A vehicle fuel consumption prediction method based on ensemble learning

Abstract

The present invention discloses a vehicle fuel consumption prediction method based on ensemble learning. The present invention obtains on-board CAN data, uses the XGBoost method to select key features, introduces an improved Pelican optimization algorithm to optimize the hyperparameters of the deep echo state network, and constructs a diesel truck fuel consumption prediction model to obtain fuel consumption prediction results, thereby achieving high-precision prediction of vehicle fuel consumption and further improving the practicality and economy of the fuel consumption prediction method.

Description

Technical Field

[0001] The present invention relates to a fuel consumption prediction method, and in particular to a vehicle fuel consumption prediction method based on ensemble learning. Background Art

[0002] With the rapid development of society and the continuous growth of transportation demand, China's road freight transportation volume has increased year by year. This makes it particularly important to accurately predict vehicle fuel consumption. The fuel consumption of road freight trucks depends not only on factors such as the vehicle's own characteristics, driving conditions, cargo type and load, but also on the actual driving environment and external conditions. In particular, in traffic congestion, frequent acceleration and deceleration and low average speeds can significantly increase fuel consumption and emissions. Although existing technologies have attempted to use various sensors and specialized equipment for direct measurement and prediction, these methods are often limited by high costs, implementation complexity and real-time issues. In addition, methods that rely solely on vehicle dynamics models and neural networks for fuel consumption estimation are difficult to avoid errors due to data noise, nonlinear system characteristics, and the need for large amounts of computing resources. In complex driving environments and road conditions, the prediction bias of traditional methods is also relatively significant.

[0003] The Chinese patent application with application number 202010918106.5, "A method and device for evaluating the energy consumption of heavy-duty diesel vehicles", uses the vehicle's ignition switch signal and engine status signal to determine the mileage, and collects driving data based on the CAN bus. By refining and extracting key features, and using the Pearson correlation coefficient to screen out features that are highly correlated with energy consumption. Subsequently, a pre-trained neural network-based model is used to combine these features for energy consumption evaluation. Disadvantages: The Pearson correlation coefficient has limitations and may ignore certain nonlinear or more complex feature relationships. A large amount of data is required to train the neural network, and there may be evaluation biases for certain specific driving scenarios or uninvolved models.

[0004] Chinese patent application number 202210310009.7, "Transient Automobile Fuel Consumption Estimation Method and System Based on Correction of Strongly Correlated Parameters of Fuel Consumption," proposes a method for estimating fuel consumption based on steady-state and non-steady-state vehicle driving data. This method combines multiple data analysis techniques, such as principal component analysis, cluster analysis, and polynomial fitting, to establish a vehicle fuel consumption model. Disadvantages: It does not take into account the impact of traffic flow conditions. For example, on busy urban roads or in traffic jams, frequent vehicle starts and stops may lead to increased fuel consumption. These unaccounted actual driving scenario factors may cause fuel consumption estimates to deviate in some cases.

[0005] The Chinese patent application with application number 202111517192.X, "Engine Dynamic Fuel Consumption Prediction Method and System Based on BP Neural Network", uses the average value filtering method and VSP-3σ method to pre-process fuel consumption data and eliminate outliers. The method is divided into a steady-state module and a transient correction module. The steady-state module uses a polynomial structure fitting to output a steady-state fuel consumption rate based on the engine torque and speed. The transient correction module considers speed, acceleration and engine speed, fits through a BP neural network, and outputs the difference between the actual and steady-state fuel consumption rates for correcting the prediction. To improve accuracy, the method considers the vehicle's specific power VSP and uses two BP networks to fit data in different VSP intervals. Verification shows that the prediction method has high accuracy. Disadvantages: The BP neural network may fall into a local minimum; new fuel consumption data in actual driving may affect the prediction; two BP neural networks with the same structure are used for fitting, which means that double the training time and computing resources are required.

[0006] Patent application number CN 111460381B, "Gaussian Process Regression-Based Multi-Operating Condition Fuel Consumption Prediction Method and System for Fuel-Engine Vehicles," provides a fuel consumption prediction solution for fuel-electric vehicles. This solution uses a Gaussian process regression model for training and a sequential sampling algorithm to optimize the dataset. This algorithm is repeated until a predetermined stopping condition is met, after which the prediction result is output. However, the repeated use of sequential sampling can increase computation time, and the Gaussian process-based model can require significant computational resources.

