the runtime data according to the brake handle level. The value of the brake handle level is several discrete quantities: {EB, REL, 1A, 1B, 2, 3, 4, 5, 6, 7, 8, OC}, EB is the emergency brake position, represented by label 0, REL is the operating position, that is, the traction position, represented by label 1, the commonly used brake positions are: 1A, 1B, 2, 3, 4, 5, 6, 7, 8, which are represented by 2, 3, 4, 5, 6, 7, 8, 9, 10 respectively in the present invention, OC is to ignore the emergency brake position activated by the passenger, represented by label 11, thereby mapping the brake handle level to: {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11}, thereby dividing the data into twelve categories. Since the present invention is related to energy consumption, the data involved here is in the operating state;

[0081] Step 2.2, each category of data is unbalanced. Unbalanced data sets will lead to large errors in classifier training, so undersampling is used to take the least data as the standard, and other data are undersampled. Since the data with brake handle level 11 represents the passenger activating the emergency brake position, this category is eliminated because the amount of such data is too small and does not contribute to model training. During the data preprocessing process, 11 data with brake level 0 were eliminated, and finally 60 data were used as the minority data set. The undersampling method was used to randomly extract samples from the majority of the brake handle levels from 1 to 10. 60 data were extracted for each brake handle level, and finally a balanced data set with 660 data in the training set was obtained.

[0082] Step 3 is a process of multiple feature selection. This study uses three feature selection methods to select features of the balanced data set obtained in step 2. The three feature selection methods are variance method, mutual information method and XGBOOST feature selection method. The data obtained after the feature selection method will be used to train the SA-GRU model. The specific operations are as follows:

[0083] Step 3.1, feature selection based on variance, is essentially to filter the class of features through the variance of the feature itself. For example, if the variance of a feature itself is very small, it means that there is basically no difference in the samples on this feature. It is possible that most of the values ​​in the feature are the same, or even the values ​​of the entire feature are the same, then this feature has no effect on sample differentiation. So no matter what the next feature engineering is to do, it is necessary to give priority to eliminating features with a variance of 0 or too small a variance. In the training set obtained in step 2, there are 47 features. Feature selection is performed from the 47 features, leaving K (1≤K≤47) features with large variances, to achieve the purpose of reducing the number of features and dimensionality reduction, making the model more generalizable and reducing overfitting. Calculate the f-value of the linear regression test of a single feature and the brake handle level label, adjust the number of features to be selected, and perform feature selection from high to low according to the scoring criteria. In the present invention, the value of K is selected as 35;

[0084] Step 3.2, the mutual information method is to calculate the two events of whether a certain feature appears and whether the classification is correct when performing single feature selection, and retain the features with higher scores. Feature selection based on mutual information uses discrete feature sets or continuous feature sets to calculate the mutual information value (also the information gain value) of each feature set and the brake handle level label, and after arranging them in descending order according to the mutual information value, K (1≤K≤47) features with large information gain values ​​are retained to complete the feature selection. The K value in the present invention is selected as 35;

[0085] Step 3.3, merge the features selected by the two feature selection methods in step 3.1 and step 3.2, and obtain 38 feature attributes after merging the two;

[0086] Step 3.4: Encode the 38 feature attributes obtained in step 3.3. Each code corresponds to a feature attribute. The feature attribute coding range is 1-38. Code 0 is the brake handle level label. Use XGBOOST to perform feature selection on the data set, and rank the relative importance of each feature attribute and the brake handle level. Finally, 34 features including speed, acceleration, distance to station, ambient temperature, gear temperature, traction power consumption, etc. are selected as the feature attributes of the final training set.

[0087] Step 4: The data obtained in step 3 is used as a training set to train the SA-GRU classification model of the present invention. This classification model consists of three layers: the first layer is the gru layer, which uses a gating mechanism to control input, memory and other information and makes predictions at the current time step. The second layer is the 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 (0, 1) real numbers (probability), while ensuring that their sum is 1;

[0088] Step 4.1: Use the final data obtained in step 3 as the input data of the SA-GRU classification model. The parameters are set as follows: the input dimension input_size is 35, the output dimension output_size is 1, the hidden layer dimension hidden_size is 35, the batch_size of the number of data samples captured for one training is 30, the number of SA-GRU layers is 2, and training is performed in the GRU under the above structure.

[0089] GRU consists of a reset gate, an update gate, candidate hidden states, and hidden states.

