AC Advisor — IEEE-Published ML System for Automobile A/C Energy Prediction

IEEE Publication: AC Advisor: A Comparative Evaluation of ML and State Space Models for Automobile Air Conditioning Energy Prediction — Read on IEEE Xplore →

A production ML system that predicts automobile A/C energy consumption from OBD-II telemetry data, with SHAP explainability and an interactive Streamlit "What-If" coach.

---

Live Demo

→ Try the live Streamlit app

Upload or simulate a drive, adjust A/C parameters, and get real-time energy predictions with SHAP feature explanations.

---

Research Summary

This project was accepted and published at IEEE in 2026. The study addresses a gap in automotive A/C energy modelling: most prior work uses physics-based simulations that don't generalize across vehicle types and real-world drive conditions.

Dataset: 37,000+ OBD-II telemetry records collected across urban, highway, and mixed driving cycles.

Key Finding: CatBoost with hyperparameter tuning achieved the best balance of accuracy and inference speed for production deployment (MAPE < 10%), while the Transformer model showed competitive accuracy but impractical latency for real-time use.

---

Model Comparison Results

Model	MAPE (%)	RMSE	R²	Inference Time
CatBoost (tuned)	7.4%	0.31	0.94	2ms
XGBoost	8.1%	0.34	0.93	1ms
Random Forest	9.2%	0.39	0.91	4ms
Linear Regression	14.7%	0.61	0.79	<1ms
LSTM	8.8%	0.37	0.92	18ms
Transformer	7.9%	0.33	0.93	47ms

Winner: CatBoost — MAPE < 10% threshold met, best R², practical inference time for real-time edge deployment.

---

Architecture

OBD-II Telemetry (37K+ records)
        │
        ▼
Feature Engineering (features.py)
  ├── Vehicle speed, RPM, throttle position
  ├── Ambient temp, cabin setpoint, compressor load
  ├── Rolling window aggregates (5s, 30s, 60s)
  └── Derived: ΔT (cabin vs ambient), load factor
        │
        ▼
Model Training & Comparison
  ├── CatBoost  ← production model
  ├── XGBoost
  ├── Random Forest
  ├── Linear Regression (baseline)
  ├── LSTM
  └── Transformer
        │
        ▼
SHAP Explainability
  ├── Global feature importance
  ├── Per-prediction SHAP waterfall
  └── Interaction effects (speed × ΔT)
        │
        ▼
Streamlit App (What-If Coach)
  ├── Replay historical drives
  ├── Simulate A/C setting changes
  └── Real-time SHAP explanations

---

SHAP Feature Importance

Top features driving A/C energy consumption (from SHAP analysis across 37K records):

Rank	Feature	Contribution
1	ΔT (cabin setpoint − ambient)	34%
2	Vehicle speed (rolling 30s avg)	21%
3	Compressor duty cycle	17%
4	Engine RPM	11%
5	Solar load proxy	9%
6	Cabin volume (vehicle class)	8%

Key insight: The temperature differential (ΔT) is by far the dominant driver. Reducing setpoint by 2°C during highway driving reduces predicted energy draw by ~18%.

---

Quickstart

Option A — Conda (recommended)

conda env create -f environment.yml
conda activate ac-advisor
streamlit run app.py

Option B — Docker

docker build -t ac-advisor .
docker run --rm -p 8501:8501 ac-advisor

Option C — pip

pip install -r requirements.txt
streamlit run app.py

---

Project Structure

Ac-advisor-ieee/
├── app.py                    # Streamlit UI — replay, what-if, SHAP coach
├── ac_advisor/
│   ├── features.py           # Feature engineering — single source of truth
│   ├── predictor.py          # Model loading, predict_now(), predict_sim()
│   ├── comfort.py            # ΔT comfort bands and scoring
│   └── coach.py              # Nearest-context recall and action logging
├── models/
│   └── catboost_model.cbm    # Trained CatBoost model
├── data/
│   └── EnergyPredictionDataset_ReadyForModel.csv
├── assets/                   # SHAP plots, architecture diagrams
├── Dockerfile
├── requirements.txt
└── README.md

---

Key Design Decisions

Why CatBoost over XGBoost? CatBoost handles categorical features (vehicle class, drive mode) natively without manual encoding, and showed better calibration on the tail of the distribution — exactly where dangerous high-load predictions occur.

Why heuristic features over raw signals? Raw OBD-II signals are noisy at 10Hz. Rolling window aggregates (5s, 30s, 60s) capture the thermal inertia of the A/C system and reduce prediction variance by ~30%.

Why Streamlit over FastAPI for the demo? The target users (automotive engineers, fleet operators) need interactive what-if exploration, not programmatic API access. Streamlit enables zero-install deployment for non-technical stakeholders.

---

Citation

If you use this work, please cite:

@inproceedings{lokam2026acadvisor,
  title     = {AC Advisor: A Comparative Evaluation of ML and State Space Models 
               for Automobile Air Conditioning Energy Prediction},
  author    = {Lokam, Kiranmayee and Tulabandula, Hemanth Kumar},
  booktitle = {IEEE},
  year      = {2026},
  url       = {https://ieeexplore.ieee.org/document/11393777}
}