[0007] Patent application number 202210511971.7, "Method and Apparatus for Predicting Vehicle Fuel Consumption Based on Machine Learning and GPS Positioning," proposes a fuel consumption prediction method that combines GPS and the XGBoost machine learning algorithm. This method first collects a training dataset through the GPS vehicle management system and extracts the required algorithm dataset from it. After model training with the XGBoost algorithm, this method can obtain and predict the vehicle's actual fuel consumption in real time. Disadvantages: Relying on the accuracy of GPS data, data bias may lead to prediction errors; predictions based on historical data may be inaccurate when faced with new road conditions.

[0008] The existing technology has the following deficiencies:

[0009] 1. Data Accuracy and Training Requirements: Most methods rely heavily on data accuracy. For example, GPS-based fuel consumption prediction methods can suffer from data bias, which can lead to prediction errors. Furthermore, some methods, such as neural network-based models, require extensive data training and may have biased assessments for specific driving scenarios or unrelated vehicle models.

[0010] 2. Computational resource and time constraints: Many methods, such as Gaussian process-based models and BP neural networks, require significant computational resources, which can lead to extended computation times and make them difficult to meet the demands of real-time applications. For example, using two BP neural networks with the same structure for fitting requires twice the training time and computational resources.

[0011] 3. Local feature issues: Some methods use technical limitations that may ignore certain nonlinear or more complex feature relationships. For example, using the Pearson correlation coefficient may ignore certain nonlinear or complex feature relationships, while BP neural networks may fall into local minima.

[0012] 4. External environmental factors not considered: Some methods fail to fully account for the potential variability of the actual driving environment, such as traffic flow conditions and unexpected situations. For example, transient vehicle fuel consumption estimation methods based on the correction of parameters strongly related to fuel consumption fail to consider the impact of the vehicle's nonlinear system characteristics, which may lead to prediction errors in specific actual driving scenarios.

[0013] Therefore, developing a fuel consumption prediction method that can combine multiple data, reduce hardware dependence, and ensure high real-time performance and high accuracy has become a research and application need. Summary of the Invention

[0014] Purpose of the invention: The purpose of the present invention is to provide a fuel consumption prediction method that can combine multiple data, reduce hardware dependence, and ensure high real-time performance and high accuracy.

[0015] Technical solution: A vehicle fuel consumption prediction method based on ensemble learning, characterized by comprising the following steps:

[0016] Step 1: Obtain basic parameters of a vehicle with CAN function;

[0017] Step 2: Collect vehicle CAN data in real time;

[0018] Step 3: Preprocess the acquired vehicle CAN data according to the time window;

[0019] Step 4: Filter key features that are highly correlated with fuel consumption from the pre-processed vehicle CAN data as prediction input features;

[0020] Step 5: Divide the preprocessed CAN data into training set and test set in a ratio of 7:3;

[0021] Step 6: Build an improved Pelican optimization algorithm that introduces Tent chaos mapping, nonlinear weight factors, Cauchy mutation strategy, and sparrow alert mechanism;

[0022] Step 7: Use the output of step 6 as the optimal hyperparameters to construct a deep echo state network;

[0023] Step 8: Based on the training set in step 5, the network constructed in step 7 is trained so that its accuracy in the test set reaches a threshold of 90%, thereby obtaining an integrated learning model for predicting diesel truck fuel consumption and outputting the prediction results.

[0024] Step 9: Construct an echo state network, support vector machine, and random forest, compare the prediction results, and select the best result for output.

[0025] Furthermore, the preprocessing in step 3 includes checking time jumps, filling in missing data, and adjusting or eliminating abnormal or illogical data.

[0026] Furthermore, in step 4, the XGBoost method is used to select 10 key features that are highly correlated with fuel consumption from the preprocessed vehicle CAN data as prediction input features.