[0090] The relevant parameter calculation and update process is as shown in equations (2)-(5):

[0091] R t =σ(I t W ir +H t -1W hr +b r ) (2)

[0092] Z t =σ(I t W iz +H t -1W hz +b z ) (3)

[0093]

[0094]

[0095] Assume that the number of hidden units is h and the input at time step t is I t ∈R I×d (I×d represents the input vector dimension), the hidden state at the previous time step t-1 is H t-1 ∈R I×h , W ir , W iz , W ih ∈R d×h and W hr, W hz , W hh ∈R h×h represents the weight parameter matrix, b r , b z , b h ∈R I×h is the bias parameter, σ represents the sigmoid activation function, ⊙ represents element-wise multiplication, R t , Z t , H t Respectively represent the reset gate, update gate, candidate hidden state and hidden state output. The reset gate controls how the hidden state of the previous time step flows into the candidate hidden state of the current time step, which helps to capture short-term dependencies in the text sequence. The update gate controls how the hidden state should be updated by the candidate hidden state containing the current time step information, which helps to capture short-term handle level relationships during the train's travel. The present invention uses a variant of GRU, the bidirectional GRU. The output of the hidden state of each time step of the bidirectional GRU is spliced ​​by the forward hidden state and the backward hidden state of the current time step. Therefore, it is jointly affected by the information transmission in the forward and backward directions and contains richer information.

[0096] Step 4.2: The 34 features selected in step 3 are processed by bidirectional GRU as the input of the fully connected layer, using x1, x2, ..., x n In order to calculate faster, the fully connected layer uses ReLu as the activation function and adds a dropout mechanism. During the training process, the neural network unit is temporarily discarded from the network with a probability of 0.5, so that it does not work. The purpose of this is to prevent the trained model from overfitting and improve the generalization ability of the model. Specifically, as shown in formula (6):

[0097]

[0098] w j is regarded as the weight of the j-th feature, that is, the influence of the importance of each dimension feature on the final score, b j is the bias term, and the score z of each category is obtained by weighted summing of the features. j ;

[0099] Step 4.3: Finally, the classification result is obtained through the Softmax classifier, that is, the probability distribution of the classification labels of different brake handle levels, as shown in formula (7):

[0100]

[0101] Among them, x is the input of the fully connected layer, W n×K is the weight, b is the bias term, is the output probability of Softmax. The calculation method of Softmax is as shown in formula (8) and formula (9):

[0102]

[0103] If the probability of each label category is split into the following:

[0104]

[0105] The most likely The label category indicated is the braking operation to be performed next.

[0106] Step 5 requires first selecting a route for testing. The selected route needs to include eleven different types of brake handle levels, and then calculate the traction power consumption of the original running data. The specific calculation process is as follows:

[0107] Assume that the energy consumption of the selected route is j1 at the starting position and j2 at the end position. Assume that the total energy consumption on this route is J, and the expression is as follows:

[0108] J=j2-j1 (10)

[0109] Among them, except for the traction handle, the same brake handle level can be divided into three operating conditions: acceleration, constant speed, and deceleration. In the acceleration stage, excessive traction needs to be applied to make the high-speed train reach a certain speed. At this time, the traction power consumption is the highest. In the constant speed stage, the high-speed train needs to overcome air resistance and body friction to keep the speed unchanged. At this time, the traction power consumption is second. No power is consumed in the braking stage, and the traction power consumption is 0 at this time.

[0110] According to the changes in speed and acceleration, the number of acceleration, uniform speed and deceleration of the same brake handle is counted, and the percentage of each operation condition is calculated based on the statistical number. The percentages of acceleration, uniform speed and deceleration are recorded as p respectively. i，0 , p i，1 , p i，2 , i represents the brake handle level label, and its value is 0, 1, 2, ..., 10.

[0111] Step 6 uses the selected route as a test set for prediction. The test set needs to go through multiple feature selections and then input into the trained SA-GRU model to obtain the predicted braking strategy and then calculate the energy consumption. The specific process is as follows:

[0112] Step 6.1, the test route data is subjected to the variance method and mutual information method to obtain thirty-seven features, and then the feature selection is performed by XGBOOST to obtain thirty-four features. The data containing the thirty-four features is input into the trained SA-GRU model to obtain the predicted braking strategy;

[0113] Step 6.2: Count the number of each type of brake handle level, using b0, b1, ..., b 10 The number of different operating conditions of each type of brake handle level is calculated, and the number of brake handle levels that consume electrical energy among the 11 types of brake handle levels is obtained, as shown in formula (11):

[0114] L=(p 0，0 +p 0，1 )×b0+b1+(p 2，0 +p 2，1 )×b2+…+(p 10，0 +p 10，1 )×b 10 (11)

