Genetic simulated annealing-markov chain method for vehicle driving cycle construction

Abstract

The application discloses a genetic simulated annealing-Markov chain method for vehicle driving working conditions, which comprises the following steps: collecting vehicle driving data based on a vehicle terminal and a vehicle networking platform, preprocessing the vehicle driving data, and performing micro-trip clustering to generate speed sequences of different driving categories; then, generating speed state codes and acceleration state codes based on the speed sequences, establishing a state transition probability matrix, and constructing an initial working condition population; then, performing iterative optimization by using selection operators, crossover operators, search operators and update operators to obtain optimal working conditions of each driving category; and finally, linearly combining the optimal working conditions of each driving category to form a complete optimal driving working condition.

Description

Technical Field

[0001] This invention relates to the field of constructing automotive driving conditions, specifically to a genetic simulated annealing-Markov chain method for constructing automotive driving conditions. Background Technology

[0002] Currently, the mainstream methods for constructing vehicle driving conditions can be classified into two categories: the micro-travel method and the Markov chain method.

[0003] The micro-journey method divides the work into segments according to the definition of micro-journeys and synthesizes work case sequences from these segments. It is often combined with clustering methods to better analyze the relationships between micro-journey segments, thereby improving the accuracy of the constructed work cases. Furthermore, some researchers have introduced genetic algorithms to optimize micro-journey combinations, further improving the representativeness of the work cases and their accuracy in emission estimation. Although the use of various optimization methods has improved the accuracy of the micro-journey method, work case design methods based on micro-journey combinations still struggle to accurately control the work case length, and the problem of combinatorial explosion occurs when the data volume is too large, severely impacting the efficiency and accuracy of the algorithm.

[0004] The Markov chain method for constructing driving conditions typically involves first discretizing driving data into a state space, and then constructing a Markov chain model based on second-level state transitions. This method can precisely control the duration of driving conditions and effectively reflect the dynamic characteristics of the driving process. Compared with the micro-stroke method, the Markov chain method avoids the combinatorial explosion problem caused by segment combinations and can further introduce parameters such as acceleration and gradient to more comprehensively simulate actual driving conditions. The Markov chain method is essentially a random sampling technique. When the state space is large and the number of states is numerous, it is often difficult to obtain high-precision and representative driving conditions within a limited design time.

[0005] In recent years, the combination of Markov chains and optimization algorithms has become an emerging research hotspot. However, related research mainly focuses on the performance analysis of stationary distributions. Research on algorithms that accelerate convergence to satisfactory sequences for fixed-length sample sequences under known stationary distributions and state transition matrices remains limited. The selection and mutation processes of genetic algorithms exhibit high randomness, and crossover operations typically only drive the population as a whole towards a better direction, limiting their local search capabilities and thus restricting further improvements in the accuracy and efficiency of condition construction. Summary of the Invention

[0006] The purpose of this invention is to provide a genetic simulated annealing-Markov chain method for constructing chains under automotive driving conditions, comprising the following steps:

[0007] Step 1) Collect vehicle driving data based on the vehicle terminal and vehicle network platform, preprocess the vehicle driving data, perform micro-trip clustering on the preprocessed vehicle driving data, and generate speed sequences for different driving categories.

[0008] Step 2) Based on the velocity sequence, generate velocity state codes and acceleration state codes, and establish a state transition probability matrix;

[0009] Step 3) Construct an initial working condition population by random simulation based on the state transition probability matrix, and let the number of iterations be . , The initial value is 1;

[0010] Step 4) Calculate the first objective function based on the multi-objective function. The overall error of each individual under each working condition in the population;

[0011] The working condition population is iteratively optimized using selection, crossover, and search operators to obtain the... Generational population;

[0012] Step 5) Determine the first Is the comprehensive error of a single working condition in the first generation population less than that of the second generation? The comprehensive error of a single working condition corresponding to the generation population, if so, then the first generation... The working condition in the generation population is updated to the first If the corresponding working condition is in the population, then the update probability is determined, and the single working condition with the smallest comprehensive error is selected from the updated population to update it as the optimal working condition for each driving category.

[0013] Step 6) Let ;like If the number of iterations is less than the maximum number of iterations, return to step 4); If the number of iterations is greater than or equal to the maximum number of iterations, proceed to step 7).

[0014] Step 7) Linearly combine the optimal conditions for each driving category to form a complete optimal driving condition.

[0015] Furthermore, in step 1), the preprocessing includes: treating the driving data of congested road sections as idling, smoothing out periods of abnormal acceleration, and ensuring that the idling time is no more than t seconds;

[0016] The micro-trip clustering steps are as follows: extract the feature values ​​of the preprocessed vehicle driving data by dimensionality reduction using principal component analysis, group micro-trip segments with similar feature values ​​into the same type based on K-means clustering, and determine the optimal number of clusters;

[0017] The driving categories include low-speed, medium-speed, and high-speed conditions.

[0018] Furthermore, in step 2), The velocity encoding at any given time is shown below:

[0019] (1)

[0020] In the formula, Let be the velocity at time t. For the speed step size, To round down;

[0021] The acceleration encoding at time t is shown below:

[0022] (2)

[0023] In the formula, for Acceleration at all times For the minimum acceleration, This is the acceleration step size;

[0024] The joint state encoding at time t is shown below:

[0025] (3)

[0026] (4)

[0027] In the formula, This represents the maximum speed. Number of speed state codes.

[0028] Furthermore, the state transition probability matrix As shown below:

[0029] (5)

[0030] In the formula, For from the first The state transitions to the first state. The probability of each state; ; ; This represents the number of status codes.

[0031] Furthermore, from the first The state transitions to the first state. The probabilities of each state are as follows:

[0032] (6)

[0033] (7)

[0034] In the formula, For from the first The state transitions to the first state. The number of times each state is counted.

[0035] Furthermore, in step 4), the multi-objective function is as follows:

[0036] (8)

[0037] In the formula, The population consists of all operating conditions. This represents the overall error for a single operating condition.

[0038] Furthermore, the first The overall error for each individual under each working condition in the population is shown below:

[0039] (9)

[0040] In the formula, The number of feature parameters, For a single operating condition, The absolute error of the j-th characteristic of a single working condition;

[0041] The absolute error of the j-th feature of a single working condition is shown below:

[0042] (10)

[0043] In the formula, To construct the first working condition One feature parameter, For the first part of the original data Each feature parameter.