[0027] Furthermore, j )·z i (α), i=1, 2,..., N, j=1, 2,..., m

[0028]

[0029] Where x i，j is the value of the jth dimension of the i-th pelican, l j and u j are the lower and upper bounds of the dimension, respectively, z i (α) is the chaotic sequence, α is a constant. N is the population size; m is the dimension;

[0030] 6.3 Calculate the fitness value of each pelican individual as follows:

[0031]

[0032] Where X is the entire pelican population, and each X i Each is a candidate solution, F is the fitness function value, N is the population size, and m is the dimension;

[0033] 6.4 Determine whether the number of iterations has reached the maximum value T. If not, proceed to step 6.5. If so, proceed to step 6.14.

[0034] 6.5 Randomly select prey individuals and record their fitness values;

[0035] 6.6 introduces a nonlinear weighting factor ω during the exploration phase to adjust the relevance of the pelican's position update to the current position information, as follows:

[0036]

[0037] Where, is the updated value of the i-th candidate solution in the j-th dimension, p j is the position of the prey in the ,th dimension, I is a random number, F p and F i are the objective function values ​​of the prey and candidate solution, respectively. rand is a random number between [0, 1]. ω is small in the early stages of the algorithm iteration, which is beneficial to global search. It becomes larger as the number of iterations increases, improving local search capabilities.

[0038] 6.7 Compare the individual fitness value with the prey fitness value and calculate the new individual position;

[0039] 6.8 Determine whether the pelican has passed through the water surface. If so, proceed to step 6.9. If so, proceed to step 6.10.

[0040] 6.9 During the water-skimming phase (development phase), the Cauchy mutation strategy is introduced to increase diversity, as follows:

[0041]

[0042] Where, is the updated value of the i-th candidate solution in the dimension, Cauchy(σ) is the Cauchy distribution, σ is the mutation intensity, and t is the iteration counter. Then execute step 6.11;

[0043] 6.10 During the water skimming phase (development phase), the original update method is used, as follows:

[0044]

[0045] In the formula, R is a constant and the coefficient The neighborhood radius of the population member decreases as the number of iterations increases, and then step 6.11 is executed

[0046] 6.11 adds a sparrow alert mechanism, which allows pelicans that are aware of danger to quickly move to a safe area. Pelicans in the middle of the flock move randomly to speed up convergence and improve global search capabilities. The details are as follows:

[0047] 1) For the pelican in danger (marginal individual), the update formula is as follows:

[0048] x i,j =x i,j +rand·(x best，j -x i,j )

[0049] Where x best，j is the current optimal solution position of the population.

[0050] 2) For the pelican in a safe state (the middle individual), the update formula is as follows:

[0051]

[0052] Where randn is a standard normal distribution random number, is the mean position of the population.

[0053] 6.12 Calculate the fitness value of the new individual and update the individual position as follows:

[0054]

[0055] Where, and are the new positions after the exploration and development phases, and are their objective function values ​​respectively;

[0056] 6.13 The number of algorithm iterations t increases by 1 and step 6.4 is executed;

[0057] 6.14 Output the optimal pelican individual x best And the optimal fitness value F best .

[0058] Furthermore, the method of step 7 is as follows:

[0059] 7.1 At each time step, for the first layer, the state equation of DeepESN is updated as follows:

[0060]

[0061] Where x is the internal state of the network at a given time point, u is the external input signal received by the network, a is the leakage rate, and W (1) is the input weight matrix, W (i) is the inter-layer connection weight matrix, is the recursive weight matrix, a (i) is the leakage rate, f is the activation function,

[0062] 7.2 The states of all layers at each time step are used as input to the output layer to calculate the output, as follows:

[0063] y(t)=W out x(t)

[0064] Where W out is the output layer weight matrix.

[0065] Furthermore, in step 9, the comparison indicators are mean absolute error (MAE), mean absolute percentage error (MAPE) and goodness of fit R2, as follows:

[0066]

[0067] Where yi is the true value of the i-th real fuel consumption, is the predicted value of fuel consumption for the i-th time, is the average value of the true fuel consumption, and n is the number of test set samples.

[0068] The present invention obtains vehicle-mounted CAN data, uses the XGBoost method to select key features, introduces an improved Pelican optimization algorithm to optimize the hyperparameters of the deep echo state network, and constructs a diesel truck fuel consumption prediction model to obtain fuel consumption prediction results, thereby achieving high-precision prediction of vehicle fuel consumption.