[0115] Since the handle level is 1, it is traction, and only consumes electric energy to accelerate, so there is no need to distinguish whether b1 is divided into acceleration, constant speed, and deceleration;

[0116] Step 6.3, calculate the energy consumption of each type of brake handle level in different operating conditions obtained in step 6.2. First, use Excel software to calculate the instantaneous energy consumption of each label in the initial operation data, then add the instantaneous energy consumption of each brake handle level, divide the sum by the number of each brake handle level, and get the average energy consumption increase value of each brake handle level. The average energy consumption increase value of each brake handle level is expressed as: q0, q1, ..., q 10 , then the energy consumption required by the model strategy predicted by the final classification model can be expressed as formula (12):

[0117] Q=q0(p 0，0 +p 0，1 )×b0+q1b1+q0(p 2，0 +p 2，1 )×b2+…+q 10 (p 10，0 +p 10，1 )×b 10 (12)

[0118] Step 7 is to compare and analyze the traction power consumption in step 5 and step 6. According to the feedback of the results, the network model is adjusted using the simulated annealing algorithm to modify it to the most suitable model and seek the best energy-saving strategy. The specific steps are as follows:

[0119] Step 7.1: From Step 5 and Step 6, we can get the power consumption of the test set in the initial runtime data, i.e., J, after the SA-GRU classification model

[0120] The energy consumption under the formulated operation strategy B is Q, and the relationship between the two is as follows:

[0121]

[0122] Step 7.2, if the energy consumption relationship is J>Q, the strategy is determined to be the initial energy-saving strategy B. In the actual operation process, the train needs to be accelerated to a speed close to the speed limit, and then the initial energy-saving strategy B is implemented. If the train speed becomes 0 but has not reached the terminal, the uniform speed running time of the train is increased so that the high-speed train is exactly equal to the test route length when the speed becomes 0. The purpose of doing this is to meet the punctuality of the train in actual situations;

[0123] Step 7.3, if the energy consumption relationship is J≤Q, the learning rate should be adjusted by adjusting the number of training times, batch_size and using the simulated annealing algorithm to achieve an energy consumption relationship of J>Q, thereby obtaining an energy-saving strategy. The simulated annealing algorithm is a random optimization algorithm, and its starting point is based on the similarity between the annealing process of solid materials in physics and the general combinatorial optimization problem. Starting from a relatively high initial temperature, with the continuous decrease of the temperature parameter, the simulated annealing algorithm combines the probability jump characteristics to randomly search for the global optimal solution of the objective function in the solution space, that is, the local optimal solution can probabilistically jump out and eventually tend to the global optimum. The present invention refers to the use of a simulated annealing algorithm to train the model with a trend of reducing the learning rate by 0.1 every 50 training times.

