Fuel consumption prediction method based on automobile driving behavior classification
By building a driving style classification and recognition model and a personalized fuel prediction model, the problem of the existing technology lacking consideration of driving behavior differences is solved, and accurate prediction of fuel consumption and personalized guidance are achieved.
Patent Information
- Application Number
- CN202511077824.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-01
- Publication Date
- 2025-09-16
AI Technical Summary
Existing fuel prediction methods lack consideration of individual differences in driving behavior, making it difficult to accurately predict individualized fuel consumption.
By obtaining vehicle operation data and preprocessing it, a driving style classification and recognition model is constructed. Driving style types are divided using principal component analysis and K-means clustering. A personalized fuel prediction model is constructed by combining a feedforward neural network and a Bayesian regularized backpropagation algorithm.
It achieves accurate prediction of fuel consumption based on driving behavior, provides guidance for subsequent fuel economy and driver operating habits, and improves the accuracy and personalization of predictions.
Smart Images
Figure CN120645972A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of intelligent driving assistance, and in particular to a fuel consumption prediction method based on automobile driving behavior classification. Background Art
[0002] In the development of vehicle energy-saving management and intelligent driving assistance systems, fuel consumption prediction is of great significance for optimizing driving strategies and formulating energy-saving policies. At present, mainstream fuel prediction methods are mostly based on real-time vehicle operating data, such as speed, acceleration, and rotational speed. By establishing regression models or machine learning models, the fuel consumption level per unit time or unit distance is predicted. However, the existing technologies generally have the following shortcomings: lack of consideration for individual differences in driving behavior. Different drivers may have significant differences in fuel consumption performance due to different operating styles in similar traffic environments and vehicle conditions. Traditional models often regard all driving behaviors as homogeneous, making it difficult to accurately predict individualized fuel consumption. Summary of the Invention
[0003] The purpose of the present invention is to provide a fuel consumption prediction method based on automobile driving behavior classification, comprising the following steps:
[0004] 1) Obtain the original speed time series and perform preprocessing to obtain the preprocessed speed time series;
[0005] 2) Construct a driving style classification and recognition model;
[0006] 3) Extracting features from the preprocessed speed time series to obtain feature parameters, which are then input into a driving style classification and recognition model to determine the current driving style type;
[0007] 4) Build a fuel prediction model under different driving styles;
[0008] 5) Obtain vehicle operating data and input it into the fuel prediction model corresponding to the current driving style type to obtain the instantaneous fuel consumption result.
[0009] Further, in step 1), the step of pre-processing includes:
[0010] Delete the data segments whose missing time length is greater than or equal to the preset value, and use the Lagrange interpolation method to fill the data segments whose missing time length is less than the preset value;
[0011] Eliminate data segments whose idle time is greater than or equal to the preset value;
[0012] Replace speed outliers that exceed the speed limit with the speed of the previous second.
[0013] Furthermore, the data completed by Lagrange interpolation is as follows:
[0014]
[0015] Where: P n (x) is the interpolation polynomial, y i is the dependent variable value of the i-th known point, L i (x) is the i-th Lagrange basis function; x is the independent variable of the point to be interpolated; x i 、x j , is the independent variable of the i / jth known point.
[0016] Furthermore, in step 2), the step of constructing a driving style classification and recognition model includes:
[0017] 2.1) Acquire multiple original velocity time series and perform preprocessing to obtain multiple preprocessed velocity time series;
[0018] 2.2) Perform feature extraction on each preprocessed velocity time series to obtain a feature parameter matrix;
[0019] 2.2) Use principal component analysis to reduce the dimension of the feature parameter matrix to obtain the characteristic principal components and the corresponding principal component scores;
[0020] 2.3) K-means clustering is used to cluster the principal component score matrix, thereby classifying the driving style types; the driving style types include aggressive, normal, and calm.
[0021] 2.4) Based on the feature parameters and the corresponding driving style types, the random forest model is trained to obtain a driving style classification and recognition model.
