Saving Fuel Without Draining the Battery: A New AI Approach to Hybrid Car Management

Researchers developed an artificial intelligence strategy that checks the likely immediate consequences of a power decision before putting it into action. Their approach combines learning from experience with predictions of battery charge and temperature.

A hybrid car makes countless small decisions about when to use its engine and when to draw power from its battery. Saving fuel depends on getting those decisions right, with battery temperature, remaining charge and engine switching all adding to the challenge. A choice that looks efficient for a few seconds can leave the battery hotter or more depleted by the end of a journey.

Research published in Energies, titled "A Dual-Evaluation Collaborative Decision Deep Reinforcement Learning Strategy for Multi-Constraint Energy Management of Hybrid Electric Vehicles," explores a way to manage these competing demands. Quanhong Zeng and colleagues developed an artificial intelligence strategy that checks the likely immediate consequences of a power decision before putting it into action. Their approach combines learning from experience with predictions of battery charge and temperature.

The researchers tested their system through computer simulations covering urban driving, higher starting temperatures and sustained highway loads. In the normal-temperature comparison, it reduced fuel consumption by 33.58% against a conventional deep Q-network controller, a type of reinforcement-learning system. Its broader contribution was a more balanced relationship between fuel savings, battery conditions and engine operation, with different strengths emerging across the tests.

Giving Each Power Decision a Second Check

Many learning-based controllers choose actions using a score that estimates their longer-term benefits, with unwanted battery conditions reflected in penalties after an action occurs. This leaves the controller without a direct comparison of the immediate battery consequences of its available choices. The new strategy, called RAEP-KAN-DQN, brings that comparison into the decision itself.

For each available engine-power choice, the system maps the command to an executable action and predicts the battery's next charge level, internal temperature and surface temperature. A combined electrical and thermal model links battery current, heat generation and temperature-dependent electrical behaviour. The controller weighs these short-term risks against the longer-term value learned from previous experience.

The learning network uses Kolmogorov–Arnold Networks, or KAN, to represent complex relationships within the vehicle's energy system. A "dueling" structure separates the value of the current situation from the relative benefit of each possible action, helping distinguish choices with similar outcomes. Together, these components guide power sharing between the engine and battery in a simulated series-parallel hybrid containing an engine, motor, generator and lithium-ion battery.

Fuel Savings That Do Not Depend on Draining the Battery

The study compared the complete strategy with conventional DQN and two versions containing only part of the proposed improvements. Testing these alternatives helped reveal the separate contributions of the revised network and the predictive risk checks. Across three random training seeds, the complete system reached a stronger, more consistent learning performance and showed less variation between runs.

In the dedicated 25°C comparison, the new controller used 2.395 litres per 100 kilometres, compared with 3.606 litres for conventional DQN. Peak battery core and surface temperatures reached about 27.98°C and 27.67°C, with the paper reporting reductions of 11.73% and 10.99% against DQN. The battery's final charge remained close to the 60% target, with a reported charge deviation of 0.002.

The separate results averaged across three seeds recorded consumption of 2.372 litres per 100 kilometres and a final battery charge of 59.7%. The risk-aware controller without KAN consumed less fuel, at 2.040 litres per 100 kilometres, but finished with only 57.5% charge, showing how greater battery use can make fuel figures look better. The complete strategy averaged about 47 engine starts, compared with roughly 91 for KAN-DQN and 82 for conventional DQN, giving it a useful balance between charge preservation and engine switching.

Hot Batteries and Highway Driving Change the Demands

Raising the starting temperature to 40°C gave the researchers a tougher test of heat management. The new strategy limited core and surface temperature rises to approximately 2.67°C and 2.47°C, around 62% below the increases recorded for conventional DQN. Fuel consumption reached 2.069 litres per 100 kilometres, a 13.61% reduction from its own 25°C result, with a larger final charge deviation of 0.020.

These results show reduced heat accumulation under the simulated conditions, rather than a guarantee that battery temperatures stay below every reference limit. The battery surface exceeded the study's 40°C reference, and another controller maintained charge more closely in the hot-start test. The proposed method's strength was its combination of relatively low fuel consumption and restrained temperature growth.

Highway testing used the HWFET driving cycle, which places more sustained demands on the powertrain. The new controller consumed 3.439 litres per 100 kilometres, 15.42% less than KAN-DQN, but more than conventional DQN and the risk-aware version without KAN. Its charge deviation of 0.111 was 54.32% below DQN's, showing that the lower fuel consumption of some alternatives came with greater battery depletion.

The highway strategy favoured longer periods of moderate engine output, replacing the shorter interventions used during urban driving. It averaged about 26 engine starts, roughly 41% fewer than conventional DQN, and kept peak core and surface temperatures below KAN-DQN's levels. Highway charge deviation remained substantial, making the result a practical trade-off rather than a win on every measure.

What Needs to Happen Before Real Cars Can Use It

The researchers tested how strongly the controller should weigh immediate risk against learned future value. Moderate settings offered a useful balance, with a risk weight of four delivering the best combined reward in the sensitivity analysis. Larger weights sometimes lowered fuel use by drawing more energy from the battery, showing that increasing caution in the scoring formula does not automatically improve every outcome.

Computing each decision took an average of about 7.5 milliseconds in the tested environment, compared with roughly 0.1 milliseconds for conventional DQN. The slowest observed decision took about 24.7 milliseconds, and performance on automotive hardware remains unverified. Planned work includes hardware-in-the-loop or vehicle testing, checks involving sensor noise and execution delays, and changes to reduce computing demands.

The study offers evidence that checking battery consequences before choosing an action can improve how a hybrid shares its energy. Its simulations do not establish real-world fuel savings, longer battery life or guaranteed thermal safety. For future vehicle development, the promising idea is a controller that treats fuel consumption, battery charge, heat and engine behaviour as parts of the same decision.