[0069] Compared with the prior art, the present invention has the following beneficial effects:

[0070] 1. Comprehensive Data Utilization and Real-Time Performance: This invention utilizes an ensemble learning approach, unlike traditional methods that rely heavily on data volume and features. It automatically updates on-board CAN data and reads it in real time. By comprehensively analyzing massive amounts of data, it ensures real-time and accurate fuel consumption prediction.

[0071] 2. Efficient feature screening: The present invention uses the XGBoost method to screen key vehicle condition features that are highly correlated with fuel consumption, significantly reducing feature redundancy and avoiding the problem of local feature limitations.

[0072] 3. Integrating the advantages of multiple models: This method effectively reduces the bias of a single model by combining multiple models for prediction. This approach combines the advantages of each model, with each model providing favorable data support for the others, resulting in more reliable and accurate prediction results.

[0073] 4. Simplicity and Cost-Effectiveness: Compared to other methods, this method only requires acquiring vehicle CAN data, eliminating the need for additional sensors or specialized monitoring equipment, thus reducing implementation costs. Furthermore, because ensemble learning significantly accelerates model training and prediction through parallel computing, it reduces reliance on extensive computing resources, further enhancing the method's practicality and cost-effectiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 is a flow chart of the method of the present invention;

[0075] Figure 2This is a flow chart of step 6 of the present invention. DETAILED DESCRIPTION

[0076] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be further described below.

[0077] Figure 1 The following is a specific process of the vehicle fuel consumption prediction method based on ensemble learning. Figure 1 Describe the relevant steps in detail:

[0078] Step 1. Obtain basic parameters of a vehicle with CAN functionality.

[0079] Step 2. Collect vehicle CAN data in real time.

[0080] Step 3. Preprocess the acquired vehicle CAN data according to the time window, including but not limited to: checking time jumps, filling missing data, and adjusting or eliminating abnormal or illogical data.

[0081] Step 4. Use the XGBoost method to select 10 key features that are highly correlated with fuel consumption from the preprocessed vehicle CAN data as prediction input features.

[0082] Step 5. Divide the preprocessed CAN data into training set and test set in a ratio of 7:3.

[0083] Step 6. Build an improved Pelican Optimization Algorithm (IPOA) that introduces Tent chaos mapping, nonlinear weight factors, Cauchy mutation strategy, and sparrow alert mechanism. The process is as follows: Figure 2 As shown, the details are as follows:

[0084] 6.1 Set the maximum number of algorithm iterations T.

[0085] 6.2 Use the Tent Chaos Map to initialize the pelican population within the specified range, as follows:

[0086] x i，j =l j +(u j -l j )·z i (α),i=1,2,…,N,j=1,2,…,m

[0087]

[0088] Where x i，j is the value of the jth dimension of the i-th pelican, lj and u j are the lower and upper bounds of the dimension, respectively, z i (α) is the chaotic sequence, α is a constant. N is the population size; m is the dimension;

[0089] 6.3 Calculate the fitness value of each pelican individual as follows:

[0090]

[0091] Where X is the entire pelican population, and each X i Each is a candidate solution, F is the fitness function value, N is the population size, and m is the dimension.

[0092] 6.4 Determine whether the number of iterations has reached the maximum value T. If not, proceed to step 6.5. If so, proceed to step 6.14.

[0093] 6.5 Randomly select prey individuals and record their fitness values.

[0094] 6.6 introduces a nonlinear weighting factor ω during the exploration phase to adjust the relevance of the pelican's position update to the current position information, as follows:

[0095]

[0096] Where, is the updated value of the i-th candidate solution in the j-th dimension, p j is the position of the prey in the ,th dimension, I is a random number, F p and F i are the objective function values ​​of the prey and candidate solution respectively, rand is a random number between [0, 1], ω is small in the early stage of the algorithm iteration, which is conducive to global search, and becomes larger as the number of iterations increases, improving the local search ability.

[0097] 6.7 Compare the individual fitness value with the prey fitness value and calculate the new individual position.

[0098] 6.8 Determine whether the pelican is skimming the water surface. If so, proceed to step 6.9. If so, proceed to step 6.10.

[0099] 6.9 During the water-skimming phase (development phase), the Cauchy mutation strategy is introduced to increase diversity, as follows:

[0100]

[0101] Where, is the updated value of the i-th candidate solution in the j-th dimension, Cauchy(σ) is the Cauchy distribution, σ is the mutation intensity, and t is the iteration counter. Then execute step 6.11.