[0044] Furthermore, the selection operator is as follows:

[0045] (11)

[0046] In the formula, The first in the population Operating conditions, A number randomly generated between 0 and 1. For the first Fitness value for each working condition;

[0047] No. The fitness values ​​for the following working conditions are shown below:

[0048] (12)

[0049] In the formula, It is a constant. Let be the comprehensive error of the i-th working condition;

[0050] The crossover operator is shown below:

[0051] (13)

[0052] (14)

[0053] In the formula, and Two working conditions are randomly selected from the population, and subsequences of equal length are extracted.

[0054] For two subsequences to be swapped, the following condition must be met:

[0055] (15)

[0056] In the formula, for Transferred to The probability, for Transferred to The probability, for Transferred to The probability, for Transferred to The probability of;

[0057] The search operator refers to selecting a random-length subsequence of velocity from the optimal operating condition velocity sequence of the current population as the operation interval, with the initial state of the operation interval as the starting point. and termination state For fixed endpoints, the velocity subsequence between the two endpoints is regenerated based on the state transition probability matrix and replaced with the original velocity subsequence.

[0058] Furthermore, in step 5), the update probability judgment means that if the update probability is greater than the discriminant factor, then the first... The working condition in the generation population is updated to the first The corresponding working condition in the generation population; otherwise, no update is performed.

[0059] Furthermore, the first Generational population update probability As shown below:

[0060] (16)

[0061] (16)

[0062] In the formula, The attenuation coefficient is... This is the initial decay number. For the first The generation The overall error of the working conditions For the first The generation The overall error of each working condition; For the first The corresponding attenuation number; For the first The corresponding decay number.

[0063] The technical effects of this invention are undeniable, and its beneficial effects are as follows:

[0064] (1) A genetic simulated annealing optimization algorithm is introduced into the framework of the Markov chain method, and a search and update operator design that satisfies the Markov property is proposed. The search operator replaces the highly random mutation operation in the traditional genetic algorithm with neighborhood reconstruction of fixed endpoints, which significantly enhances the local fine search capability and stability of the algorithm; the update operator adaptively controls the convergence direction based on the Metropolis criterion, maintains population diversity in the early stage of iteration to escape local optima, and suppresses random perturbations in the later stage of iteration to lock the global optimum. Combining micro-travel clustering and velocity-acceleration dual-parameter state coding provides a new approach for high-precision working condition construction.

[0065] (2) Verification based on real-vehicle big data from multiple cities in China shows that the GSAMC algorithm has a significant advantage in the accuracy of condition construction. Under the condition of a maximum number of iterations of 500, the maximum error of each characteristic parameter in the constructed 1800-second condition is no more than 2.70% compared with the original data, the MAPE is as low as 1.01%, the average error of the VA distribution is only 0.22%, and the similarity coefficient is as high as 99.27%. Comparative analysis shows that compared with MCMC and MCE, the MAPE of the GSAMC algorithm is reduced by 7.11% and 4.73% respectively, and the error distribution of the VA distribution is more uniform, effectively overcoming the defect of the MCE algorithm in having too large errors in key parameters such as average acceleration. In addition, this invention achieves a good balance between computational efficiency and construction accuracy under the search interval setting of 10% sequence length.

[0066] (3) Based on the actual driving data of Chongqing, Guangzhou and Hangzhou, the constructed regional operating conditions can accurately capture the traffic flow characteristics of different cities. The average error of the characteristic parameters of the operating conditions in the three cities is controlled at 0.68%, 1.18% and 2.53% respectively, all below 3%. This shows that the present invention can generate operating condition curves that are highly consistent with the driving behavior of users in specific areas, making up for the shortcomings of general regulatory operating conditions such as WLTC and CLTC in reflecting the characteristics of specific areas. It provides a high-precision data foundation for vehicle calibration, energy consumption assessment and emission certification for specific scenarios. Attached Figure Description

[0067] Figure 1 The search process for satisfying the Markov property in this invention;

[0068] Figure 2 This is an enlarged view of the effective area of ​​the search operator of the present invention; wherein, (a) is a schematic diagram of the search operator, and (b) is an enlarged view of the effective area of ​​the search operator;

[0069] Figure 3 This is the population update process of the present invention;

[0070] Figure 4 This is the process for constructing the working conditions of this invention;

[0071] Figure 5 The vehicle's driving trajectory is shown in the figure.

[0072] Figure 6 This is a micro-stroke segment of the present invention;

[0073] Figure 7 This represents the contribution rate of the principal components in this invention.

[0074] Figure 8 This invention shows the variation trends of SSE and ASC with the number of clusters.

[0075] Figure 9 This illustrates the trend of the working error and time consumption of the present invention as a function of the search interval length.

[0076] Figure 10 The iterative change trend of the optimal solutions of GSAMC and MCE in this invention;

[0077] Figure 11 The results are constructed under different working conditions according to the present invention; where (a) is the GSAMC construction condition, (b) is the MCE construction condition, and (c) is the MCMC construction condition.

[0078] Figure 12 The following is a comparison of the VA distribution of different working conditions constructed by different methods of the present invention; wherein, (a) is the VA distribution of the original data, (b) is the VA distribution of the working conditions constructed by GSAMC, (c) is the VA distribution of the working conditions constructed by MCE, and (d) is the VA distribution of the working conditions constructed by MCMC. Detailed Implementation

[0079] The present invention will be further described below with reference to embodiments, but it should not be construed that the scope of the present invention is limited to the following embodiments. Various substitutions and modifications made based on ordinary technical knowledge and common practices in the art without departing from the above-described technical concept of the present invention should be included within the scope of protection of the present invention.

[0080] Example 1:

[0081] See Figures 1 to 12 The genetic simulated annealing-Markov chain method for constructing a system for vehicle driving conditions includes the following steps:

[0082] Step 1) Collect vehicle driving data based on the vehicle terminal and vehicle network platform, preprocess the vehicle driving data, perform micro-trip clustering on the preprocessed vehicle driving data, and generate speed sequences for different driving categories.