[0022] Furthermore, in step 2.2), the step of reducing the dimension of the feature parameter set using principal component analysis includes:
[0023] 2.2.1) Perform region center normalization on the characteristic parameter matrix A to obtain the normalized matrix Z m×n ;
[0024] Among them, the characteristic parameter matrix A is as follows:
[0025]
[0026] Where: A m×n is a matrix composed of different characteristic parameters of each fragment; a mn is the characteristic parameter;
[0027] Normalized matrix Z m×n As shown below:
[0028]
[0029] Where z ij is the standardized characteristic parameter; a ij is the characteristic parameter; is the average value of the characteristic parameter; s j is the standard deviation;
[0030] 2.2.2) Calculate the normalized matrix Z m×n The covariance matrix C is:
[0031]
[0032] Where cov(n,n) is the covariance;
[0033] 2.2.3) Calculate the correlation coefficient matrix R, that is:
[0034]
[0035] Where: r mn is the correlation coefficient; j, k = 1, 2, ..., n;
[0036] 2.2.2) Calculate the characteristic parameter λ of the correlation coefficient matrix R i , and calculate the contribution rate The top k characteristic parameters with the highest contribution rate are selected as principal components, and the contribution rate is used as the principal component score.
[0037] Furthermore, the characteristic parameters include speed mean, speed standard deviation, acceleration absolute value mean, acceleration absolute value standard deviation, Jerk, Jerk standard deviation, throttle opening position mean, throttle opening position standard deviation, throttle opening position average change rate, and throttle opening position change rate standard deviation.
[0038] Furthermore, in step 4), the step of constructing a fuel prediction model includes:
[0039] 4.1) Obtain multiple sets of vehicle operating data under the same driving style type and the corresponding instantaneous fuel consumption;
[0040] 4.2) Using normalization method to standardize the retained vehicle operation data;
[0041] 4.3) Quantify the correlation between the standardized vehicle operation data and instantaneous fuel consumption to obtain a gray correlation index, and retain vehicle operation data with a correlation index greater than a threshold;
[0042] 4.4) Constructing a training set and a test set based on the vehicle operation data and instantaneous fuel consumption retained in step 4.3);
[0043] 4.5) Train and test the feedforward neural network using the training and test sets to obtain a fuel prediction model for the current driving style.
[0044] 4.6) Change the driving style type and return to step 4.1) until a fuel prediction model corresponding to each driving style type is generated.
[0045] Furthermore, in step 4.1), the vehicle operation data includes GPS vehicle speed, fan speed, oil temperature, water temperature, oil pressure, intake pressure, relative intake pressure, actual torque percentage, engine speed, friction torque percentage, and gearbox output shaft speed.
[0046] Furthermore, during the training process of the feedforward neural network, the parameters of the feedforward neural network are updated using the Bayesian regularized back-propagation algorithm.
[0047] Furthermore, in step 4.3), the grey relational degree is as follows:
[0048]
[0049] Where, Δ i (k) is the difference between the reference sequence and the comparison sequence, ρ is the resolution coefficient; γ i is the grey relational degree; i (k) is the correlation coefficient of the i-th comparison sequence at the k-th moment; min i,k is the global minimum deviation; max i,k is the global maximum deviation.
[0050] The technical effect of the present invention is unquestionable. The present invention can judge the driver's driving behavior and predict fuel consumption based on the driving behavior, providing guidance for subsequent fuel economy, driver's operating habits, etc. BRIEF DESCRIPTION OF THE DRAWINGS
[0051] Figure 1 is the preprocessed data;
[0052] Figure 2 This is the classification model training result;
[0053] Figure 3 is the confusion matrix of the prediction results;
[0054] Figure 4 The results are the influencing factors;
[0055] Figure 5 Predict results for the model;
[0056] Figure 6 The results of model validation. DETAILED DESCRIPTION
[0057] The present invention will be further described below with reference to the following examples, but it should not be understood that the scope of the present invention is limited to the following examples. Without departing from the above technical ideas of the present invention, various substitutions and modifications can be made according to common technical knowledge and customary means in the art, and all should be included in the scope of protection of the present invention.
