Employing Deep Continuous Reinforcement Learning Based Control to Optimize the Regenerative Braking Process in Electric Vehicles: Employing Deep Continuous Reinforcement Learning Based Control to Optimize the Regenerative Braking Process in Electric Vehicles
Authors
Kraft, Simon Bonderup ; Dalsgard, Filip ; Johansen, Simon
Term
4. term
Education
Publication year
2026
Submitted on
2026-05-29
Abstract
This thesis explores how deep, continuous reinforcement learning (RL) can be used to control regenerative braking (RB) in electric vehicles. Regenerative braking means that the car uses braking force to recover energy and recharge the battery, instead of relying only on mechanical brakes. Traditional rule-based controllers struggle to meet three conflicting goals at the same time: providing a comfortable driving experience, maximizing energy recovery, and protecting battery lifetime. This work aims to find a better balance between these goals. A simulation environment representing a 2016 Nissan Leaf was built in MATLAB/Simulink. Within this environment, two RL methods, TD3 and SAC, were trained. The RL agents used carefully designed reward functions (known as reward shaping) that include safety, comfort, energy recovery, and long-term battery health. RL proved particularly effective at handling constraints and optimizing several objectives at once. For example, a TD3 agent achieved a 15.5% increase in driving range compared with a similar constant braking strategy. By adjusting the reward function, the RL agent could be guided to prioritize driver comfort, energy recovery, or a compromise between the two. To test whether the approach also works outside simulation, a physical motor-generator test bench was built and modelled. A new RL agent was trained on this test bench model. The measured behaviour of the physical system matched the simulated behaviour closely, showing that RL-based control transfers reliably to real hardware. Based on these results, the thesis concludes that reinforcement learning is a viable and promising control strategy for regenerative braking in electric vehicles. The method can find satisfactory compromises between comfort, energy recovery, and battery lifetime in both simulated and practical experiments.
Denne specialeopgave undersøger, hvordan dyb, kontinuerlig reinforcement learning (RL) kan bruges til at styre den regenerative bremsning (RB) i elbiler. Regenerativ bremsning betyder, at bilen bruger bremsekraften til at genvinde energi og lade batteriet, i stedet for kun at bruge mekaniske bremser. Traditionelle, regelbaserede styresystemer har svært ved samtidig at opfylde tre modstridende mål: at give føreren en behagelig køreoplevelse, at maksimere energigenvindingen og at beskytte batteriets levetid. Denne afhandling forsøger at finde en bedre balance mellem disse mål. Der blev opbygget et simuleringsmiljø i MATLAB/Simulink, som modellerer en Nissan Leaf fra 2016. I dette miljø blev to RL-metoder, TD3 og SAC, trænet. RL-agenternes belønningsfunktioner blev omhyggeligt designet (såkaldt reward shaping) til at tage højde for sikkerhed, komfort, energigenvinding og batteriets langsigtede helbred. RL viste sig særligt velegnet til at håndtere begrænsninger og til at optimere flere mål på én gang. En TD3-agent opnåede fx 15,5 % længere rækkevidde end en sammenlignelig strategi med konstant bremsning. Ved at ændre belønningsfunktionen kunne man styre, om RL-agenten lagde mest vægt på komfort, på energigenvinding eller på et kompromis mellem de to. For at teste om metoden også virker uden for simuleringen blev der opbygget en fysisk motor-generator testbænk, som også blev modelleret. En ny RL-agent blev trænet på denne testbænksmodel. Målinger på det fysiske system stemte godt overens med simulationerne, hvilket viser, at RL-baseret styring kan overføres pålideligt til rigtig hardware. På baggrund af disse resultater konkluderes det, at reinforcement learning er en lovende og anvendelig kontrolstrategi til regenerativ bremsning i elbiler. Metoden kan finde tilfredsstillende kompromiser mellem komfort, energigenvinding og batteriets levetid, både i simulerede forsøg og i praktiske eksperimenter.
[This abstract has been rewritten with the help of AI based on the project's original abstract]