[0083] Vehicle driving data includes continuous speed and time series;

[0084] Step 2) Based on the velocity sequence, generate velocity state codes and acceleration state codes, and establish a state transition probability matrix;

[0085] Step 3) Construct an initial working condition population by random simulation based on the state transition probability matrix, and let the number of iterations be . , The initial value is 1;

[0086] Step 4) Calculate the first objective function based on the multi-objective function. The overall error of each individual under each working condition in the population;

[0087] The working condition population is iteratively optimized using selection, crossover, and search operators to obtain the... Generational population;

[0088] Step 5) Determine the first Is the comprehensive error of a single working condition in the first generation population less than that of the second generation? The comprehensive error of a single working condition corresponding to the generation population, if so, then the first generation... The working condition in the generation population is updated to the first If the corresponding working condition is in the population, then the update probability is determined, and the single working condition with the smallest comprehensive error is selected from the updated population to update it as the optimal working condition for each driving category.

[0089] Step 6) Let ;like If the number of iterations is less than the maximum number of iterations, return to step 4); If the number of iterations is greater than or equal to the maximum number of iterations, proceed to step 7).

[0090] Step 7) Linearly combine the optimal conditions for each driving category to form a complete optimal driving condition.

[0091] Example 2:

[0092] The main structure of this embodiment is the same as that of embodiment 1. Further, in step 1), the preprocessing includes: treating the driving data of the congested road section as idling, smoothing the abnormal acceleration period and idling time not exceeding t seconds.

[0093] A congested road section is defined as a road section where the speed is less than 2 m / s for 7 consecutive seconds.

[0094] An abnormal acceleration refers to an acceleration greater than 3m. -2 Or deceleration less than -3m -2 ;

[0095] The purpose of setting T=180 and idling time not to exceed 180 seconds is to avoid the impact of excessively long idling time on the working condition construction results. Idling data exceeding 180 seconds of continuous idling is directly removed.

[0096] The micro-trip clustering steps are as follows: extract the feature values ​​of the preprocessed vehicle driving data by dimensionality reduction using principal component analysis, group micro-trip segments with similar feature values ​​into the same type based on K-means clustering, and determine the optimal number of clusters;

[0097] The driving categories include low-speed, medium-speed, and high-speed conditions.

[0098] Example 3:

[0099] The main structure of this embodiment is the same as any one of embodiments 1-2. Further, in step 2), The velocity encoding at any given time is shown below:

[0100] (1)

[0101] In the formula, Let be the velocity at time t. For the speed step size, To round down;

[0102] The acceleration encoding at time t is shown below:

[0103] (2)

[0104] In the formula, for Acceleration at all times For the minimum acceleration, This is the acceleration step size;

[0105] The joint state encoding at time t is shown below:

[0106] (3)

[0107] (4)

[0108] In the formula, This represents the maximum speed. Number of speed state codes.

[0109] Example 4:

[0110] The main structure of this embodiment is the same as any one of embodiments 1-3. Furthermore, the state transition probability matrix... As shown below:

[0111] (5)

[0112] In the formula, For from the first The state transitions to the first state. The probability of each state; ; ; This represents the number of status codes.

[0113] Example 5:

[0114] The main structure of this embodiment is the same as any one of embodiments 1-4. Furthermore, starting from the... The state transitions to the first state. The probabilities of each state are as follows:

[0115] (6)

[0116] (7)

[0117] In the formula, For from the first The state transitions to the first state. The number of times each state is counted.

[0118] Example 6:

[0119] The main structure of this embodiment is the same as any one of embodiments 1-5. Further, in step 4), the multi-objective function is as follows:

[0120] (8)

[0121] In the formula, The population consists of all operating conditions. This represents the overall error for a single operating condition.

[0122] Example 7:

[0123] The main structure of this embodiment is the same as any one of embodiments 1-6. Furthermore, the... The overall error for each individual under each working condition in the population is shown below:

[0124] (9)

[0125] In the formula, The number of feature parameters, For a single operating condition, The absolute error of the j-th characteristic of a single working condition;

[0126] The absolute error of the j-th feature of a single working condition is shown below:

[0127] (10)

[0128] In the formula, To construct the first working condition One feature parameter, For the first part of the original data Each feature parameter.

[0129] Example 8:

[0130] The main structure of this embodiment is the same as any one of embodiments 1-7. Furthermore, the operator selection is as follows:

[0131] (11)

[0132] In the formula, The first in the population Operating conditions, A number randomly generated between 0 and 1. For the first Fitness value for each working condition;

[0133] No. The fitness values ​​for the following working conditions are shown below:

[0134] (12)

[0135] In the formula, It is a constant. ; Let be the comprehensive error of the i-th working condition;

[0136] The crossover operator is shown below:

[0137] (13)

[0138] (14)

[0139] In the formula, and Two working conditions are randomly selected from the population, and subsequences of equal length are extracted.

[0140] For two subsequences to be swapped, the following condition must be met:

[0141] (15)

[0142] In the formula, for Transferred to The probability, for Transferred to The probability, for Transferred to The probability, for Transferred to The probability of;

[0143] The search operator refers to selecting a random-length subsequence of velocity from the optimal operating condition velocity sequence of the current population as the operation interval, with the initial state of the operation interval as the starting point. and termination state For fixed endpoints, the velocity subsequence between the two endpoints is regenerated based on the state transition probability matrix and replaced with the original velocity subsequence.

[0144] Example 9:

[0145] The main structure of this embodiment is the same as any one of embodiments 1-8. Further, in step 5), the update probability judgment means that if the update probability is greater than the discriminant factor, then the first... The working condition in the generation population is updated to the first The corresponding working condition in the generation population; otherwise, no update is performed.

[0146] Example 10:

[0147] The main structure of this embodiment is the same as any one of embodiments 1-9. Furthermore, the... Generational population update probability As shown below:

[0148] (16)

[0149] (16)

[0150] In the formula, ; The attenuation coefficient is... ; This is the initial decay number. =1000; For the first The generation The overall error of the working conditions For the first The generation The overall error of each working condition; For the first The corresponding attenuation number; For the first The corresponding decay number.

[0151] Example 11:

[0152] 1. Markov chain method based on genetic simulated annealing

[0153] 1.1 Initial Population Construction

[0154] 1.1.1 Two-parameter state coding

[0155] Markov chains require discrete states, thus necessitating state partitioning and encoding of the original velocity sequence data. A state transition probability matrix is ​​constructed using two parameters: velocity and acceleration, with the velocity ranging from 0 to... The acceleration range is arrive When the step size is too large, the accuracy of the working condition construction is poor. When the step size is too small, the number of states increases significantly and the computational efficiency decreases. To balance computational efficiency and accuracy, a speed step size is set. equals 0.5 m·s -1 acceleration step size equals 0.5 m·s -2 Finally, the state is encoded.