[0058] Example 1:
[0059] See also Figures 1 to 6 , a fuel consumption prediction method based on automobile driving behavior classification, comprising the following steps:
[0060] 1) Obtain the original speed time series and perform preprocessing to obtain the preprocessed speed time series;
[0061] 2) Construct a driving style classification and recognition model;
[0062] 3) Extracting features from the preprocessed speed time series to obtain feature parameters, which are then input into a driving style classification and recognition model to determine the current driving style type;
[0063] 4) Build a fuel prediction model under different driving styles;
[0064] 5) Obtain vehicle operating data and input it into the fuel prediction model corresponding to the current driving style type to obtain the instantaneous fuel consumption result.
[0065] Example 2:
[0066] A fuel consumption prediction method based on automobile driving behavior classification, the technical content of which is the same as that of Example 1, further, in step 1), the step of performing preprocessing includes:
[0067] Delete the data segments whose missing time length is greater than or equal to the preset value, and use the Lagrange interpolation method to fill the data segments whose missing time length is less than the preset value;
[0068] Eliminate data segments whose idle time is greater than or equal to the preset value;
[0069] Replace speed outliers that exceed the speed limit with the speed of the previous second.
[0070] Example 3:
[0071] A fuel consumption prediction method based on automobile driving behavior classification, the technical content of which is the same as any one of Examples 1-2, further, the data supplemented by the Lagrange interpolation method is as follows:
[0072]
[0073] Where: P n(x) is the interpolation polynomial, y i is the dependent variable value (speed) of the i-th known point, L i (x) is the i-th Lagrange basis function; x is the independent variable (time) of the point to be interpolated; x i 、x j , is the independent variable of the i / jth known point.
[0074] Example 4:
[0075] A fuel consumption prediction method based on automobile driving behavior classification, the technical content of which is the same as any one of Examples 1-3, further, in step 2), the step of constructing a driving style classification and recognition model includes:
[0076] 2.1) Acquire multiple original velocity time series and perform preprocessing to obtain multiple preprocessed velocity time series;
[0077] 2.2) Extract features from each preprocessed velocity time series to obtain a feature parameter matrix;
[0078] 2.2) Use principal component analysis to reduce the dimension of the feature parameter matrix to obtain the characteristic principal components and the corresponding principal component scores;
[0079] 2.3) K-means clustering is used to cluster the principal component score matrix, thereby classifying the driving style types; the driving style types include aggressive, normal, and calm.
[0080] 2.4) Based on the feature parameters and the corresponding driving style types, the random forest model is trained to obtain a driving style classification and recognition model.
[0081] Example 5:
[0082] A fuel consumption prediction method based on automobile driving behavior classification, the technical content of which is the same as any one of Examples 1-4, further, in step 2.2), the step of reducing the dimension of the feature parameter set using principal component analysis includes:
[0083] 2.2.1) Perform region center normalization on the characteristic parameter matrix A to obtain the normalized matrix Z m×n ;
[0084] Among them, the characteristic parameter matrix A is as follows:
[0085]
[0086] Where: A m×n is a matrix composed of different characteristic parameters of each fragment; a mn is the characteristic parameter;
[0087] Normalized matrix Zm×n As shown below:
[0088]
[0089] Where z ij is the standardized characteristic parameter; a ij is the characteristic parameter; is the average value of the characteristic parameter; s j is the standard deviation;
[0090] 2.2.2) Calculate the normalized matrix Z m×n The covariance matrix C is:
[0091]
[0092] Where cov(n,n) is the covariance;
[0093] 2.2.3) Calculate the correlation coefficient matrix R, that is:
[0094]
[0095] Where: r mn is the correlation coefficient; j, k = 1, 2, ..., n;
[0096] 2.2.2) Calculate the characteristic parameter λ of the correlation coefficient matrix R i , and calculate the contribution rate The top k characteristic parameters with the highest contribution rate are selected as principal components, and the contribution rate is used as the principal component score.
