WO2021114742A1 - Comprehensive energy prediction and management method for hybrid electric vehicle - Google Patents

Abstract

A comprehensive energy prediction and management method for a hybrid electric vehicle, which combines vehicle speed prediction and passenger prediction and more comprehensively considers future short-term power requirements in real application scenarios of vehicles. Using a model prediction control-based method can achieve a real-time online application of vehicle energy management. The method can achieve prediction for the real-time online energy management of vehicles with more comprehensive working conditions and more accurate prediction accuracy, increase the energy-saving potential of a hybrid electric vehicle, and is easy to implement online and in real time, and has good engineering application prospects.

Description

A comprehensive predictive energy management method for hybrid electric vehicles Technical field

The invention relates to the technical field of energy management of hybrid electric vehicles, in particular to an online comprehensive predictive energy management method based on model predictive control and combining the speed of hybrid electric buses and the number of passengers.

Background technique

Although traditional hybrid electric vehicle energy management can perform global optimization to obtain the optimal solution, it cannot be applied online in real time, and the practical application value of energy management is lost. Then someone proposed online predictive energy management based on model predictive control. Normally, this predictive energy management only predicts vehicle speed; however, the energy management of hybrid electric vehicles is not only affected by vehicle speed and road conditions, but also by vehicle load. , This has an important impact on vehicle power demand and power allocation strategy. Take a hybrid electric bus in actual operation as an example. The full-load mass of the hybrid electric bus is 18 tons, the curb weight is 13 tons, and the mass of passengers carried is 5 tons. When fully loaded, the mass of passengers accounts for about half of the full-load mass of the vehicle. 30%. Sometimes when the vehicle is overloaded, the mass of passengers accounts for a larger proportion of the full vehicle mass. The mass of passengers is determined by the number of passengers. The load status of the vehicle, including the number of occupants, has a great influence on the power distribution and energy consumption of the vehicle. Therefore, considering the quality of the occupants when studying the energy management of hybrid electric vehicles is of great significance to the energy saving of hybrid electric vehicles.

Summary of the invention

In order to solve the above-mentioned technical problems in the prior art, the present invention provides a comprehensive predictive energy management method for hybrid electric vehicles, which specifically includes the following steps:

S01. Collect vehicle driving data and occupant data, and establish a database of vehicle historical driving conditions and a vehicle model;

S02. Forecast the future short-term vehicle speed and the number of future short-term occupants based on the deep neural network;

S03. Regard the hybrid electric vehicle as a multi-constrained nonlinear optimization problem, select the state quantity and control quantity to solve the problem, and construct an optimization objective function, aiming at the optimal energy consumption of the whole vehicle;

S04. Based on the vehicle model, using the future short-term vehicle speed and occupant data predicted in step S02, according to the optimization objective function, global optimization in the prediction time domain is used to obtain the optimal control sequence, and the first of the optimal control sequence A control instruction is input to the vehicle model for execution, so that the vehicle state quantity is updated and fed back to the next time for calculation;

S05. Repeat the process of steps S01S04 to finally complete the comprehensive predictive energy management of the entire vehicle driving condition.

Further, the vehicle travel data collected in step S01 includes: the vehicle's speed, acceleration, slope, steering, geographic location (latitude and longitude and altitude), travel date, including bus routes or driving routes of any travel route; The occupant data includes: the number of occupants, the location of the stop, and the time, and the time is marked as the peak period of the working day, the normal period of the working day, the low period of the working day, and the weekend.

Further, step S02 specifically includes establishing a deep neural network predictor to perform the prediction. The input of the predictor is the collected vehicle driving data and occupant data; the predicted future short-term vehicle speed is a short-time scale prediction, measured in seconds. As a unit of prediction, the predicted number of short-term occupants in the future is a long-time scale prediction, and the prediction is made in the unit of the stop location.

Further, step S03 specifically includes establishing an optimizer for the multi-constrained nonlinear optimization problem, and the optimization goal is the optimal energy consumption of the entire vehicle in the short-term prediction time domain.

Further, step S04 specifically includes the comprehensive predictive energy management of the hybrid electric vehicle described under the framework of the model predictive control method to perform working condition prediction, objective function optimization and state quantity feedback in the rolling forward time domain, and each step is based on Starting from the current moment, predict the vehicle speed and the number of occupants in the H time domain in the future, and input the predicted vehicle speed and the number of occupants in the H time domain to the optimizer for global optimization in the H time domain.

The H time domain generally takes a value in the range of 3-30 seconds, which not only guarantees the accuracy of prediction but also ensures the real-time performance of optimization.