[0156] Speed ​​of time Status coding :

[0157] (1)

[0158] acceleration at any moment Status coding :

[0159] (2)

[0160] For the state transition probability matrix with two parameters, velocity and acceleration... Joint state coding at time step :

[0161] (3)

[0162] (4)

[0163] In the formula —Number of speed coding states.

[0164] 1.1.2 State transition probability matrix

[0165] In each step of a Markov chain, the system can change from one state to another or remain in the current state. Each of these operations is recorded as a state transition, and the matrix consisting of all the state transition counts is the state transition matrix. The probability of changing or maintaining the state is called the state transition probability. The matrix consisting of all state transition probabilities is the state transition probability matrix. .

[0166] (5)

[0167] In the formula —From the first The state transitions to the first state. The number of times each state is represented;

[0168] — Number of status codes.

[0169] (6)

[0170] (7)

[0171] In the formula —From the first The state transitions to the first state. The probability of each state.

[0172] 1.1.3 State Decoding

[0173] Before generating the actual velocity sequence, it is necessary to calculate the state encoding sequence. The calculation formula for the state encoding based on the state transition probability matrix is ​​shown in equation (8):

[0174] (8)

[0175] In the formula, —Next state encoding; —A random number between 0 and 1.

[0176] pass The corresponding state code can be obtained, and the speed code can be calculated from the state code. The speed code represents the speed value within a certain range. The speed code is decoded using equation (9):

[0177] (9)

[0178] In the formula, —Predicted velocity value; —The speed state code corresponding to the state code after decoding.

[0179] 1.1.4 Construction of Working Condition Sets

[0180] To ensure the representativeness of driving conditions, their duration must be sufficiently long to fully cover various operational segments such as idling, acceleration, deceleration, and constant speed, and to ensure that the speed-time distribution accurately reflects the characteristics of actual traffic flow. If the condition length is too short, it cannot cover a sufficiently rich range of driving scenarios; if it is too long, it will significantly increase testing time and cost, and the improvement in condition representativeness will gradually decrease. In the current regulatory test cycle, the WLTC condition is divided into four stages: low speed, medium speed, high speed, and ultra-high speed, while the CLTC condition consists of three intervals: low speed, medium speed, and high speed. The total duration of both is fixed at 1800 seconds. Based on this, this paper adopts the same time benchmark as the aforementioned mainstream regulatory conditions, with the total condition length set at 1800 seconds. At the same time, a set of condition individuals is randomly generated as the initial population according to the state transition probability matrix, and the population size is set to 100 to balance the optimization effect of the algorithm and computational efficiency.

[0181] 1.2 Objective Function Design

[0182] The selection of reasonable driving characteristic parameters is the foundation for evaluating the accuracy of the driving condition construction and the optimization of the driving condition. This paper uses the statistical parameters listed in Table 1 to characterize the kinematic characteristics of each micro-stroke segment, covering key indicators such as average speed, average driving speed, average acceleration, average deceleration, acceleration ratio, idle ratio, and the ratio of each speed range.

[0183] The objective function is used to evaluate the performance of individual driving conditions in the algorithm and guide the convergence direction. To ensure that the aforementioned driving characteristic parameters can converge effectively during the optimization process, this paper constructs a multi-feature error unified objective function. This function comprehensively reflects the degree of difference between the characteristic parameters of the constructed operating conditions and the original data. Its design aims to obtain the operating condition sequence with the minimum comprehensive characteristic error. Objective Function By integrating the relative errors of each characteristic parameter into a single evaluation index, multi-objective convergence can be achieved under a unified dimension.

[0184] (10)

[0185] (11)

[0186] (12)

[0187] In the formula, —A single working condition constructed; —A population comprised of all operating conditions; —Comprehensive error of a single operating condition; —The number of feature parameters; —The first construction condition One feature parameter; —The first of the original data Each feature parameter.

[0188] Table 1 Feature Parameter Set

[0189]

[0190] 1.3 Design of a Genetic Simulated Annealing Algorithm Satisfying Markov Property

[0191] 1.3.1 Selection Operator

[0192] The selection process determines which conditions in the population will be used for evolution and to produce the next generation. This paper employs roulette wheel selection. In this method, each condition... The probability of being selected is directly proportional to its fitness value, meaning that conditions with higher fitness have a greater chance of being selected. The selection operator based on the roulette wheel selection method is shown in equation (13):

[0193] (13)

[0194] In the formula —the first in the population Operating conditions; ——No. Fitness value for each working condition; —A number randomly generated between 0 and 1; = .

[0195] In the design of this optimization algorithm, the fitness value is negatively correlated with the operating condition error; that is, the operating condition sequence with lower error should have a higher probability of being selected. Therefore, based on the distribution characteristics of operating condition errors, this paper constructs an exponential fitness function. This is designed to significantly enhance the adaptability advantage under low-error operating conditions.

[0196] (14)

[0197] 1.3.2 Crossover Operator

[0198] Crossover can recombine different candidate work cases to attempt to construct new work cases with lower feature errors. The work cases generated after the crossover operation should strictly follow the rules of Markov chains and be designed using crossover operators that satisfy Markov properties.

[0199] First, two work cases are randomly selected from the population. Then, subsequences of equal length are extracted from each of the two work cases. and If two subsequences satisfy the requirements of equation (16), then the two sequences can be exchanged. During the crossover process, the population is divided into two equal groups, and each time a working condition is selected from each of the two groups for crossover. The working condition sequence that has been selected will not be crossed over again.

[0200] (15)

[0201] (16)

[0202] 1.3.3 Search Operators

[0203] Traditional genetic simulated annealing algorithms generate new solutions by randomly perturbing a certain position in the solution through mutation, thus preventing the algorithm from getting trapped in local optima. Mutation operators that satisfy the Markov property first randomly select a mutation point in the sequence of operating conditions to be mutated; then, starting from the state code corresponding to that point, a new operating condition segment ending at the idling state is generated based on a pre-constructed state transition probability matrix. However, the mutation point in this method is not fixed. If the point is too early, it can easily lead to drastic changes in operating conditions and low computational efficiency; if the point is too late, the mutation effect may be insignificant.