[0102] 6.10 During the water skimming phase (development phase), the original update method is used, as follows:

[0103]

[0104] In the formula, R is a constant and the coefficient Represents the neighborhood radius of the population member, which decreases as the number of iterations increases. Then execute step 6.11

[0105] 6.11 adds a sparrow alert mechanism, which allows pelicans that are aware of danger to quickly move to a safe area. Pelicans in the middle of the flock move randomly to speed up convergence and improve global search capabilities. The details are as follows:

[0106] 1) For the pelican in danger (marginal individual), the update formula is as follows:

[0107] x i,j =x i,j +rand·(x best，j -x i，j )

[0108] Where x best，j is the current optimal solution position of the population.

[0109] 2) For the pelican in a safe state (the middle individual), the update formula is as follows:

[0110]

[0111] Where randn is a standard normal distribution random number, is the mean position of the population.

[0112] 6.12 Calculate the fitness value of the new individual and update the individual position as follows:

[0113]

[0114] Where, and are the new positions after the exploration and development phases, and are their objective function values ​​respectively.

[0115] 6.13 The number of algorithm iterations t increases by 1 and step 6.4 is executed.

[0116] 6.14 Output the optimal pelican individual x best And the optimal fitness value F best .

[0117] Step 7. Use the output from step 6.14 as the optimal hyperparameters to construct a Deep Echo State Network (DeepESN), as follows:

[0118] 7.1 At each time step, for the first layer, the state equation of DeepESN is updated as follows:

[0119]

[0120] Where x is the internal state of the network at a given time point, u is the external input signal received by the network, a is the leakage rate, and W (1) is the input weight matrix, W (i) is the inter-layer connection weight matrix, is the recursive weight matrix, a (i) is the leakage rate, and f is the activation function.

[0121] 7.2 The states of all layers at each time step are used as input to the output layer to calculate the output, as follows:

[0122] y(t)=W out x(t)

[0123] Where W out is the output layer weight matrix.

[0124] Step 8. Based on the training set in step 5, train the DeepESN network constructed in step 7 so that its accuracy in the test set reaches the 90% threshold, obtain the integrated learning model for predicting diesel truck fuel consumption, and output the prediction results.

[0125] Step 9. Construct an Echo State Network (ESN), a Support Vector Machine (SVM), and a Random Forest (RF), compare their prediction results, and select the best result for output. The comparison indicators are the Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and goodness of fit R2, as follows:

[0126]

[0127] Where yi is the true value of the i-th real fuel consumption, is the predicted value of fuel consumption for the i-th time, is the average value of the true fuel consumption, and n is the number of test set samples.

[0128] The above description is merely a preferred embodiment of the present invention and does not limit the present invention in any way. Any person skilled in the art who, without departing from the scope of the present invention, makes any equivalent substitution, modification, or other changes to the technical solution and technical content disclosed in the present invention shall be deemed to be within the scope of the present invention and still fall within the scope of protection of the present invention.

Claims

1. A vehicle fuel consumption prediction method based on ensemble learning, characterized in that: The steps include: Step 1: Obtain basic parameters of a vehicle with CAN function; Step 2: Collect vehicle CAN data in real time; Step 3: Preprocess the acquired vehicle CAN data according to the time window; Step 4: Filter key features that are highly correlated with fuel consumption from the pre-processed vehicle CAN data as prediction input features; Step 5: Divide the preprocessed CAN data into training and test sets in a ratio of 7:3; Step 6: Build an improved Pelican optimization algorithm that introduces Tent chaos mapping, nonlinear weight factors, Cauchy mutation strategy, and sparrow alert mechanism; Step 7: Use the output of step 6 as the optimal hyperparameters to construct a deep echo state network; Step 8: Based on the training set in step 5, train the network constructed in step 7 so that its accuracy in the test set reaches the 90% threshold, obtain an integrated learning model for predicting diesel truck fuel consumption, and output the prediction results; Step 9: Construct the echo state network, support vector machine and random forest, compare the prediction results, and select the best result for output. 6.9 During the water skimming phase (development phase), a Cauchy mutation strategy is introduced to increase diversity, as follows: ； Where, For the The candidate solution is The updated value of the dimension, is the Cauchy distribution, is the variation intensity, is the iteration counter, then execute step 6.11; 6.10 During the water skimming phase (development phase), the original update method is used as follows: ； Where, is a constant, the coefficient The neighborhood radius of the population member decreases as the number of iterations increases, and then step 6.11 is executed 6.11 Added a sparrow alert mechanism to make pelicans that are aware of danger move quickly to a safe area. Pelicans in the middle of the flock move randomly to speed up convergence and improve global search capabilities. The details are as follows: 1) For the pelican in a dangerous state (marginal individual), the update formula is as follows: ； Where, is the current optimal solution position of the population; 2) For the pelican in a safe state (the middle individual), the update formula is as follows: ； Where, is a standard normally distributed random number, is the mean position of the population.