By continuing to predict the current working conditions-objective function optimization-state feedback and enter the next step of vehicle speed and number of occupants prediction, then objective function optimization, state feedback, loop in turn, complete the online comprehensive predictive energy management of the vehicle's entire driving route.

Compared with the prior art, the comprehensive predictive energy management method for hybrid electric vehicles provided by the present invention has at least the following advantages:

(1) Through comprehensive prediction of the future short-term vehicle speed and the number of occupants, the disturbance of the power demand of the real application scenarios of the vehicle is considered, and the future short-term power demand of the vehicle is more truly and comprehensively reflected. It can not only be applied online in real time, but also better Allocate the power output of each power source of the power system to increase the energy-saving potential.

(2) Using deep learning based on a deep neural network framework to predict vehicle speed and passenger numbers, it can mine the deep potential relationship between historical operating condition information and future operating condition information in complex and changeable vehicle driving conditions, and improve prediction accuracy , It can more accurately predict the short-term power demand of the whole vehicle in the future, and provide a strong guarantee for the comprehensive forecast energy management of the whole vehicle.

Description of the drawings

Fig. 1 is a schematic flow chart of the method provided by the present invention.

Detailed ways

The technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work shall fall within the protection scope of the present invention.

As shown in Figure 1, in a preferred embodiment of the present invention, the following steps are specifically performed:

S01. Collect vehicle driving data and occupant data, and establish a database of vehicle historical driving conditions and a vehicle model;

Vehicle driving data collection is mainly by installing an inertial navigation system on a hybrid electric vehicle that needs to collect vehicle driving data to collect vehicle speed, acceleration, slope, steering, geographic location (latitude, longitude and altitude), driving date, etc. ; Or read the speed, acceleration, slope, steering, geographic location (latitude and longitude and altitude), driving date and other data on the CAN bus through Kavaser, but the speed, acceleration, slope and steering on the CAN bus read through Kavaser are the system The estimated vehicle speed, acceleration, slope, steering, and some deviations from the actual vehicle speed can still be used.

The occupant data collection is mainly through the real-time recording of the number, time, station, etc. of the occupants getting on and off the vehicle during the driving of the hybrid electric vehicle, and calculating the number of on-board passengers at each station based on the number of occupants getting on and off the vehicle; or recording through the on-board video system Extract time, site or geographic location, number of passengers on board, etc., and mark the recorded time as peak hours on weekdays, normal hours on weekdays, low hours on weekdays, and weekends.

S02, based on the deep neural network to predict the future short-term vehicle speed and the number of occupants; through the occupant data to establish the relationship between the occupant quality and energy consumption and the final energy management and distribution;

The determination of the structure of the established deep neural network predictor includes the number of network layers, the number of nodes, activation functions, etc.; the number of network layers includes input layers, n hidden layers, and output layers. The greater the number of hidden layers n, the depth of the deep neural network The deeper, n is determined by the training effect; the number of nodes is also determined by the training effect; activation functions commonly used include SIGMOID, TANH, RELU, LEAKY RELU, etc., among which RELU is the most widely used, because nodes with activation values equal to zero do not participate in the reverse Propagation will not cause the problem of slow training speed, and has the effect of sparse network; at the same time, adjust the number of iterations appropriately according to the training effect. If the number of iterations is too few, the algorithm is easily not suitable, and if the number of iterations is too much, the algorithm is likely to overfit.

The steps to train a deep neural network are as follows:

(1) Filter multi-dimensional historical data to eliminate unreasonable data noise caused by collection errors;

(2) Normalized multi-dimensional data (such as historical speed, instantaneous acceleration, slope, date and geographic location);

(3) Construct a neural network framework, including the number of network layers, the number of nodes, activation functions (hidden layer ReLU, output layer sigmoid), loss function (binary_crossentropy), optimizer (Adam), etc.;

(4) Send the sample data to the deep learning network to obtain the neuron activation value and actual output error required for parameter update;

(5) Top-down supervised learning, that is, using the back propagation of the supervised learning error, the parameters of each layer are updated from top to bottom, and the error of output and input is further reduced;

(6) Complete the network training, and use the test data to test the prediction results of the network. If the prediction result meets the requirements, the trained network can be used for prediction. If the requirements are not met, the network can be returned to the third step to adjust the network structure parameters;

(7) Denormalize the forecast data and obtain the forecast result.

