New energy bus range estimation method facing working condition and driver

Abstract

The application provides a new energy bus driving range estimation method for working conditions and drivers, which extracts, performs principal component analysis and clustering on operation characteristic parameters of different personnel driving vehicles in sequence to obtain different working condition types which are related to objective factors such as road conditions and environment and subjective factors such as specific driver behaviors and habits, avoids the disadvantages in artificial explanation of principal components, and realizes as comprehensive coverage as possible of real complex working conditions by using machine learning. Through training and online application of a unit distance average energy consumption prediction model based on a deep belief network, the new energy bus vehicle can rapidly identify real-time working conditions according to actual driver and vehicle operation parameters, more accurately estimate the remaining driving range of the vehicle under the current working condition, and update the driving range estimation result as the vehicle position and working condition change.

Description

Technical Field

[0001] This invention belongs to the field of new energy vehicle range estimation technology, specifically involving a method for determining real-time operating conditions and quickly estimating the driving range by utilizing big data on the operation of new energy buses and combining it with driver information. Background Technology

[0002] Because new energy vehicles are affected by various factors such as the external environment and driver behavior during operation, the vehicle's operating conditions and energy consumption performance change constantly. This increases the difficulty of accurately estimating the remaining driving range, which is not conducive to alleviating the "range anxiety" that is prevalent in pure electric vehicles. Current technologies for estimating the driving range of pure electric new energy vehicles often rely on simple calculations based on real-time State of Charge (SOC), lacking consideration of actual operating conditions and thus suffering from severe accuracy deficiencies. Furthermore, the results fluctuate significantly over time. Some existing technologies first use vehicle operating data to classify several main operating condition types and their corresponding energy consumption models, then identify the corresponding operating condition based on real-time vehicle operating parameters to obtain the corresponding driving range estimate. However, since the main operating condition types often cannot comprehensively cover real-world conditions, and the differences in driving behavior and habits among different people facing the same external traffic and environmental conditions further increase the complexity of operating condition classification, this method clearly cannot meet the required estimation accuracy. Summary of the Invention

[0003] Considering that the daily driving routes and drivers of new energy buses are relatively fixed, the external environment in which the vehicles operate is relatively stable, and it is relatively easy to extract vehicle operation segments using bus stops, it is possible to comprehensively consider the external environment and the unique driving behaviors of different people in different environments, thereby achieving more accurate identification of complex operating conditions and calculation of remaining driving range. Based on this consideration, this invention addresses the technical problems existing in this field by providing a method for estimating the driving range of new energy buses based on operating conditions and drivers, specifically including the following steps:

[0004] Step 1: Extract real-vehicle operation data of a new energy bus driven by a specific driver on a specific bus route within a certain period and upload it to the big data platform; On the big data platform, based on the start and end time of the vehicle's journey from any stop to the next adjacent stop, the received real-vehicle operation data is segmented into different... m One running segment;

[0005] Step 2: Extract the relevant data from the actual vehicle operation data for each operation segment. n Each feature parameter, for m The average energy consumption per unit distance of the vehicle was calculated for each running segment. e avg;

[0006] Step 3: For data extracted from actual vehicle operation data... n Each runtime segment contains several characteristic parameters, and these parameters are standardized to eliminate dimensional differences; the standardized parameters are then used to eliminate dimensional differences. n Each feature parameter is used as a row element to construct a row vector, and then... m The row vectors of each running segment are constructed into a normalized feature parameter matrix of the following form. Q m×n The elements in the matrix q ij Indicates the first i The first running segment j One standardized feature parameter, i =1,2,…, m , j =1,2,…, n ;

[0007] Principal component analysis is performed on the standardized feature parameters, first calculating the matrix. Q m×n The elements in q ij Correlation coefficients and constructing a correlation coefficient matrix R Then, using the correlation coefficient matrix, the contribution rate and principal component score matrix of each principal component are calculated sequentially, thereby... n dimensionality reduction and determination of feature parameters are achieved. p One principal component;

[0008] Step 4, based on p Clustering is performed on the score vectors corresponding to each principal component, and the corresponding different working condition types are determined based on the obtained cluster centers;

[0009] Step 5: Extract feature parameters from historical operation segments for different operating conditions, and combine them with the average energy consumption per unit distance of the segments obtained in Step 2 to form an operation feature training set; use the average energy consumption per unit distance of the segments as the output and each feature parameter as the input to establish a prediction model for the average energy consumption per unit distance based on a deep belief network; the established deep belief network specifically consists of an input layer, an RBM layer, a BP layer, and an output layer.