[0097] Example 6:
[0098] A fuel consumption prediction method based on automobile driving behavior classification, with technical content similar to any one of Examples 1-5, further comprising: characteristic parameters including speed mean, speed standard deviation, acceleration absolute value mean, acceleration absolute value standard deviation, Jerk, Jerk standard deviation, throttle opening position mean, throttle opening position standard deviation, throttle opening position average change rate, and throttle opening position change rate standard deviation.
[0099] Example 7:
[0100] A fuel consumption prediction method based on automobile driving behavior classification, the technical content of which is the same as any one of Examples 1-6, further, in step 4), the step of constructing a fuel prediction model includes:
[0101] 4.1) Obtain multiple sets of vehicle operating data under the same driving style type and the corresponding instantaneous fuel consumption;
[0102] 4.2) Using normalization method to standardize the retained vehicle operation data;
[0103] 4.3) Quantify the correlation between the standardized vehicle operation data and instantaneous fuel consumption to obtain a gray correlation index, and retain vehicle operation data with a correlation index greater than a threshold;
[0104] 4.4) Constructing a training set and a test set based on the vehicle operation data and instantaneous fuel consumption retained in step 4.3);
[0105] 4.5) Train and test the feedforward neural network using the training and test sets to obtain a fuel prediction model for the current driving style.
[0106] 4.6) Change the driving style type and return to step 4.1) until a fuel prediction model corresponding to each driving style type is generated.
[0107] Example 8:
[0108] A fuel consumption prediction method based on automobile driving behavior classification, the technical content of which is the same as any one of Examples 1-7, further, in step 4.1), the automobile operation data includes GPS vehicle speed, fan speed, oil temperature, water temperature, oil pressure, intake pressure, relative intake pressure, actual torque percentage, engine speed, friction torque percentage, and gearbox output shaft speed.
[0109] Example 9:
[0110] A fuel consumption prediction method based on automobile driving behavior classification, the technical content of which is the same as any one of Examples 1-8, further, during the feedforward neural network training process, the parameters of the feedforward neural network are updated using a Bayesian regularized backpropagation algorithm.
[0111] Example 10:
[0112] A fuel consumption prediction method based on automobile driving behavior classification, the technical content of which is the same as any one of Examples 1-9, further, in step 4.3), the gray correlation degree is as follows:
[0113]
[0114]
[0115] Where, Δ i (k) is the difference between the reference sequence and the comparison sequence, ρ is the resolution coefficient; γ i is the grey relational degree; i (k) is the correlation coefficient of the i-th comparison sequence at the k-th moment; min i,k is the global minimum deviation; max i,k is the global maximum deviation.
[0116] Example 11:
[0117] A fuel consumption prediction method based on vehicle driving behavior classification comprises the following steps:
[0118] 1. Establish a data processing model
[0119] During the data collection process, the loss of GPS signal will cause signal discontinuity, and long-term idling will cause abnormal collected data. The original data needs to be preprocessed to eliminate abnormal or invalid data and generate a new speed time series.
[0120] (1) Due to the obstruction of complex geographical environments such as high-rise buildings and tunnels, GPS signals may be lost, resulting in missing data during the data collection process. Shorter missing segments can be interpolated and supplemented. The Lagrange interpolation method is generally used for completion. The formula is as follows; larger missing data segments are deleted.
[0121]
[0122] Where: P n (x) is the interpolation polynomial, y i is the value of the i-th known point, L i (x) is the i-th Lagrangian basis function.
[0123] (2) Abnormal data collected during long-term idling. When a car idles for a long time, a lot of useless data will be generated, which will occupy computing resources and need to be eliminated.
[0124] (3) Clean up outliers in speed. The domestic speed limit is no more than 120 km / h. Speeds greater than 120 km / h are removed and replaced with the speed of the previous second.
[0125] 2. Establish a driving style classification and recognition model
[0126] When calculating characteristic parameters, it's necessary to determine the length of the driving style recognition cycle. Based on application requirements, the driving style recognition cycle was set to 8 seconds, and the raw data was divided into recognition segments, each of which was equal in length to the driving style recognition cycle. To accurately describe the state and characteristics of each kinematic segment, nine characteristic parameters were selected to characterize the vehicle driving condition characteristic evaluation system, as shown in Table 1.