The above-mentioned specific implementation of predicting the future short-term vehicle speed based on the trained deep neural network is as follows:

Input_v={V _tk ,V _t-k+1 ,...,V _t }

Or Input_v={V _tk ,V _t-k+1 ,...,V _t ,S _tk ,S _t-k+1 ,...,S _t }

Or Input_v={V _tk ,V _t-k+1 ,...,V _t ,D _tk ,D _t-k+1 ,...,D _t }

Or Input_v={V _tk ,V _t-k+1 ,...,V _t ,G _tk ,G _t-k+1 ,...,G _t }

Or Input_v=(V _tk ,V _t-k+1 ,...,V _t ,S _tk ,S _t-k+1 ,...,S _t ,D _tk ,D _t-k+1 ,... .,D _t }

Or Input_v=(V _tk ,V _t-k+1 ,...,V _t ,S _tk ,S _t-k+1 ,...,S _t ,G _tk ,G _t-k+1 ,... .,G _t }

Or Input_v=(V _tk ,V _t-k+1 ,...,V _t ,D _tk ,D _t-k+1 ,...,D _t ,G _tk ,G _t-k+1 ,... .,G _t }

or

Input_v={V _tk ,V _t-k+1 ,...,V _t ,S _tk ,S _t-k+1 ,...,S _t ,D _tk ,D _t-k+1 ,... ,D _t ,G _tk ,G _t-k+1 ,...,G _t }

Where G=(Lg _tk ,Lg _t-k+1 ,...,Lg _t ,La _tk ,La _t-k+1 ,...,La _t ,A _tk ,A _t-k+1 ,... .,A _t }

Output data={V _t+1 ,V _t+2 ,...,V _t+H }

Among them, V is the vehicle speed; S is the steering; D is the date of travel; G is the geographic location; Lg is the longitude; La is the latitude, A is the altitude; t is the current moment; k is the past k time period; H is the predicted time domain.

The above-mentioned specific implementation method for predicting the number of short-term occupants in the future based on a trained deep neural network is as follows:

Input_N=(B _tk ,B _t-k+1 ,...,B _t ,N _tk ,N _t-k+1 ,...,N _t ,W _tk ,W _t-k+1 ,. ..,W _t }

Output data={N _t+1 ,N _t+2 ,...,N _t+H }

Among them, B is the station; N is the number of occupants; W is the time period; t is the current moment; k is the past k time period; H is the prediction time domain.

S03: Regard the hybrid electric vehicle as a multi-constrained nonlinear optimization problem, select the state and control variables to solve the problem, and construct an optimization objective function, aiming at the optimal energy consumption of the entire vehicle;

In this embodiment, a dynamic programming algorithm is used to obtain the best energy consumption and economic performance of the entire vehicle. Take the dual-motor coaxial hybrid hybrid configuration as an example, assuming that x is the state variable, u is the control variable, d is the system disturbance, and y is the output, then there is

y=g(x,u,d)

和y的两个方程构成状态空间表达式，

表达的是状态变量x的一阶求导；控制量u如上式所述；扰动量d为 车速和乘员数量；T _eng、T _mot、T _ISG分别是发动机、电动机和启动电机的转矩；ω _eng、ω _mot分别是发动机、电动机的转速；P _bat是动力电池的功率；

是单位时间燃油消耗量，clutch＝0表示模式离合器断开，clutch＝1表示模式离合器接合；当扰动量车速和乘员数量由所述的基于深度神经网络的预测器预测出来时，构建成本函数J _fuel；T是预测时域的终止时间，O(t)是发动机启停次数，λ是惩罚函数。同时，整车动力系统需要满足动力电池、发动机、起动机、电动机的物理约束。 Among them, the state quantity x is the state of charge (SOC) of the power battery,

The two equations of and y form the state space expression,

It expresses the first-order derivation of the state variable x; the control quantity u is as described in the above formula; the disturbance quantity d is the vehicle speed and the number of occupants; T _eng , T _mot , and T _ISG are the torques of the engine, electric motor and starter motor, respectively; _eng and ω _mot are the speed of the engine and the motor respectively; P _bat is the power of the power battery;

Is the fuel consumption per unit time, clutch=0 means the mode clutch is disconnected, and clutch=1 means the mode clutch is engaged; when the disturbance amount and the number of occupants are predicted by the predictor based on the deep neural network, the cost function J is constructed _fuel ; T is the end time of the prediction time domain, O(t) is the number of engine starts and stops, and λ is the penalty function. At the same time, the vehicle power system needs to meet the physical constraints of the power battery, engine, starter, and electric motor.