Claims

1. A high-speed train energy-saving driving strategy method based on SA-GRU, characterized in that: The specific steps are as follows: Step 1: For the running data generated during the running of the train from the starting station to the terminal station, the interpolation method and the chi-square test method are used to pre-process the running time, speed, carriage electrical cabinet temperature, transformer temperature, drive side shaft temperature, gear temperature, shaft temperature, speed, duct temperature and brake handle level of a series of individual features; Step 2: Divide the data into 12 categories according to 12 different categories of brake handle levels, and map each category of brake handle levels to numbers from 0 to 11 as labels, remove the data of the emergency brake handle level with too few numbers, that is, the data with label 11, and for the remaining train running data, the labels are 0 to 10, and the undersampling method is used to construct a balanced data set operation to obtain a balanced data set with the same number of categories; Step 3, multiple feature selection is performed on the balanced data set in step 2 to remove features that have no effect or little effect on the brake handle level operation. After multiple selection, the data in the balanced data set obtained only contains thirty-four features with a large correlation with the brake handle level. This balanced data set will be used as a training set; Step 4: Use the data except the brake handle level in the training set after multiple feature selection in step 3 as the input of the SA-GRU classification model, and use the brake handle level as the output to train the SA-GRU classification model; Step 4 is as follows: The balanced data set obtained in step 3 is used as a training set to input the SA-GRU classification model for training. The SA-GRU classification model consists of three layers: the first layer is the gru layer, which uses a gating mechanism to control input, memorize information, and make predictions at the current time step; The second layer is the fully connected layer. This layer connects n Real numbers are mapped into K real number; The third layer is the softmax layer, which takes K The real number is mapped to K (0,1) real numbers, while ensuring that their sum is 1; Step 4.1: Use the final data obtained in step 3 as the input data of the SA-GRU classification model. The parameters are set as follows: the input dimension input_size is 35, the output dimension output_size is 1, the hidden dimension hidden_size is 35, the number of data samples captured for one training batch_size is 30, the number of SA-GRU layers is 2, and training is performed in the GRU under the above parameter settings; The GRU includes a reset gate, an update gate, a candidate hidden state, and a hidden state; The relevant parameter calculation and update process is as shown in equations (2)-(5): (2) (3) (4) (5) Among them, the number of hidden units is , time step The input is , Represents the input vector dimension, the previous time step The hidden state of , , , and , , represents the weight parameter matrix, , , is the bias parameter, represents the sigmoid activation function, means to do element-wise multiplication, , , , Represent the outputs of the reset gate, update gate, candidate hidden state, and hidden state respectively; Step 4.2: The 34 features selected in step 3 are processed by bidirectional GRU as the input of the fully connected layer. Indicates that the fully connected layer uses ReLu as the activation function and adds a dropout mechanism. During the training process, in order to prevent overfitting, the neural network units in the fully connected layer are temporarily discarded from the network with a probability of 0.5, making them inoperative and preventing the model from over-learning certain features. Specifically, as shown in formula (6): (6) Will Considered as The weight of the class feature, that is, the influence of the importance of each dimension feature on the final score, is the bias term, and the score of each category is obtained by weighted summing of the features ; The value has a probability of 0.5 being discarded and will not be transmitted further; Step 4.3, finally pass The classifier obtains the probability distribution of different brake handle level labels , specifically as formula (7): (7) in, is the input of the fully connected layer, is the weight, is the bias term, The calculation method of is as follows: (8) The probability of each label category is as follows: (9) The most likely The label category indicated is the braking operation to be performed next; Step 5: Select a section of route as the test set from all the data preprocessed in step 1 according to the speed change, and calculate the difference of the traction power consumption from the starting station to the terminal station recorded in the data set to obtain the traction power consumption under the original braking strategy in this section of the route. , and record them; then calculate the percentage of acceleration, uniform speed, and deceleration under each label according to the operating conditions in the initial runtime data, and record them as ; Step 6: After multiple feature selection, the test set is input into the trained SA-GRU model to obtain the initial braking strategy. Since energy is consumed only in the acceleration and constant speed stages, and no energy is consumed in the deceleration stage, the strategy obtained by the percentage in step 5 and the SA-GRU model is The number of each brake handle level label can be used to obtain the number of energy consumption labels L; finally, the initial braking strategy is obtained based on the prediction The energy consumption increase per second brought by each brake handle level calculated from the initial running data is calculated in the strategy Traction power consumption ; Step 7: Compare and analyze the traction power consumption in step 5 and step 6, adjust the network model, and seek the best energy-saving strategy; Step 7.1: From Step 5 and Step 6, the power consumption of the test set in the initial runtime data can be obtained, that is, , the operation strategy developed by the SA-GRU classification model The energy consumption is There are two types of relationships between them: Step 7.2: If the energy consumption relationship is , then the strategy is determined as the initial energy-saving strategy ; In actual operation, the train needs to be accelerated to a speed close to the speed limit before executing the initial energy-saving strategy If the train has not reached the terminal station when the speed becomes 0, the uniform speed running time of the train will be increased so that the high-speed train will be exactly equal to the length of the test route when the speed becomes 0. This is done to ensure the punctuality of the train in actual situations. Step 7.3: If the energy consumption relationship is , we need to adjust the number of trainings, batch_size and the learning rate using the simulated annealing algorithm to achieve the energy consumption relationship of , thereby obtaining an energy-saving strategy; the simulated annealing algorithm trains the model with a learning rate that decreases by 0.1 every 50 training times.

2. The high-speed train energy-saving driving strategy method based on SA-GRU according to claim 1 is characterized in that: Step 1 is as follows: Step 1.1: Remove the attributes that will not change during the operation process, that is, the data generated during the operation of the train from the starting station to the terminal station, that is, the data that is irrelevant to the train driving energy consumption, which is not only invalid for the classification model training, but also interferes with the model training; Step 1.2: Perform a chi-square test based on the relationship between each attribute and the brake handle level. The chi-square test value represents the degree of correlation between each attribute and the brake handle level. The attribute with the highest chi-square value is selected. Features retained; Step 1.3: Perform interpolation to fill missing values ​​in the acquired running data. Use the values ​​of the two timestamps above and below the missing value to calculate the average. The obtained value is used to fill in the missing value. At this time, the final complete data containing 47 features is obtained.