[0127] Table 1 Characteristic parameters
[0128]
[0129]
[0130] Principal component analysis is a very common data dimensionality reduction method. Dimensionality reduction through principal component analysis can greatly improve computational efficiency and save computing resources. In the process of constructing vehicle driving conditions, the characteristic parameters mentioned in the above table have certain correlations. Principal component analysis can use fewer characteristic parameter dimensions to represent driving big data, which can significantly improve the computational efficiency of the characteristic parameter matrix in cluster analysis. The specific steps are as follows:
[0131] (1) Center standardization
[0132] The characteristic parameter value is represented by matrix A, that is,
[0133]
[0134] Where: A m×n It is a matrix composed of different characteristic parameters of each fragment.
[0135] For matrix A m×n The elements in the area center are normalized and substituted into the following formula
[0136]
[0137] This gives the normalized matrix Z m×n ,for
[0138]
[0139] (2) Calculate the covariance matrix
[0140] The covariance matrix C calculated from the standardized Z matrix is
[0141]
[0142] The correlation coefficient matrix R is further obtained as
[0143]
[0144] Where: r mn is the correlation coefficient.
[0145] (3) Calculate matrix eigenvalues
[0146] Calculate the characteristic parameter λ of the matrix R i ,i=1,2…n;λ i The order from large to small is λ1≥λ2≥λ3…≥λ n ,definition The contribution rate of the principal component is used to measure the information richness of the variables covered by the principal component. The greater the contribution rate, the more effective information is expressed. The sum of the contribution rates of the first r principal components is defined as When the cumulative contribution rate of the current r principal components reaches more than 85%, it is considered that they can better represent the original information.
[0147] K-means clustering was used to cluster the principal component score matrix, categorizing driving behaviors into three categories: aggressive, normal, and calm. The clustered data was then classified using the random forest algorithm to build a classification model, which was then validated.
[0148] 3. Establish a fuel prediction model
[0149] First, the data after driving style classification is extracted, and 11 characteristic parameter information is selected for predicting instantaneous fuel consumption. The characteristic parameter selection is shown in Table 2 below.
[0150] Table 2 Fuel consumption related characteristic parameters
[0151]
[0152]
[0153] Grayscale correlation analysis is used to quantitatively evaluate the correlation between each characteristic variable and the target variable (instantaneous fuel consumption). The instantaneous fuel consumption is set as the reference sequence X0, and the other characteristic variables are the comparison sequences X0. i , the correlation calculation method is as follows:
[0154] 1) Initialization sequence (standardization)
[0155] Normalize the original sequence:
[0156]
[0157] 2) Calculate the correlation coefficient:
[0158] Assume the correlation coefficient at the kth moment is:
[0159]
[0160] Where: Δ i (k) is the difference between the reference sequence and the comparison sequence, and p is the resolution coefficient.
[0161] 3) Calculate the gray relational degree:
[0162]
[0163] When γ i When it is >0.5, it is considered that the variable has a strong correlation with fuel consumption and can be used for subsequent model input.
[0164] Based on a multi-layer feedforward neural network, this model aims to effectively predict instantaneous fuel consumption using multi-dimensional feature data collected during vehicle operation. Key feature parameters that influence fuel consumption are used as input variables. Subsequently, normalization is used to standardize the input features to eliminate dimensional differences between features, improving the model's training stability and generalization capabilities. In terms of model structure design, a feedforward neural network consisting of multiple hidden layers is constructed to enhance the model's ability to model nonlinear relationships in the input data. During model training, a Bayesian regularized backpropagation algorithm is used to improve model robustness and effectively prevent overfitting. Furthermore, a validation mechanism is introduced to dynamically evaluate the training process, ensuring balanced performance across the training and validation sets. After model training, the model can be applied to actual test data for real-time prediction of instantaneous fuel consumption. To ensure the rationality of the prediction results, necessary corrections are applied to abnormal prediction values. Finally, the model's performance is comprehensively evaluated using the error evaluation metrics mean squared error and coefficient of determination to verify the model's prediction accuracy and practical value.