S04: Under the framework of the model predictive control method, use the future short-term vehicle speed and the number of occupants predicted in step S02, according to the optimization objective function, global optimization in the prediction time domain, obtain the optimal control sequence, and perform optimal control The first control command of the sequence is input to the vehicle model for execution, so that the vehicle state quantity is updated and fed back to the next time for calculation;

The aforementioned predictive energy management steps based on the model predictive control architecture are as follows:

(1) Forecast of future short-term working conditions. According to the driving characteristics of the vehicle, use deep learning algorithms to predict and estimate the driving conditions of the vehicle in the future control time domain;

(2) Search for the best. Import the predicted short-term operating conditions into the vehicle model, and use dynamic programming algorithms to optimally solve the control problems in the future control time domain according to the optimization objective function;

(3) The first step is valid. Because there is a certain error in the prediction results of the working conditions in the future time domain, the control strategy calculated in step 2) also has a certain error. Therefore, in the model predictive control, only the control sequence solved in the first step is applied to the controller and the vehicle model actuating parts. The calculation results from the second step to the last step are mainly used to modify the calculation results of the first step to improve its optimal performance;

(4) Feedback loop. According to the operation results of the control sequence calculated in the first step in the controller and the vehicle model actuation components, the system state is fed back to the controller, and the prediction model re-predicts based on the error between the actual output and the predicted result, and the current state of the vehicle In the next time domain, return to step (1) to start a new model predictive control cycle.

S05, repeat the process of steps S01 to S04, and finally complete the comprehensive predictive energy management of the entire driving condition of the vehicle.

It can be seen from the above principles and implementation steps that the method provided by the present invention uses a combination of vehicle speed prediction and occupant prediction to more comprehensively consider the future short-term power demand in the real application scenarios of the vehicle; adopting the method based on model predictive control can be Realize the real-time online application of vehicle energy management. In summary, a comprehensive predictive energy management method for hybrid electric vehicles proposed above can realize real-time online energy management of vehicles with more comprehensive forecasting conditions and more accurate forecasting accuracy, and improve the energy-saving of hybrid electric vehicles. potential. The proposed method is easy to be implemented online and in real time, and has good engineering application prospects.

Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, and substitutions can be made to these embodiments without departing from the principle and spirit of the present invention. And variations, the scope of the present invention is defined by the appended claims and their equivalents.

Claims (6)

A comprehensive predictive energy management method for hybrid electric vehicles is characterized in that it specifically includes the following steps: S01. Collect vehicle driving data and occupant data, and establish a database of vehicle historical driving conditions and a vehicle model; S02. Forecast the future short-term vehicle speed and the number of future short-term occupants based on the deep neural network; S03. Regard the hybrid electric vehicle as a multi-constrained nonlinear optimization problem, select the state quantity and control quantity to solve the problem, and construct an optimization objective function, aiming at the optimal energy consumption of the whole vehicle; S04. Based on the vehicle model, using the future short-term vehicle speed and number of occupants predicted in step S02, according to the optimization objective function, global optimization in the prediction time domain is used to obtain the optimal control sequence, and the first of the optimal control sequence A control instruction is input to the vehicle model for execution, so that the vehicle state quantity is updated and fed back to the next time for calculation; S05. Repeat the process of steps S01S04 to finally complete the comprehensive predictive energy management of the entire vehicle driving condition. The method according to claim 1, characterized in that: the vehicle travel data collected in step S01 includes: vehicle speed, acceleration, slope, steering, geographic location, travel date, and driving route; and the occupant data includes : The number of occupants, the location of the stop, and the time. The time is marked as the peak period of the working day, the normal period of the working day, the low period of the working day and the weekend. The method according to claim 2, characterized in that: step S02 specifically includes establishing a deep neural network predictor to perform the prediction, and the input of the predictor is the collected vehicle driving data and occupant data; the predicted future The short-term vehicle speed is a short-time scale prediction, which is made in seconds, and the predicted number of short-term occupants in the future is a long-term scale prediction, which is made in the unit of the stop station location. The method according to claim 1, wherein step S03 specifically includes establishing an optimizer for the multi-constrained nonlinear optimization problem, and the optimization goal is the optimal energy consumption of the entire vehicle in the short-term prediction time domain. The method according to claim 1, characterized in that: step S04 specifically includes the comprehensive predictive energy management of the hybrid electric vehicle under the framework of the model predictive control method to perform working condition prediction and objective function in the rolling forward time domain. Optimization and state feedback, each step starts at the current time, predicts the vehicle speed and the number of passengers in the H time domain in the future, and inputs the predicted vehicle speed and the number of passengers in the H time domain to the optimizer for global optimization in the H time domain. The method according to claim 1, wherein the multi-constraint nonlinear optimization problem in step S03 is solved based on a dynamic programming algorithm.

Images