Adaptive Recursive Prediction of the Desired Torque of a Hybrid Powertrain

Abstract

Hybrid powertrain provides driving torque based upon the vehicle acceleration pedal position. Due to the slow response of the internal combustion (IC) engine, the electric motor(s) (EM) often has to provide additional torque during transient vehicle operations. During cold-start operations, the available battery power is significantly reduced due to cold temperatures, and due to the starting delay of the IC engine, the battery has to provide power at the level beyond its normal operation limit for the electrical motor to meet the torque demand, which would dramatically shorten the battery life. However, if the desired powertrain torque can be predicted, the IC engine could be started early when the predicted future powertrain power is higher than the battery normal operational range, and hence, the duration of battery operated over its normal operational range could be significantly reduced, leading to extended battery life. This makes predicting multistep future torque demand, based upon current and past torque demand data, an important subject. Amid many existing studies, few research studies were focused on this subject. This paper proposes an adaptive recursive prediction algorithm to estimate the future desired powertrain torque and compares the performance of the proposed algorithm with two existing prediction algorithms, i.e., step-by-step and fixed-gain algorithms. These three prediction algorithms are studied under the federal test procedure (FTP) and another four typical driving cycles. The simulation results show that the adaptive recursive prediction algorithm reduces the prediction error significantly with a 4% maximum prediction error, and in addition, the reduced computational load makes it feasible for real-time implementation. The introduced weighting factors for the past and current data are the key to improve prediction accuracy and avoid prediction calculation over- and underflow. In addition, due to the online update of regression gain, the proposed algorithm is also robust to driving behavior variations.