Page 830 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

(continued)

Item Details
Model Training LSTM (Long-Short Term Memory) and Transformer-based models trained
and Fine-Tuning on historical driving, road gradient, and battery usage data for energy fore-
casting; RL models continuously fine-tuned using real-time feedback and
usage logs.
Testbeds or Pilot Not yet deployed; simulated driving environment and historical driving data-
Deployments sets to be used in initial prototyping phase.
Code reposito- N/A (To be published upon completion of PoC and initial prototype)
ries

2 Use Case Description

2�1 Description

Context & Background:
• Problem with existing solutions: The existing solutions/Apps could tell us the route and
traffic, and cars shows us the battery level or speed limits — but none of these systems
tell us how to drive in advance to maximize our battery, reduce wear, and improve safety
based on the upcoming terrain, our battery's health, and our own driving behavior.
They’re all reactive dashboards — not predictive advisors.
• What we are doing differently: Our system takes real-time driver input, vehicle battery
condition, and terrain elevation data, and uses AI models (TST + Mistral) to predict: Ideal
speed and RPM for the next 5 km, based on battery health, terrain slope, driver behavior,
and route curvature. It presents it as a real-time, actionable advisory for the driver.

This isn’t just a map. It’s a driver- and EV-aware AI assistant that evolves and adapts across
driving sessions. Think of it like Cruise Control on Steroids, but made specifically for your
country roads, tier-2/tier-3 conditions, and personalized EV optimization.

• Real-World Example: Let’s say you’re entering a steep hill for 2 km. Your battery is
overheating, and you tend to brake aggressively. Our system will say: 'Reduce RPM to
1600, hold at 40 km/h for the next 3 km to maintain battery life and avoid overheating.
Regenerative braking will kick in on the downhill. No other system tells you this, this early,
or this personally.
Objectives & Aim: We aim to develop an AI-driven real-time endurance estimation framework
in autonomous vehicles.

• To enhance route optimization by combining route, elevation (OpenTopography), and
curvature every 5 km with integration of deep learning models with real-time contextual
data streams, including road topology, traffic conditions, environmental parameters, and
driver-specific behavioural patterns.
• To leverage time-series forecasting models (LSTM/Transformer) such as the Chronus TST
model for time-series prediction of RPM/speed, where the system continuously learns
from historical telemetry and energy consumption data to provide accurate, dynamic
predictions of battery usage.
• To incorporate Reinforcement Learning (RL) with Mistral LLM (via Ollama) to convert
model output into driver instructions to deliver adaptive, personalized recommendations
for energy-efficient driving behaviour.

794