[0010] The model is trained using the aforementioned operational feature training set, including: firstly, setting the network structure and number of neurons for each layer, as well as setting the activation function and evaluation function for the deep belief network; establishing an objective function with the goal of minimizing the root mean square error of the estimated average energy consumption per unit distance; optimizing the calculation using the Adam (Adaptive Moment Estimation) algorithm during training; and finally obtaining the trained average energy consumption prediction model per unit distance, enabling the trained model to output the average energy consumption per unit distance for each operational segment based on the feature parameters of different operating conditions.

[0011] Step 6: On the actual vehicle, use the sensors of the new energy bus to acquire real-time operating data corresponding to each feature parameter and determine the driver's identity. Based on the clustering results of the working condition type determined for the driver, identify the current working condition type of the vehicle; acquire the vehicle's location information to determine the current operating segment; and use the trained average energy consumption per unit distance prediction model to output the average energy consumption per unit distance under the same working condition type for the subsequent journey from the current route segment to the end of the route.

[0012] Step 7: Based on the estimated average energy consumption per unit distance obtained in Step 6, calculate the remaining driving range in combination with the vehicle's current remaining battery power; repeat Step 6 at certain time intervals and update the calculation of the remaining driving range.

[0013] Furthermore, in step two, the average energy consumption per unit distance of the vehicle in each segment is... e avg Specifically, it is calculated based on the vehicle's SOC or remaining battery power at the start and end points of the segment, combined with the segment distance.

[0014] Furthermore, the specific training process for the RBM layer in the deep belief network in step five is as follows:

[0015] Establish the weight matrix of the RBM layer respectively W Hidden layer bias vector b and visible layer bias vector c ;

[0016] The input data is used to assign values ​​to the visible layer and to calculate the conditional probability of each hidden layer neuron with respect to the visible layer.

[0017]

[0018] Where P(|) represents the conditional probability, v and h These are neurons in the visible layer and the hidden layer, respectively. s For the Sigmoid function, n This represents the total number of neurons in the visible layer.i and j This refers to the sequence number of a neuron in the hidden or visible layer.

[0019] Generate a random number in the range [0,1]. r j Update the values ​​of each hidden layer neuron to:

[0020]

[0021] Next, calculate the conditional probabilities of each visible layer neuron for the updated hidden layer:

[0022]

[0023] And generate a random number in the range [0,1]. s j Update the values ​​of neurons in each visible layer to:

[0024]

[0025] Then, the conditional probabilities of each hidden layer neuron with respect to the updated visible layer are recalculated:

[0026]

[0027] The aforementioned steps are repeated a predetermined number of iterations to achieve the desired weight matrix. W Hidden layer bias vector b and visible layer bias vector c Assignment update:

[0028]

[0029]

[0030]

[0031] in, l The superscript * represents the learning rate, and each assignment update process is indicated by the superscript *.

[0032] Furthermore, step five, which involves executing the Adam algorithm for optimization during training, is as follows:

[0033]

[0034]

[0035]

[0036] in, g t For the firstt Parameters at the next iteration i t That is, the gradient of the optimization objective, ▽ L (·) is the function for finding the gradient. m t For the first t The first Moment item in the next iteration. v t For the first t The second Moment item in the next iteration. or For learning rate, r 1 = 0.9 r 2 = 0.999, ∈ =10 -8 superscript ^ This represents the corresponding parameters after iterative updates.