[0204] To address the aforementioned issues, this paper proposes a neighborhood search operator that satisfies the Markov property, such as... Figure 1 As shown: A subsequence of random length is selected from the operating condition sequence as the operating interval, with the starting state of this interval... and termination state For fixed endpoints, the velocity subsequence between the two endpoints is regenerated based on the state transition probability matrix and replaced with the original sequence.

[0205] This method generates new feasible work case sequences while preserving the Markov property of the work cases. Given that the ultimate goal of work case construction is to obtain the sequence with the minimum error, and that the search operator has a relatively small impact on work case errors, to improve computational efficiency, only the work case sequence with the minimum current overall error is selected from the population for the search operation, avoiding excessive computation time caused by searching the entire population. Unlike the mutation operator, the neighborhood search operator focuses on optimizing and adjusting local segments of the work case, aiming to find better solutions within local regions, while the mutation operator may cause significant changes to the overall structure. Especially in the later stages of iteration, when the work case sequences already possess high accuracy, small-scale adjustments rather than large-scale mutations should be used to further improve accuracy.

[0206] like Figure 2 As shown, this operator achieves local reconstruction of the working condition by inserting a differentiated subsequence between a fixed starting point and an ending point. The proposed design not only preserves the algorithm's ability to escape local optima, but also significantly improves the algorithm's fine-grained search capability while strictly maintaining the Markov property consistency between the new segment and the original sequence.

[0207] 1.3.4 Updated operator design

[0208] In the later stages of iteration, the error magnitudes among different conditions within the population become similar. Due to the randomness of operators such as selection and crossover, the probability of superior individuals being eliminated increases, making it difficult for the algorithm to achieve higher accuracy in the later stages of evolution. To control the convergence direction of the algorithm, this paper proposes an update operator design based on the Metropolis criterion of the annealing algorithm.

[0209] The design of the update operator, such as Figure 3 As shown: For the old population before this iteration (the first iteration) (generation) and the new population after iteration (the first generation) The solutions from the two generations are sorted in ascending order based on the comprehensive error. Solutions with the same ranking in both generations are paired and updated accordingly. The pseudocode for this process is shown in Table 2. If the error of the solution in the new population is less than the corresponding solution in the old population, the corresponding solution is directly updated to the new solution. Otherwise, the update probability is used to determine whether to update. The calculation formula is shown in equation (17), where the discriminant factor is... A random number between 0 and 1. Greater than An update will occur if the condition is met; otherwise, no update will occur.

[0210] (17)

[0211] (18)

[0212] In the formula ——No. The generation The overall error of each working condition; ——No. The generation The overall error of each working condition; ——No. The corresponding attenuation number; ; =1000.

[0213] As shown in equation (17), the update probability depends on the operating condition error and the decay number. In the initial stage of the iteration, the decay number... With initial value When the errors are very close, the update probability is mainly governed by the operating condition error: if the operating condition error of the new individual is similar to that of the original individual, even if the error of the new individual is slightly higher than that of the original individual, the update probability still approaches 1, thus executing the update operation with a very high probability. This helps to explore optimization directions extensively in the early stages of the search and avoid getting trapped in local optima. In the later stages of iteration, Much smaller Furthermore, as the error continues to decrease, the difference in error between individuals is significantly amplified; when the error of a new individual is higher than that of the original individual, the update probability approaches 0. Therefore, the algorithm can retain the obtained superior solutions with a high probability, thus achieving adaptive control over the update frequency of the optimal solution in the population. The design of the above update operator not only promotes the convergence of the entire population towards a better solution but also effectively avoids the loss of the discovered optimal solution due to population replacement operations in the later stages of iteration.

[0214] Table 2. Update Algorithm Pseudocode

[0215]

[0216] In summary, this paper proposes a method for constructing driving conditions based on the GSAMC algorithm, the process of which is as follows: Figure 4 As shown in the diagram, firstly, the collected vehicle driving big data is preprocessed, and speed sequences corresponding to different driving modes are obtained through micro-trip clustering, serving as the foundational data for subsequent analysis. Based on this, optimal driving conditions are generated for each driving category. When constructing each driving condition, a state transition probability matrix is ​​established based on the dual parameters of speed and acceleration. Then, a set of driving conditions satisfying Markov properties is generated based on this matrix, forming the initial population. Subsequently, selection, crossover, and search operations are used to iteratively optimize the population, gradually bringing it closer to the global optimum. Simultaneously, an update operator is introduced to ensure that the optimization process always progresses along the target direction and achieves stable convergence in the later stages of optimization, thereby obtaining a more accurate solution. Finally, by linearly combining the optimal driving condition curves for each category, a complete driving condition curve is generated, ensuring that the constructed vehicle driving conditions best match the statistical characteristics and driving states of the original data.

[0217] 2. Driving Big Data Processing

[0218] 2.1 Data Preprocessing

[0219] Real-vehicle data was collected using in-vehicle terminals and a vehicle-to-everything (V2X) platform. This involved six months of driving data from 34 commercial vehicles of a certain type, covering Chongqing, Guangzhou, and Hangzhou and their surrounding areas. The data encompassed typical road types, including elevated roads and expressways, urban arterial roads, secondary roads, and local roads. The driving trajectories are shown below. Figure 5 As shown, a total of approximately 57.2 million sets of driving data were collected, with a total mileage exceeding 450,000 kilometers.

[0220] Because data acquisition may involve signal discontinuities, prolonged idling, and abrupt acceleration / deceleration, the extracted speed time series needs to be preprocessed to remove abnormal or invalid data and generate new speed time series.

[13] This mainly includes: treating congested road sections as idling, and handling accelerations greater than 3 m / s². -2 or less than -3 m·s -2 Smoothing out periods of abnormal acceleration, and ensuring that idling time does not exceed 180 seconds, etc.

[0221] 2.2 Micro-travel Extraction

[0222] A micro-travel segment is a speed-time segment that begins at one idle moment and ends at the next idle moment. It mainly includes four types of driving segments: acceleration, deceleration, constant speed, and idling. Figure 6 As shown, the velocity at the start and end of the micro-travel segment is zero, therefore this extraction method does not destroy the Markov property of the initial data. The acceleration segment is defined as an acceleration greater than or equal to 0.15 m / s². -2 The driving range includes a deceleration section where the acceleration is less than or equal to -0.15 m / s². -2 The driving range, the idling range is when the vehicle speed is equal to 0 m / s -1 The driving area.