3. The high-speed train energy-saving driving strategy method based on SA-GRU according to claim 2 is characterized in that: Step 2 is as follows: Step 2.1: Classify the runtime data according to the brake handle level. The value of the brake handle level is several discrete quantities: , EB is the emergency brake position, represented by label 0; REL is the operating position, i.e. the traction position, represented by label 1; the common brake positions 1A, 1B, 2, 3, 4, 5, 6, 7, 8 are represented by 2, 3, 4, 5, 6, 7, 8, 9, 10 respectively; OC is the emergency brake position for ignoring passenger activation, represented by label 11, thus mapping the brake handle level to: , a total of twelve categories; Step 2.2: The data with brake handle level 11 represents the category of passenger-activated emergency brake position, and 60 groups of data are randomly selected from the data with brake handle level labels of 1-10 using the undersampling method, and a total of 660 data are obtained.

4. The high-speed train energy-saving driving strategy method based on SA-GRU according to claim 3 is characterized in that: Step 3 is as follows: Step 3 uses multiple feature selection to select data feature attributes, using traditional feature selection methods and XGBOOST for feature selection. The traditional feature selection method uses variance method and mutual information method, as follows: Step 3.1: Feature selection based on variance: Perform feature selection on the 47 features in step 1.3, leaving K features with large variance, where 1≤K≤47; calculate the f-value of the linear regression test of a single feature and the brake handle level label, adjust the number of features to be selected, and select features from high to low according to the scoring criteria, with a K value of 35; Step 3.2: Feature selection based on mutual information: After arranging the mutual information values ​​of the single features and the brake handle levels in step 1 in descending order, select K features to continue the operation, 1≤K≤47, and the K value is 35; Step 3.3: Merge the features selected by the two feature selection methods in step 3.1 and step 3.2 to obtain K1 feature attributes; Step 3.4: Encode the K1 feature attributes obtained in step 3.

3. Each code corresponds to a feature attribute. The feature attribute coding range is 1-K1. Code 0 is the brake handle level label. Use XGBOOST to perform feature selection on the data set, and rank the relative importance of each feature attribute and the brake handle level. Finally, select 34 features with high importance as the feature attributes of the final training set.

5. The high-speed train energy-saving driving strategy method based on SA-GRU according to claim 1 is characterized in that: The step 5 is specifically as follows: First, a route is selected for testing. The selected route requires that the original brake handle level contains eleven different categories in the initial operation data record. Then, the traction power consumption of the original operation data is calculated. The specific calculation process is as follows: The energy consumption of the selected route at the starting position is , the value at the end point is , then the total energy consumption on this route is , the expression is as follows: (10) Among them, except for the traction handle, the same brake handle level can be divided into three operating conditions: acceleration, uniform speed, and deceleration; according to the changes in speed and acceleration, the number of acceleration, uniform speed, and deceleration of the same brake handle is counted, and the percentage of each operating condition is calculated based on the statistical number. The percentages of acceleration, uniform speed, and deceleration are recorded as , Represents the brake handle level label, with values ​​of 0, 1, 2, ..., 10.

6. The high-speed train energy-saving driving strategy method based on SA-GRU according to claim 1 is characterized in that: Step 6 is as follows: The selected route is used as a test set for prediction. The test set needs to go through multiple feature selections and then input into the trained SA-GRU model to obtain the predicted braking strategy and then calculate the energy consumption. The specific process is as follows: Step 6.1, the test route data is subjected to the variance method and mutual information method to obtain thirty-seven features, and then the feature selection is performed by XGBOOST to obtain thirty-four features. The data containing the thirty-four features is input into the trained SA-GRU model to obtain the predicted braking strategy; Step 6.2: Count the number of each type of brake handle level, using , , …, The number of different operating conditions of each type of brake handle level is calculated, and the number of brake handle levels that consume electrical energy among the 11 types of brake handle levels is obtained. , as shown in formula (11): (11) Since the handle level is 1, it is traction, and only consumes energy to accelerate, so there is no need to distinguish Whether the situation is divided into acceleration, constant speed and deceleration; Step 6.3: Calculate the energy consumption of each type of brake handle level obtained in step 6.2 for the number of different operating conditions. First, calculate the instantaneous energy consumption of each label in the initial running data with the help of Excel software. Then add the instantaneous energy consumption under each brake handle level, divide the sum by the number of each brake handle level, and obtain the average energy consumption increase value of each brake handle level. The average energy consumption increase value of each brake handle level is expressed as: , , …, ; Then the energy consumption required by the model strategy predicted by the final SA-GRU classification model can be expressed as formula (12): (12)。

Images