2. The vehicle fuel consumption prediction method based on ensemble learning according to claim 1 is characterized in that: The preprocessing in step 3 includes checking time jumps, filling in missing data, and adjusting or eliminating abnormal or illogical data.

3. The vehicle fuel consumption prediction method based on ensemble learning according to claim 2 is characterized in that: In step 4, the XGBoost method is used to select 10 key features that are highly correlated with fuel consumption from the preprocessed vehicle CAN data as prediction input features.

4. The vehicle fuel consumption prediction method based on ensemble learning according to claim 3 is characterized in that: The optimization algorithm in step 6 is as follows: 6.1 Setting the maximum number of algorithm iterations ; 6.2 Use the Tent Chaos Map to initialize the pelican population within the specified range, as follows: ； ； Where, For the The first pelican The value of the dimension, and are the lower and upper bounds of the dimension, respectively. is a chaotic sequence, is a constant, N is the population size; m is the dimension; 6.3 Calculate the fitness value of each pelican individual as follows: ； ； Where, For the entire pelican population, each is a candidate solution. is the fitness function value, is the population size, is the dimension; 6.4 Determine whether the number of iterations has reached the maximum value If not, proceed to step 6.5; if reached, proceed to step 6.14; 6.5 Randomly select prey individuals and record their fitness values; 6.6 Introducing a nonlinear weighting factor in the prey movement phase (exploration phase) , used to adjust the relevance of Pelican's location updates to the current location information, as follows: ； Where, For the The candidate solution is The updated value of the dimension, For prey in The location of the dimension, A random integer , and are the objective function values ​​of prey and candidate solutions respectively, for A random number between It is small in the early stages of algorithm iteration, which is beneficial to global search. It becomes larger as the number of iterations increases, improving local search capabilities. 6.7 Compare the individual fitness value with the prey fitness value and calculate the new individual position; 6.8 Determine whether the pelican is skimming the water surface. If so, proceed to step 6.

9. If so, proceed to step 6.

10. 6.12 Calculate the fitness value of the new individual and update the individual position as follows: ； Where, and are the new positions after the exploration and development phases, and are their objective function values ​​respectively; 6.13 Number of Algorithm Iterations Increase by 1 and execute step 6.4; 6.14 Outputting the Optimal Pelican Individual and the optimal fitness value .

5. The vehicle fuel consumption prediction method based on ensemble learning according to claim 4 is characterized in that: The method of step 7 is as follows: 7.1 At each time step, for the first layer, the state equation of DeepESN is updated as follows: ； Where, is the internal state of the network at a given point in time, is the external input signal received by the network, is the leakage rate, is the input weight matrix, is the inter-layer connection weight matrix, is the recursive weight matrix, is the leakage rate, is the activation function, 7.2 The states of all layers at each time step are used as input to the output layer to calculate the output, as follows: ； Where, is the output layer weight matrix.

6. The vehicle fuel consumption prediction method based on ensemble learning according to claim 5 is characterized in that: In step 9, the comparison indicators are mean absolute error (MAE), mean absolute percentage error (MAPE), and goodness of fit R2, as follows: ； ； ； Where, For the i The actual fuel consumption value, For the i The fuel consumption prediction value, is the average value of the actual fuel consumption, n is the number of test set samples.

7. The vehicle fuel consumption prediction method based on ensemble learning according to claim 1 is characterized in that: The vehicle is a diesel truck.

Images