[0037] The method provided by this invention extracts, performs principal component analysis and clustering on the operating characteristic parameters of vehicles driven by different personnel to classify different operating condition types that are correlated with objective factors such as road conditions and environment, as well as subjective factors such as specific driver behavior and habits. This avoids the drawbacks of manually interpreting principal components and utilizes machine learning to achieve comprehensive coverage of real and complex operating conditions. By training and applying a unit distance average energy consumption prediction model based on deep belief networks, new energy buses can quickly identify real-time operating conditions based on actual driver and vehicle operating parameters, accurately estimate the remaining driving range under the current operating condition, and update the driving range estimation results as the vehicle's location and operating conditions change. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the overall process of the method provided by the present invention. Detailed Implementation

[0039] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] The method for estimating the driving range of new energy buses based on operating conditions and drivers provided by this invention, such as... Figure 1 As shown, the specific steps include:

[0041] Step 1: Extract real-vehicle operation data of a new energy bus driven by a specific driver on a specific bus route within a certain period and upload it to the big data platform; On the big data platform, based on the start and end time of the vehicle's journey from any stop to the next adjacent stop, the received real-vehicle operation data is segmented into different... m One running segment;

[0042] Step 2: Extract the relevant data from the actual vehicle operation data for each operation segment. n Each feature parameter, for m The average energy consumption per unit distance of the vehicle was calculated for each running segment. e avg ;

[0043] Step 3: For data extracted from actual vehicle operation data... n Each runtime segment contains several characteristic parameters, and these parameters are standardized to eliminate dimensional differences; the standardized parameters are then used to eliminate dimensional differences. n Each feature parameter is used as a row element to construct a row vector, and then... m The row vectors of each running segment are constructed into a normalized feature parameter matrix of the following form. Q m×n The elements in the matrix q ij Indicates the first i The first running segment j One standardized feature parameter, i =1,2,…, m , j =1,2,…, n ;

[0044] Principal component analysis is performed on the standardized feature parameters, first calculating the matrix. Q m×n The elements in q ij Correlation coefficients and constructing a correlation coefficient matrix R Then, using the correlation coefficient matrix, the contribution rate and principal component score matrix of each principal component are calculated sequentially, thereby... n dimensionality reduction and determination of feature parameters are achieved. p One principal component;

[0045] Step 4, based on pClustering is performed on the score vectors corresponding to each principal component, and the corresponding different operating conditions are determined based on the obtained cluster centers. Unlike existing technologies, this application deliberately avoids the conventional approach of interpreting the principal components and simply classifying the operating conditions into several categories after determining the principal components, such as classifying them into smooth traffic, congestion, general traffic, or high, medium, and low speed operating conditions, and then combining them with the driver model to determine the specific operating condition. The problem with this approach is that even for the same operating condition, different drivers have vastly different behavioral habits. For example, in congestion, some drivers will choose to drive gently with the traffic flow, while others will adopt aggressive methods such as rapid acceleration and deceleration and frequent lane changes. This makes the complexity of real operating conditions even more complex and cannot be generalized. The difference between the results obtained by classifying operating conditions and driver behaviors will be further amplified and will obviously be detrimental to the accurate calculation of subsequent energy consumption and driving range.

[0046] Step 5: Extract feature parameters from historical operation segments for different operating conditions, and combine them with the average energy consumption per unit distance of the segments obtained in Step 2 to form an operation feature training set; use the average energy consumption per unit distance of the segments as the output and each feature parameter as the input to establish a prediction model for the average energy consumption per unit distance based on a deep belief network; the established deep belief network specifically consists of an input layer, an RBM layer, a BP layer, and an output layer.

[0047] The model is trained using the aforementioned operational feature training set, including: firstly, setting the network structure and number of neurons for each layer, as well as setting the activation function and evaluation function of the deep belief network; establishing an objective function with the goal of minimizing the root mean square error of the estimated average energy consumption per unit distance; using the Adam (Adaptive Moment Estimation) algorithm for optimization calculations during training; and finally obtaining the trained average energy consumption prediction model per unit distance. This allows the trained model to output the average energy consumption per unit distance for each operational segment based on the feature parameters of different operating conditions. The Adam algorithm effectively avoids the problem of getting trapped in local optima when using optimization algorithms such as gradient descent and stochastic gradient descent during neural network training. In the training process of the deep belief network, the operating conditions affected by various factors such as traffic, road conditions, temperature, and driver behavior are implicitly classified.

