Research On Hybrid System Energy Management Strategy Based On LSTM Recurrent Neural Network

Curbing carbon emissions to combat climate change is now a global consensus,and globally,surface transport is responsible for about 20%of total carbon dioxide emissions.In order to improve energy utilization efficiency and reduce carbon emissions,various new energy vehicles such as hybrid power,pure electricity and hydrogen fuel cell develop rapidly.Energy conversion and distribution between power sources require reasonable energy management,and effective energy management strategies can give full play to the energy-saving potential of new energy vehicles.At present,the scheme of using recurrent neural network as the core of energy management strategy of hybrid power system has not been studied deeply.Therefore,this paper takes the vehicle model equipped with power split hybrid system as the object,adopts the method of combining algorithm simulation and controller-in-loop test,and explores the energy management strategy scheme of power split hybrid system based on long short-term memory recurrent neural network（LSTM）.Firstly,according to the characteristic data of each part of the prototype vehicle and the mechanical and electrical relationship,the whole vehicle model of a passenger vehicle and a commercial vehicle was established respectively by backward method.The vehicle model includes vehicle dynamics model,engine model,motor model,battery model and transmission model.The difference between passenger vehicle model and commercial vehicle model is that the two dynamic models adopt conventional driving resistance method and coasting test method respectively,and the component specifications and topological structure of the power system are slightly different,but both of them are essentially power-split hybrid electric vehicles.The purpose of setting up two vehicle models respectively is to fully verify the universality of subsequent algorithms.In addition,the characteristics of commonly used standard driving cycles are compared and analyzed.The vehicle model and driving cycle above are used as the controlled object and driving scenario of the subsequent algorithm.Secondly,in order to obtain the theoretical optimal energy management scheme,the dynamic programming（DP）algorithm is used to solve the optimal energy management problem of the hybrid system.The state space of the minimum energy consumption problem of the hybrid electric vehicle model is established,the degree of discretization of the state space is selected,and the objective function is defined.By analyzing the fuel consumption and engine and motor operating characteristics of the optimal control scheme under each cycle,it is proved that the DP algorithm has strong universality for each driving condition,and the maximum difference of fuel consumption rate between the four driving cycles is only 1.8%.In order to avoid the defect that DP algorithm is unable to carry out real-time control,DP algorithm is used to provide a training set scheme for LSTM cyclic neural network in this paper.At the same time,after the LSTM neural network energy management strategy is implemented on the given driving cycle and vehicle model,DP algorithm can calculate the optimal energy management scheme and compare with the results of LSTM.Then,an energy management strategy based on LSTM neural network was constructed.The optimal control data obtained by DP algorithm was used as the training set to train the network,and the trained LSTM network was used as the core of the energy management strategy to control the vehicle model in real time.The results show that,in the two test driving cycles of CHTC-LT and C-WTVC,LSTM shows good generalization ability,SOC fluctuation is relatively stable,and the difference between the beginning and the end is less than 10%,and LSTM scheme can achieve a minimum fuel consumption gap of 13.6%compared with the theoretical optimal scheme given by DP algorithm.By comparison of control methods,LSTM and DP have certain similarities in control logic,and the minimum difference of average speed and average torque between engine and motor is 14.5%.Finally,based on the controller-in-loop test,the ARM cortex M7 processor was used as the control core and CAN bus was used to communicate with the virtual vehicle model in Simulink through serial port.The online application capability of the energy management strategy based on LSTM neural network was verified.The calculation time statistics show that the three energy management strategies of LSTM layer scale can meet the real-time requirements.In addition,several examples were calculated to investigate the effects of LSTM layer size,training iterations,initialη_bat and training set on the control effect of energy management strategies.Then the average working efficiency and fuel consumption of engine,motor and vehicle are analyzed.The main conclusions are as follows:LSTM neural network model is insensitive to the type of driving cycle in the training set;The optimal training amount is 300 to 500 iterations of the optimal energy management scheme for a single driving cycle.Too low number of iterations will lead to the decrease of the stability of neural network model for SOC.The same model has similar efficiency arrangement for each component under different test driving cycles.Relative to the optimal LSTM layer size and training iterations of 128 layers and 500,respectively,the model using this configuration has a fuel economy advantage ranging from 3%to 10%compared with other configurations.