[0223] 2.3 Micro-trip clustering

[0224] Clustering algorithms can group micro-travel segments with similar feature values ​​into the same type, compressing the state space dimension when modeling based on Markov chains in the future, and improving the efficiency and accuracy of the working condition construction process.

[0225] The dimensionality of feature parameters significantly affects the efficiency and results of clustering. Principal Component Analysis (PCA) is used to reduce the dimensionality of the features. Taking Chongqing driving data as an example, PCA is applied to the feature parameters, and the contribution rates of each principal component are as follows: Figure 7 As shown, when the cumulative contribution rate of the principal components exceeds 95%, the contribution rate of subsequent components decreases rapidly, indicating that the current principal components can fully reflect the data characteristics. The first 11 principal components are extracted as feature values ​​of the micro-journey segment.

[0226] The core of classifying micro-travel segments lies in determining the optimal number of clusters. The clustering performance under different numbers of clusters is evaluated by calculating the sum of squared errors (SSE) and the average silhouette coefficient (ASC). In the K-means algorithm, a smaller SSE and a larger ASC indicate better clustering performance. The trends of SSE and ASC for clustering with different numbers of clusters are shown below. Figure 8 As shown.

[0227] As the number of clusters increases, the SSE gradually decreases. When the number of clusters is greater than or equal to 3, the decrease in SSE decreases significantly. The ASC reaches its maximum value when the number of clusters is equal to 3, indicating that the current clustering effect is good. To balance clustering accuracy and computational efficiency, clustering analysis was performed with a cluster number of 3, resulting in 3 types of micro-travel segments. By splicing the micro-travel segments within each cluster together end-to-end, 3 types of speed sequences were reconstructed. Referring to the current regulatory test cycle, based on the average speed, average driving speed, and speed interval ratio differences of the clustered speed sequences, they were named low-speed, medium-speed, and high-speed driving modes, respectively. The statistical characteristic parameters of the three driving modes are shown in Table 3.

[0228] Table 3 Characteristic parameters of three types of driving modes

[0229]

[0230] 3. Working Condition Construction Analysis and Comparison

[0231] 3.1 Other working condition construction methods

[0232] 3.1.1 Markov Chain Stochastic Simulation Method

[0233] Markov Chain Monte Carlo (MCMC) Stochastic Simulation Method

[24] This method treats the vehicle speed-time series as a stochastic process with no aftereffect. First, it divides the original driving data into states, then constructs corresponding state transition probability matrices for driving condition construction. A Markov chain model based on a two-parameter "velocity-acceleration" model is used, where the state intervals for velocity and acceleration are set to 0.5 m·s⁻¹ and 0.5 m·s⁻², respectively. Driving condition sequences are constructed through stochastic simulation as typical driving conditions for various driving modes, and finally merged into complete driving conditions.

[0234] 3.1.2 Markov Chain Evolutionary Algorithm

[0235] Markov Chain Evolution Algorithm (MCE)

[26] Genetic algorithms are introduced to optimize the MCMC (Most Optimal Driving Condition) system. Specifically, the algorithm first generates an initial population of a certain size through random simulation. Then, iterative optimization is performed on the population using selection, crossover, and mutation operations, driving the population to converge toward the global optimum. Finally, the optimal solutions obtained in the iteration process for each driving condition are extracted and merged to form the complete optimal driving condition.

[0236] 3.2 Vehicle Driving Condition Construction

[0237] Using the categorized low-speed, medium-speed, and high-speed operating conditions as raw data, operating conditions were constructed using MCMC, MCE, and the proposed GSAMC algorithms, respectively. The maximum number of iterations for both the MCE and GSAMC methods was limited to 500. The experimental platform used an R7-8845H processor, and timing was implemented using MATLAB's built-in timer. The actual construction length for each operating condition was [not specified]. The total length of the velocity sequence for each type of working condition after clustering It accounts for the total length of the three types of velocity sequences The proportion is determined as shown in equation (19):

[0238] (19)

[0239] Based on vehicle driving data in Chongqing, driving conditions were constructed, with the durations of the three types of driving condition segments being: 276 seconds for low-speed driving condition, 619 seconds for medium-speed driving condition, and 905 seconds for high-speed driving condition.

[0240] In the GSAMC algorithm, the sequence operated on by the search operator is randomly selected. To compare the construction time and accuracy of the working condition under different search lengths, the maximum length of the search interval is set to 3%, 5%, 10%, 20%, and 50% of the length of the complete sequence, respectively. Figure 9 As shown, when the upper limit is less than 10%, although the computation time is shorter, the error of the constructed working condition is slightly higher. This is because the short search interval limits the range of changes in the fragments, resulting in a lack of significant improvement in accuracy. When the upper limit is greater than 20%, the excessively long interval can easily cause drastic changes in fragment features, introducing large errors and causing a sharp increase in the time required for the sequence selection process that satisfies the Markov property. Further comparison of the average time of the MCE algorithm under a 10% mutation probability shows that in the interval where the search ratio is less than 20%, GSAMC takes much less time than MCE, effectively improving computational efficiency.

[0241] When the search ratio is set to 10% of the complete sequence length, the GSAMC algorithm simultaneously ensures search efficiency and operational accuracy: it avoids the search efficiency bottleneck caused by insufficient dynamic information in short intervals, while also suppressing the risk of drastic operational changes caused by long intervals. Setting the upper limit of the search operator to 10% of the complete sequence length for operational condition construction, taking high-speed operational conditions as an example... Figure 10The iterative changes of the optimal solutions of MCE and GSAMC during the process are shown. The results indicate that both methods can effectively reduce the operating error when the initial operating error is relatively large in the early stages of iteration. The MCE algorithm has a higher error reduction rate in the early stages, but as the iteration progresses, the individual errors within the population gradually converge, and the overall error is reduced to a low level. Influenced by random factors such as selection and mutation operators, the error curve exhibits oscillations, making it difficult to further improve accuracy. In contrast, the GSAMC algorithm shows some oscillations in the early stages of iteration, but the fluctuations significantly weaken and stabilize in the later stages, with the operating error showing a steady downward trend. Its final operating accuracy is significantly higher than that of the MCE algorithm.