[0048] Step 6: On the actual vehicle, the new energy bus uses sensors to acquire real-time operating data corresponding to each feature parameter and determine the driver's identity. Based on the clustering results of the working condition type determined for the driver, the current working condition type of the vehicle is identified. The vehicle's location information is obtained to determine the current operating segment. The trained average energy consumption per unit distance prediction model is used to output the average energy consumption per unit distance under the same working condition type for the subsequent journey from the current route segment to the end of the route.

[0049] Step 7: Based on the estimated average energy consumption per unit distance obtained in Step 6, calculate the remaining driving range in combination with the vehicle's current remaining battery power; repeat Step 6 at certain time intervals and update the calculation of the remaining driving range.

[0050] In a preferred embodiment of the present invention, step two involves considering the average energy consumption per unit distance of the vehicle across each segment. e avg Specifically, it is calculated based on the vehicle's SOC or remaining battery power at the start and end points of the segment, combined with the segment distance.

[0051] In a preferred embodiment of the present invention, the specific training process for the RBM layer in the deep belief network in step five is as follows:

[0052] Establish the weight matrix of the RBM layer respectively W Hidden layer bias vector b and visible layer bias vector c ;

[0053] The input data is used to assign values ​​to the visible layer and to calculate the conditional probability of each hidden layer neuron with respect to the visible layer.

[0054]

[0055] Where P(|) represents the conditional probability, v and h These are neurons in the visible layer and the hidden layer, respectively. s For the Sigmoid function, n This represents the total number of neurons in the visible layer. i and j This refers to the sequence number of a neuron in the hidden or visible layer.

[0056] Generate a random number in the range [0,1]. r j Update the values ​​of each hidden layer neuron to:

[0057]

[0058] Next, calculate the conditional probabilities of each visible layer neuron for the updated hidden layer:

[0059]

[0060] And generate a random number in the range [0,1]. s j Update the values ​​of neurons in each visible layer to:

[0061]

[0062] Then, the conditional probabilities of each hidden layer neuron with respect to the updated visible layer are recalculated:

[0063]

[0064] The aforementioned steps are repeated a predetermined number of iterations to achieve the desired weight matrix. W Hidden layer bias vector b and visible layer bias vector c Assignment update:

[0065]

[0066]

[0067]

[0068] in, l The superscript * represents the learning rate, and each assignment update process is indicated by the superscript *.

[0069] In a preferred embodiment of the present invention, step five, which involves performing optimization calculations using the Adam algorithm during training, is as follows:

[0070]

[0071]

[0072]

[0073] in, g t For the first t Parameters at the next iteration i t That is, the gradient of the optimization objective, ▽ L (·) is the function for finding the gradient. m t For the first t The first Moment item in the next iteration. v t For the first t The second Moment item in the next iteration. or For learning rate, r 1 = 0.9 r 2 = 0.999, ∈ =10 -8 superscript ^ This represents the corresponding parameters after iterative updates.