[0165] 4. Establish a result verification model
[0166] To verify the accuracy of the proposed method, validation data is used to verify the accuracy of the training model.
[0167] 1. Collect information related to vehicle speed, mark it accordingly, input the marked data into the classification model, conduct comparative analysis, and verify the accuracy of the classification model.
[0168] 2. Collect information related to the vehicle's instantaneous fuel consumption, input the collected information into the prediction model, compare the predicted data with the collected data, and verify the accuracy of the prediction model.
[0169] Example 12:
[0170] A fuel consumption prediction method based on vehicle driving behavior classification is verified, comprising the following steps:
[0171] A data collection experiment was conducted using a certain type of commercial vehicle as the test object, collecting information such as speed, fuel consumption, and altitude on a city's public roads. The pre-processed data is as follows: Figure 1 shown.
[0172] Based on the preprocessed vehicle driving data, kinematic segments were divided. A total of 18,127 kinematic segments were obtained. To accurately characterize the state and characteristics of each kinematic segment, the following nine characteristic parameters were selected to construct an evaluation system for vehicle driving condition characteristics. The characteristic parameter values for each segment were calculated and constructed into a characteristic parameter matrix. The parameter values are shown in Table 3.
[0173] Table 3 Feature parameter matrix
[0174]
[0175] Dimensionality reduction through principal component analysis can greatly improve computational efficiency and save computing resources. During the construction of vehicle driving conditions, the characteristic parameters mentioned in the above table have certain correlations. Principal component analysis can use fewer characteristic parameter dimensions to represent driving data, significantly improving the computational efficiency of the characteristic parameter matrix in cluster analysis. The principal component score matrix is obtained and used as the basis for subsequent clustering, as shown in Table 4.
[0176] Table 4 Principal component score matrix
[0177]
[0178]
[0179] The K-means clustering algorithm is used to cluster the solved principal component score matrix. The clustering results are shown in Table 5.
[0180] Table 5 Clustering results
[0181]
[0182] As can be seen from the table, in category 1, the average speed is high, the standard deviation and jerk of acceleration are minimal, and the accelerator pedal is moderate, indicating a calm driving state. In category 2, the average speed is high, the standard deviation and jerk of acceleration are small, and the mean, standard deviation, and rate of change of the accelerator pedal are large, indicating a normal driving state. In category 3, the average speed is low, and the standard deviation and jerk of acceleration are large, indicating an aggressive driving state.
[0183] The dataset containing features and labels is divided into training set and test set according to the preset ratio. In order to enhance the generalization ability of the model, the random forest classifier is selected as the modeling tool and the model is trained using the training set. The model adopts an integrated structure composed of multiple decision trees and enables the out-of-bag (OOB) verification mechanism to achieve self-evaluation of the model's performance during the training process. The training results are as follows: Figure 2 shown.
[0184] By plotting the out-of-bag error of the random forest model during training, we can observe how the model's performance changes with the number of decision trees. The results show that as the number of trees increases, the out-of-bag error decreases significantly and eventually stabilizes, validating the classification accuracy and generalization ability of the constructed model and demonstrating its good adaptability and robustness in the driving style recognition task. The figure shows the confusion matrix of the driving style recognition model. As can be seen, the model achieved over 99% accuracy for all driving styles on the test data, with 607 and 276 correctly classified examples for class 1 and class 2, respectively, and very few false positives. This demonstrates the good classification performance and stability of the constructed random forest model in the real-world driving style classification task.
[0185] The sample data that has not participated in the model training is input into the classification model as a test set, and the model is used to predict its corresponding driving style category. The classification results are as follows: Figure 3 As shown in the figure, the original cluster labels generated during the unsupervised clustering phase for the test data are obtained as the "true labels." By comparing the consistency between the predicted and true cluster categories and calculating the number of prediction errors, the accuracy and recognition ability of the model are evaluated.
