Application of machine learning for optimizing energy management strategy

Abstract

The increase of CO2 emissions impacts greatly the climate change and causes manyconcerns for the future of our planet. One way to counteract this development isby reducing the fuel consumption of passenger vehicles. Hybrid Electric Vehicles (HEVs) combine advantages of both electric and conventional (combustion) vehicles, and due to their high appeal they represent an important milestone in the shift from fossil fuels to clean, regenerative energies. As HEVs continue to use fossil fuel as one of the energy sources, there is a need to optimize the energy consumption. Fuel economy and emission reduction in HEV rely strongly on the energy management strategy. The strategy aims at minimizing the vehicles fuel consumption by controlling the flow between the internal combustion engine and the electric system. The two main solutions for this problem are Dynamic Programming (DP)and the Equivalent Consumption Minimization Strategy (ECMS), both of which require the knowledge of the driving situation beforehand, in order to calculate the optimal operating points for the two power supplies. This thesis proposes Machine Learning (ML) for optimizing the strategy by recognizing the driving situation through available data such as vehicle speed, torque and battery State of Charge, with the aim of improving fuel economy and reducing computation time. In particular, the proposed strategy combines ECMS with a machine learning algorithm which yields an adaptable cost factor needed for ECMS. This cost factor is updated periodically and is determined only by analyzing past and present data.The ML models researched for this task were Linear Regression (LR), Support Vector Regression (SVR) and a Neural Network (NN). During modeling of the cost factor both the charge sustaining mode as well as the energy optimisation were taken into account. The selection of the best ML algorithm was made by comparing them against each other. Further, in order to verify the method of the solution, the ECMS with the adaptable cost factor from our solution was compared against the ECMS with a standard cost factor constant. The evaluation was made in regard of the fulfilment of the battery charge sustaining as well as the fuel consumption. In order to justify the difference in the battery state of charge at the end of a mission, a fuel post-correction was applied.The solution proposal with the NN shows an overall decrease in fuel consumption by an average of ≈ 2.5%, while showing better charge sustaining results compared to the constant cost factor. This results show the great potential of utilizing machine learning for reducing CO2 emissions and optimizing the energy- management strategy, when combined with big data.