[0242] The results show that the update operator effectively enhances population diversity in the early stages of GSAMC algorithm iteration, helping the algorithm escape local optima. In the later stages of iteration, the operator helps maintain population stability, promoting the algorithm to converge smoothly to the global optimum. As the optimal solution gradually stabilizes, the search operator becomes more prominent, driving the accuracy of the optimal working condition in the population to gradually approach the theoretical optimum by strengthening the local fine search.

[0243] 3.3 Error Analysis of Build-up Working Conditions

[0244] The optimal driving conditions constructed by each algorithm in Section 3.2 are combined into a complete driving condition, and the result is as follows: Figure 11 As shown in Table 4, to verify the rationality of the constructed working condition, error analysis was conducted based on the characteristic parameters listed in Section 1.2, and the corresponding results are summarized in Table 4. This paper uses the Mean Absolute Percentage Error (MAPE) as a comprehensive evaluation index to quantify the degree of agreement between the constructed working condition and the overall statistical characteristics of the original data.

[0245] (20)

[0246] (twenty one)

[0247] In the formula ——No. The relative error of each characteristic parameter; ——Original data One feature parameter; —Construction Condition # Each feature parameter.

[0248] A comparative analysis of the accuracy of the three working condition construction methods on multi-dimensional feature parameters reveals that the GSAMC algorithm exhibits the best overall performance, with the lowest mean absolute percentage error (MAPE) of only 1.01%, and maintains high fitting accuracy across all feature parameters, with the maximum error for each parameter not exceeding 2.70%. The MCE algorithm has the second-highest overall accuracy (MAPE = 5.74%). While it shows lower errors on some parameters, its fitting error on key dynamic parameters such as mean acceleration is significant, reaching 17.68%, reflecting poor uniformity in the feature error distribution of the working conditions constructed by this method. The MCMC algorithm has the largest overall error (MAPE = 8.12%), with particularly prominent errors on features such as speed range distribution. In conclusion, the GSAMC algorithm demonstrates superior performance in both accuracy and feature balance, providing a more reliable technical approach for constructing high-precision vehicle driving conditions.

[0249] Table 4 Comparison of errors in constructing working condition characteristics using different methods

[0250]

[0251] Figure 12 The velocity-acceleration joint probability (VA) distributions of the original data and those constructed using three different methods were compared. To quantify the degree of similarity in distribution patterns, a VA distribution similarity coefficient was used.

[26] An evaluation was conducted. The results showed that the similarity coefficients of the GSAMC, MCE and MCMC construction conditions were 99.27%, 98.05% and 98.13%, respectively, all remaining at a high level, indicating that all three can well reflect the joint statistical characteristics of speed and acceleration during actual driving.

[0252] Further analysis of the spatial characteristics of the distribution error revealed that the maximum local errors of GSAMC, MCE, and MCMC were 1.90%, 2.74%, and 2.57%, respectively, and the average distribution errors were 0.22%, 0.28%, and 0.32%, respectively. Overall, the GSAMC algorithm showed a balanced performance in terms of overall VA similarity and local error control, and the driving conditions it constructed had high accuracy and reliability, accurately reproducing the VA distribution of actual driving characteristics.

[0253] (twenty two)

[0254] (twenty three)

[0255] In the formula —Similarity coefficient; —The first dimensional arrangement of the original data's VA distribution One component; —The first one-dimensional arrangement of the VA distribution of the construction conditions One component; —The total number of components after the one-dimensional arrangement of the VA distribution; — Average distribution error; —Construction Condition # The proportion of each distribution interval; ——Original data The proportion of each distribution interval; — The number of distribution intervals.

[0256] 3.4 Feasibility and Applicability Analysis

[0257] To verify the feasibility of the GSAMC method in engineering applications and demonstrate the necessity of constructing customized driving conditions for different regional driving conditions, driving conditions for each city were constructed using the GSAMC algorithm based on actual vehicle driving data collected from Chongqing, Hangzhou, and Guangzhou. The results were then compared and evaluated using characteristic parameter errors. Taking the eight key characteristic parameters listed in Table 1 (such as average speed, average acceleration / deceleration, and their ratios) as examples, the results are shown in Table 5. Significant differences were observed in these characteristic parameters across the three cities, directly demonstrating the significant impact of regional driving behavior and traffic environment on vehicle operating characteristics.

[0258] Table 5 further shows that the average error of each characteristic parameter for the three urban driving conditions constructed using the GSAMC method is less than 3% (0.68% in Chongqing, 1.18% in Guangzhou, and 2.53% in Hangzhou). This indicates that the method can maintain high accuracy in constructing driving conditions under different urban traffic flow characteristics and has good generalization ability. In summary, the GSAMC method can effectively construct high-precision driving conditions for specific users or regions, thus providing more reliable and accurate data support for vehicle energy consumption assessment, emission policy formulation, and related engineering application development.

[0259] Table 5. Construction Working Condition Characteristics and Errors in Different Cities

[0260]

[0261] 4. Conclusion

[0262] (1) A vehicle driving condition construction algorithm integrating genetic simulated annealing and Markov chains is proposed. The genetic simulated annealing optimization algorithm is introduced into the Markov chain method framework, and a search and update operator design satisfying Markov property is proposed. The search operator replaces the highly random mutation operation in the traditional genetic algorithm with neighborhood reconstruction at fixed endpoints, significantly enhancing the algorithm's local fine-grained search capability and stability. The update operator adaptively controls the convergence direction based on the Metropolis criterion, maintaining population diversity in the early iteration stage to escape local optima and suppressing random disturbances in the later iteration stage to lock in the global optimum. Combining micro-trip clustering and velocity-acceleration dual-parameter state encoding provides a new approach for high-precision driving condition construction.

[0263] (2) Verification based on real-vehicle big data from multiple cities in China shows that the GSAMC algorithm has a significant advantage in the accuracy of condition construction. Under the condition of a maximum number of iterations of 500, the maximum error of each characteristic parameter in the constructed 1800-second condition is no more than 2.70% compared with the original data, the MAPE is as low as 1.01%, the average error of the VA distribution is only 0.22%, and the similarity coefficient is as high as 99.27%. Comparative analysis shows that compared with MCMC and MCE, the MAPE of the GSAMC algorithm is reduced by 7.11% and 4.73% respectively, and the error distribution of the VA distribution is more uniform, effectively overcoming the defect of the MCE algorithm in having too large errors in key parameters such as average acceleration. In addition, the algorithm achieves a good balance between computational efficiency and construction accuracy under the search interval setting of 10% sequence length.