[0186] As can be seen from the figure, the model's recognition accuracy for various driving styles on sample data that did not participate in model training is over 90%. Among them, the number of correct classifications for class 1 and class 3 are 1024 and 16 respectively, and the number of incorrectly identified samples for class 2 is relatively small, verifying that the constructed random forest model has good classification performance and stability in the actual driving style classification task.
[0187] Based on the classified driving style data, one category was selected for verification. Based on the grey correlation analysis method, the key parameter factors that have a significant impact on instantaneous fuel consumption were identified. The data mainly includes GPS vehicle speed, fan speed, oil temperature, water temperature, oil pressure, intake pressure, relative intake pressure, actual torque percentage, engine speed, friction torque percentage and transmission output shaft speed, totaling 11 characteristic parameters. The calculated correlation degree was plotted as a bar chart and the ranking results were output as follows: Figure 4 shown.
[0188] According to the analysis results, the selected features have a high correlation with instantaneous fuel consumption, with a mean value greater than 0.5, and are therefore selected as inputs for subsequent models.
[0189] First, the data is divided into training set and test set, 80% of the data is used as training samples, and the remaining 20% is used as test samples. In order to improve the training effect of the model, the input features of the training set and the test set are normalized respectively. A multi-layer feedforward neural network is constructed. The network contains three hidden layers, and the number of nodes is 100, 50 and 10 respectively. The model training function is set to the Bayesian Regularization method (trainbr), and the data in the training set is divided into training, verification and test subsets at a ratio of 80%, 10% and 10%. The maximum number of training cycles and the maximum number of verification failures are set to control the model training process, and the learning rate is set to stabilize the network convergence. The prediction effect of the training model is as follows: Figure 5 shown.
[0190] The prediction results show that the mean square error (MSE) of the trained model is about 0.7870, and the coefficient of determination (R 2 ) is 0.9985, indicating that the model has high prediction accuracy and stability, and can better realize the real-time prediction of fuel consumption.
[0191] In order to further verify the generalization performance of the neural network model, another set of samples that did not participate in the training was used as an external test set to input the model for prediction. The verification results are as follows: Figure 6 The validation results show that the mean square error (MSE) is about 0.2022, and the coefficient of determination (R 2 ) is 0.9996. The results show that the predicted value is highly consistent with the actual fuel consumption value, the trend is basically the same, the average error remains at a low level, and the determination coefficient is also close to 1, which verifies that the model has good promotion ability and practical application value.
Claims
1. A fuel consumption prediction method based on automobile driving behavior classification, characterized in that: The following steps are involved: 1) Obtain the original speed time series and perform preprocessing to obtain the preprocessed speed time series; 2) Construct a driving style classification and recognition model; 3) Extracting features from the preprocessed speed time series to obtain feature parameters, which are then input into a driving style classification and recognition model to determine the current driving style type; 4) Build a fuel prediction model under different driving styles; 5) Obtain vehicle operating data and input it into the fuel prediction model corresponding to the current driving style type to obtain the instantaneous fuel consumption result.
2. The fuel consumption prediction method based on automobile driving behavior classification according to claim 1 is characterized in that: In step 1), the pre-processing step includes: Delete the data segments whose missing time length is greater than or equal to the preset value, and use the Lagrange interpolation method to fill the data segments whose missing time length is less than the preset value; Eliminate data segments whose idle time is greater than or equal to the preset value; Replace speed outliers that exceed the speed limit with the speed of the previous second.
3. The fuel consumption prediction method based on automobile driving behavior classification according to claim 2 is characterized in that: The data completed by Lagrange interpolation is as follows: Where: P n (x) is the interpolation polynomial, y i is the dependent variable value of the i-th known point, l i (x) is the i-th Lagrange basis function; x is the independent variable of the point to be interpolated; x i 、x j , is the independent variable of the i / jth known point.
