Project Overview

This is an early project about optimizing energy efficiency of hybrid electric vehicles with respect to an open-ended world of driving conditions.

The two key questions are:

• how do we represent a diverse set of driving conditions/profiles, and how can it possibly scale to open-ended scenarios?
• how to design the energy allocation so the system learns to adapt its strategy based on the driving condition it's experiencing?

To answer these questions, we:

• Designed a VAE-based driving cycle generative model that performed kinematic feature extraction and resampling on large-scale real vehicle driving condition segments, achieved low-cost simulation of diverse driving conditions (both normal and extreme scenarios).
• Designed a TCN-based driving condition forecasting model, achieving a 23% improvement in forecast accuracy and enabling proactive energy allocation decisions.
• Developed a distributed, driving condition-aware agent adaptive switching mechanism that dynamically selects between specialised agents based on real-time driving scenarios (urban/highway/congested traffic), achieving a 29% reduction in vehicle energy consumption under real stochastic driving conditions.

From a more recent perspective, the methods used in this project seemed primitive, but the core ideas of modelling of long-tail scenarios using a limited set of real-world data, adaptive learning, still shred light on what we are excited to explore in the years to come.