[0264] (3) Based on the actual driving data of Chongqing, Guangzhou and Hangzhou, the constructed regional operating conditions can accurately capture the traffic flow characteristics of different cities. The average error of the characteristic parameters of the operating conditions in the three cities is controlled at 0.68%, 1.18% and 2.53% respectively, all below 3%. This shows that the method can generate operating condition curves that are highly consistent with the driving behavior of users in specific areas, making up for the shortcomings of general regulatory operating conditions such as WLTC and CLTC in reflecting the characteristics of specific areas. It provides a high-precision data foundation for vehicle calibration, energy consumption assessment and emission certification for specific scenarios.

Claims

1. A genetic simulated annealing-Markov chain method for constructing models based on vehicle driving conditions, characterized in that... Includes the following steps: Step 1) Collect vehicle driving data based on the vehicle terminal and vehicle network platform, preprocess the vehicle driving data, perform micro-trip clustering on the preprocessed vehicle driving data, and generate speed sequences for different driving categories. Step 2) Based on the velocity sequence, generate velocity state codes and acceleration state codes, and establish a state transition probability matrix; Step 3) Construct an initial working condition population by random simulation based on the state transition probability matrix, and let the number of iterations be . , The initial value is 1; Step 4) Calculate the first objective function based on the multi-objective function. The overall error of each individual under each working condition in the population; The working condition population is iteratively optimized using selection, crossover, and search operators to obtain the... Generational population; Step 5) Determine the first Is the comprehensive error of a single working condition in the first generation population less than that of the second generation? The comprehensive error of a single working condition corresponding to the generation population, if so, then the first generation... The working condition in the generation population is updated to the first The corresponding working conditions in the generational population; Otherwise, perform an update probability judgment, and select the single working condition with the smallest comprehensive error from the updated population to update it as the optimal working condition for each driving category; Step 6) Let ;like If the number of iterations is less than the maximum number of iterations, return to step 4); If the number of iterations is greater than or equal to the maximum number of iterations, proceed to step 7). Step 7) Linearly combine the optimal conditions for each driving category to form a complete optimal driving condition.

2. The genetic simulated annealing-Markov chain method for constructing systems based on vehicle driving conditions as described in claim 1, characterized in that: In step 1), the preprocessing includes: treating the driving data of congested road sections as idling, smoothing out periods of abnormal acceleration, and ensuring that the idling time is no more than t seconds; the micro-trip clustering step is: extracting the feature values ​​of the preprocessed vehicle driving data by reducing the dimensionality using principal component analysis, grouping micro-trip segments with similar feature values ​​into the same type based on K-means clustering, and determining the optimal number of clusters; The driving categories include low-speed, medium-speed, and high-speed conditions.

3. The genetic simulated annealing-Markov chain method for constructing systems based on vehicle driving conditions as described in claim 1, characterized in that: In step 2), The velocity encoding at any given time is shown below: （1） In the formula, Let be the velocity at time t. For the speed step size, To round down; The acceleration encoding at time t is shown below: （2） In the formula, for Acceleration at all times For the minimum acceleration, This is the acceleration step size; The joint state encoding at time t is shown below: （3） （4） In the formula, This represents the maximum speed. Number of speed state codes.

4. The genetic simulated annealing-Markov chain method for constructing systems based on vehicle driving conditions as described in claim 3, characterized in that: State transition probability matrix As shown below: （5） In the formula, For from the first The state transitions to the first state. The probability of each state; ; ; This represents the number of status codes.

5. The genetic simulated annealing-Markov chain method for constructing systems based on vehicle driving conditions as described in claim 4, characterized in that: From the The state transitions to the first state. The probabilities of each state are as follows: （6） （7） In the formula, For from the first The state transitions to the first state. The number of times each state is counted.

6. The genetic simulated annealing-Markov chain method for constructing systems based on vehicle driving conditions as described in claim 1, characterized in that: In step 4), the multi-objective function is as follows: （8） In the formula, The population consists of all operating conditions. This represents the overall error for a single operating condition.

7. The genetic simulated annealing-Markov chain method for constructing systems based on vehicle driving conditions as described in claim 6, characterized in that: No. The overall error for each individual under each working condition in the population is shown below: （9） In the formula, The number of feature parameters, For a single operating condition, The absolute error of the j-th characteristic of a single working condition; The absolute error of the j-th feature of a single working condition is shown below: （10） In the formula, To construct the first working condition One feature parameter, For the first part of the original data Each feature parameter.

8. The genetic simulated annealing-Markov chain method for constructing systems based on vehicle driving conditions as described in claim 6, characterized in that: The operators are selected as follows: （11） In the formula, The first in the population Operating conditions, A number randomly generated between 0 and 1. For the first Fitness value for each working condition; No. The fitness values ​​for the following working conditions are shown below: （12） In the formula, It is a constant. Let be the comprehensive error of the i-th working condition; The crossover operator is shown below: （13） （14） In the formula, and Two working conditions are randomly selected from the population, and subsequences of equal length are extracted. For two subsequences to be swapped, the following condition must be met: （15） In the formula, for Transferred to The probability, for Transferred to The probability, for Transferred to The probability, for Transferred to The probability of; The search operator refers to selecting a random-length subsequence of velocity from the optimal operating condition velocity sequence of the current population as the operation interval, with the initial state of the operation interval as the starting point. and termination state For fixed endpoints, the velocity subsequence between the two endpoints is regenerated based on the state transition probability matrix and replaced with the original velocity subsequence.

9. A genetic simulated annealing-Markov chain method for constructing models oriented towards vehicle driving conditions, as described in claim 1, is characterized in that... In step 5), the update probability judgment means that if the update probability is greater than the discriminant factor, then the first... The working condition in the generation population is updated to the first The corresponding working conditions in the generational population; Otherwise, no update will be performed.

10. A genetic simulated annealing-Markov chain method for constructing models oriented towards vehicle driving conditions, as described in claim 9, is characterized in that... No. Generational population update probability As shown below: （16） （16） In the formula, The attenuation coefficient is... This is the initial decay number. For the first The generation The overall error of the working conditions For the first The generation The overall error of each working condition; For the first The corresponding attenuation number; For the first The corresponding decay number.

Images