4. The fuel consumption prediction method based on automobile driving behavior classification according to claim 1 is characterized in that: In step 2), the steps of constructing a driving style classification and recognition model include: 2.1) Acquire multiple original velocity time series and perform preprocessing to obtain multiple preprocessed velocity time series; 2.2) Extract features from each preprocessed velocity time series to obtain a feature parameter matrix; 2.2) Use principal component analysis to reduce the dimension of the feature parameter matrix to obtain the characteristic principal components and the corresponding principal component scores; 2.3) K-means clustering is used to cluster the principal component score matrix, thereby classifying the driving style types; the driving style types include aggressive, normal, and calm. 2.4) Based on the feature parameters and the corresponding driving style types, the random forest model is trained to obtain a driving style classification and recognition model.
5. The method for predicting fuel consumption based on automobile driving behavior classification according to claim 4, characterized in that: In step 2.2), the steps of reducing the dimension of the feature parameter set using principal component analysis include: 2.2.1) Perform region center normalization on the characteristic parameter matrix A to obtain the normalized matrix Z m×n ; Among them, the characteristic parameter matrix A is as follows: Where: A m×n is a matrix composed of different characteristic parameters of each fragment; a mn is the characteristic parameter; Normalized matrix Z m×n As shown below: Where z ij is the standardized characteristic parameter; a ij is the characteristic parameter; is the average value of the characteristic parameter; s j is the standard deviation; 2.2.2) Calculate the normalized matrix Z m×n The covariance matrix C is: Where cov(n,n) is the covariance; 2.2.3) Calculate the correlation coefficient matrix R, that is: Where: r mn is the correlation coefficient; j, k = 1, 2, ..., n; 2.2.2) Calculate the characteristic parameter λ of the correlation coefficient matrix R i , and calculate the contribution rate The top k characteristic parameters with the highest contribution rate are selected as principal components, and the contribution rate is used as the principal component score.
6. The method for predicting fuel consumption based on automobile driving behavior classification according to claim 4, characterized in that: The characteristic parameters include speed mean, speed standard deviation, acceleration absolute value mean, acceleration absolute value standard deviation, Jerk, Jerk standard deviation, throttle opening position mean, throttle opening position standard deviation, throttle opening position average change rate, and throttle opening position change rate standard deviation.
7. The method for predicting fuel consumption based on automobile driving behavior classification according to claim 1, characterized in that: In step 4), the steps of constructing the fuel prediction model include: 4.1) Obtain multiple sets of vehicle operating data under the same driving style type and the corresponding instantaneous fuel consumption; 4.2) Using normalization method to standardize the retained vehicle operation data; 4.3) Quantify the correlation between the standardized vehicle operation data and instantaneous fuel consumption to obtain a gray correlation index, and retain vehicle operation data with a correlation index greater than a threshold; 4.4) Constructing a training set and a test set based on the vehicle operation data and instantaneous fuel consumption retained in step 4.3); 4.5) Train and test the feedforward neural network using the training and test sets to obtain a fuel prediction model for the current driving style. 4.6) Change the driving style type and return to step 4.1) until a fuel prediction model corresponding to each driving style type is generated.
8. The method for predicting fuel consumption based on automobile driving behavior classification according to claim 7, characterized in that: In step 4.1), the vehicle operation data includes GPS vehicle speed, fan speed, oil temperature, water temperature, oil pressure, intake pressure, relative intake pressure, actual torque percentage, engine speed, friction torque percentage, and gearbox output shaft speed.
9. The method for predicting fuel consumption based on automobile driving behavior classification according to claim 7, characterized in that: During the training process of the feedforward neural network, the Bayesian regularized back-propagation algorithm is used to update the parameters of the feedforward neural network.
10. The method for predicting fuel consumption based on automobile driving behavior classification according to claim 7, characterized in that: In step 4.3), the grey relational degree is as follows: Where, Δ i (k) is the difference between the reference sequence and the comparison sequence, ρ is the resolution coefficient; γ i is the grey relational degree; i (k) is the correlation coefficient of the i-th comparison sequence at the k-th moment; min i,k is the global minimum deviation; max i,k is the global maximum deviation.