[0074] It should be understood that the sequence number of each step in the embodiments of the present invention does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0075] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for estimating the driving range of new energy buses based on operating conditions and drivers, characterized by: Specifically, the following steps are included: Step 1: Extract real-vehicle operation data of a new energy bus driven by a specific driver on a specific bus route within a certain period and upload it to the big data platform; on the big data platform, based on the start and end time of the vehicle's journey from any stop to the next adjacent stop, segment the received real-vehicle operation data into different... m One running segment; Step 2: Extract the relevant data from the actual vehicle operation data for each operation segment. n Each feature parameter, for m Calculate the average energy consumption per unit distance of the vehicle on each running segment. e avg ; Step 3: For data extracted from actual vehicle operation data... n Each characteristic parameter is identified, and the dimensional differences of the characteristic parameters are eliminated through standardization. Each running segment contains, after standardization n Each feature parameter is used as a row element to construct a row vector, and then... m The row vectors of each running segment are constructed into a normalized feature parameter matrix of the following form. Q m×n The elements in the matrix q ij Indicates the first i The first running segment j One standardized feature parameter, i =1,2,…, m , j =1,2,…, n ; against Perform principal component analysis using standardized feature parameters, first calculating the matrix. Q m×n The elements in q ij Correlation coefficients and constructing a correlation coefficient matrix R Then, using the correlation coefficient matrix, the contribution rate and principal component score matrix of each principal component are calculated sequentially, thereby... n dimensionality reduction and determination of feature parameters are achieved. p One principal component; Step 4, based on p Clustering is performed on the score vectors corresponding to each principal component, and the corresponding different working condition types are determined based on the obtained cluster centers; Step 5: Extract feature parameters from historical operation segments for different operating conditions, and combine them with the average energy consumption per unit distance of the segments obtained in Step 2 to form an operation feature training set; use the average energy consumption per unit distance of the segments as the output and each feature parameter as the input to establish a prediction model for the average energy consumption per unit distance based on a deep belief network; the established deep belief network specifically consists of an input layer, an RBM layer, a BP layer, and an output layer. The model is trained using the aforementioned operational feature training set, including: firstly, setting the network structure and number of neurons for each layer, as well as setting the activation function and evaluation function of the deep belief network; establishing an objective function with the goal of minimizing the root mean square error of the estimated average energy consumption per unit distance; optimizing the calculation using the Adam algorithm during training; and finally obtaining the trained average energy consumption prediction model per unit distance, so that the trained model can output the average energy consumption per unit distance for each operational segment based on the feature parameters of different operating conditions. Step 6: On the actual vehicle, use the sensors of the new energy bus to acquire real-time operating data corresponding to each feature parameter and determine the driver's identity. Based on the clustering results of the working condition type determined for the driver, identify the current working condition type of the vehicle; acquire the vehicle's location information to determine the current operating segment; and use the trained average energy consumption per unit distance prediction model to output the average energy consumption per unit distance under the same working condition type for the subsequent journey from the current route segment to the end of the route. Step 7: Based on the predicted average energy consumption per unit distance of the vehicle during the subsequent journey obtained in Step 6, calculate the remaining driving range in combination with the vehicle's current remaining battery power; repeat Step 6 at certain time intervals and update the calculation of the remaining driving range.

2. The method as described in claim 1, characterized in that: Step two focuses on the vehicle's average energy consumption per unit distance across each segment. e avg Specifically, it is calculated based on the vehicle's SOC or remaining battery power at the start and end points of the segment, combined with the segment distance.

3. The method as described in claim 1, characterized in that: The specific training process for the RBM layer in the deep belief network in step five is as follows: Establish the weight matrix of the RBM layer respectively W Hidden layer bias vector b and visible layer bias vector c ; The input data is used to assign values ​​to the visible layer and to calculate the conditional probability of each hidden layer neuron with respect to the visible layer. Where P(|) represents the conditional probability, v and h These are neurons in the visible layer and the hidden layer, respectively. σ For the Sigmoid function, n This represents the total number of neurons in the visible layer. i and j This refers to the sequence number of a neuron in the hidden or visible layer. Generate a random number in the range [0,1]. r j Update the values ​​of each hidden layer neuron to: Next, calculate the conditional probabilities of each visible layer neuron for the updated hidden layer: And generate a random number in the range [0,1]. s j Update the values ​​of neurons in each visible layer to: Then, the conditional probabilities of each hidden layer neuron with respect to the updated visible layer are recalculated: The aforementioned steps are repeated a predetermined number of iterations to achieve the desired weight matrix. W Hidden layer bias vector b and visible layer bias vector c Assignment update: in, λ The superscript * represents the learning rate, and each assignment update process is indicated by the superscript *.

4. The method as described in claim 1, characterized in that: Step five, the process of executing the Adam algorithm for optimization calculations during training, is as follows: in, g t For the first t Parameters at the next iteration θ t That is, the gradient of the optimization objective, ▽ L (·) is the function for finding the gradient. m t For the first t The first Moment item in the next iteration. v t For the first t The second Moment item in the next iteration. η For learning rate, ρ 1 = 0.9 ρ 2 = 0.999, ∈ =10 -8 superscript ^ This represents the corresponding parameters after iterative updates.

Images