🌍 Carbon-Aware ML Training Controller — CAML-TC

---

⚡ Executive Summary

A reinforcement learning–driven carbon-aware scheduling system that dynamically optimises machine learning workloads using real-time and forecasted UK electricity grid carbon intensity.

📄 Peer-reviewed research — Accepted at IEEE EEEIC 2026 (Q1, Scopus-indexed, Web of Science)
🔗 Try the live app — uses live UK National Grid data, no login required
📦 pip install caml-tc — use as a Python library in your own ML pipeline

The system integrates:

• 🔮 Carbon forecasting with uncertainty-aware confidence weighting
• 🧠 Heuristic optimisation with multi-window holistic scoring
• 🤖 Risk-aware reinforcement learning — Q-learning MDP trained on 8,000 episodes of real UK grid data

to shift compute workloads into low-emission periods, achieving 28–34% CO₂ reduction under realistic conditions and >50% under optimal conditions — without changing a single line of model code.

---

💡 Problem Statement

The UK electricity grid fluctuates between 50 and 260 gCO₂/kWh within a single day. The same ML training job at 3am versus 7pm evening peak can produce 3–5× different carbon emissions.

Modern machine learning workloads are executed without any environmental awareness. This results in:

• ❌ Unnecessary CO₂ emissions from poorly timed execution
• ❌ No integration between carbon intelligence and scheduling decisions
• ❌ Lack of sustainability in ML pipelines despite available grid data
• ❌ Existing solutions (Google, Meta, Microsoft) are proprietary and inaccessible

Google has done this internally since 2020. Microsoft released a partial SDK. Nobody built an open, deployable version for individual researchers and engineers. Until now.

---

📊 Validated Results

Tested on real UK National Grid data. Emissions measured with CodeCarbon.

Strategy	Carbon Reduction	Conditions
Baseline (immediate execution)	0%	—
Heuristic Scheduler	~28%	Realistic UK grid
RL Scheduler (CAML-TC)	28–34%	Realistic UK grid
RL Scheduler (CAML-TC)	>50%	Optimal low-variability

RL outperforms heuristics under high grid variability — the conditions where static rules fail and adaptive learning matters most.

---

🚀 Install & Quickstart

pip install caml-tc

from camltc import CarbonScheduler

scheduler = CarbonScheduler(duration_minutes=90, urgency="low")
result = scheduler.recommend()

print(result.best_window)        # "03:00 — 04:30 UTC"
print(result.carbon_saving_pct)  # 31.4
print(result.strategy)           # "rl" or "heuristic"
print(result)                    # full formatted summary

Urgency levels

# Low — aggressive carbon optimisation, longer delay allowed
CarbonScheduler(duration_minutes=120, urgency="low").recommend()

# Medium — balanced (default)
CarbonScheduler(duration_minutes=60, urgency="medium").recommend()

# High — run as soon as a clean-enough window appears
CarbonScheduler(duration_minutes=30, urgency="high").recommend()

---

🧠 Key Innovations

• 🌱 Carbon intensity used as a first-class scheduling signal — not an afterthought
• 🤖 Hybrid scheduling architecture — heuristic for stable grids, RL for volatile ones
• ⚖️ Multi-objective RL reward design — jointly optimises carbon intensity, forecast uncertainty, delay penalty, and deadline constraints
• 📉 Exponential forecast confidence decay — wₜ = exp(−λ·Δt) — prevents over-commitment to unreliable long-horizon predictions
• 🎲 Stochastic noise injection — forces RL agent to learn policies robust to real grid variability, not just noise-free signals
• 📡 End-to-end closed-loop pipeline — from raw National Grid API to scheduling decision in one system
• 🧪 CodeCarbon-integrated emissions measurement — validated, reproducible results

---

⚙️ System Architecture

┌──────────────────────────────────────┐
│   UK National Grid Carbon API        │
│   Real-time + 24h Forecast           │
└──────────────────┬───────────────────┘
                   ↓
┌──────────────────────────────────────┐
│   Carbon Intelligence Layer          │
│   Confidence decay · Volatility      │
│   Peak/low detection · Uncertainty   │
└──────────────────┬───────────────────┘
                   ↓
┌──────────────────────────────────────┐
│   Scheduling Engine                  │
│   ├─ Heuristic Optimizer             │
│   │    Window scoring · Delay penalty│
│   └─ RL Policy Agent (Q-Learning MDP)│
│        State: (μ, σ, delay)          │
│        Reward: carbon+uncertainty    │
│               +delay+deadline        │
└──────────────────┬───────────────────┘
                   ↓
┌──────────────────────────────────────┐
│   ML Workload Simulator              │
│   Synthetic training load            │
│   CodeCarbon emissions tracking      │
└──────────────────┬───────────────────┘
                   ↓
┌──────────────────────────────────────┐
│   Analytics Dashboard (Streamlit)    │
│   Forecast Explorer · Simulation Lab │
│   Strategy Comparison · KPI View     │
└──────────────────────────────────────┘

---

✨ Key Features

• 📡 Real-time UK carbon intensity integration via National Grid API
• 🔮 24-hour carbon forecasting with uncertainty modelling
• 🧠 Heuristic + RL scheduling engine — automatically selects best strategy
• 🤖 Risk-aware RL agent trained on 8,000 episodes of real UK grid data
• 🧪 ML workload simulation with CodeCarbon emissions tracking
• 📊 Interactive multi-page Streamlit dashboard
• 📈 Carbon peak/low detection and optimal window identification
• ⚙️ Multi-strategy emissions comparison (baseline vs heuristic vs RL)
• 📦 pip-installable Python library for direct pipeline integration

---

🧪 System Modules

📈 core/carbon_api.py — Carbon Intelligence Layer

• Real-time and forecast carbon intensity from UK National Grid
• Exponential confidence weighting on future predictions
• Peak/low detection and optimal window identification

⚙️ core/scheduler.py — Heuristic Scheduling Engine

• Holistic window scoring (carbon intensity + uncertainty + delay penalty)
• Urgency-parameterised delay penalty coefficient
• Strong interpretable baseline that RL must genuinely beat

🤖 core/rl_agent.py — Reinforcement Learning Agent

• Q-learning with ε-greedy exploration and linear epsilon decay
• State: (μₜ, σₜ, dₜ) — mean CI, uncertainty proxy, accumulated delay
• Multi-objective reward: carbon + uncertainty + delay + deadline violation
• 8,000 training episodes · 4-phase iterative development

🧪 core/simulator.py — Emissions Measurement

• Synthetic ML workload generation
• CodeCarbon-based emissions tracking per strategy
• Reproducible experimental pipeline

📊 demo/ — Interactive Dashboard

• app.py — multi-page Streamlit application
• pages/forecast.py — 24h carbon intensity explorer
• pages/simulation.py — strategy comparison lab
• pages/overview.py — KPI and results summary

---

🧰 Tech Stack

• Python 3.9+
• Streamlit — interactive dashboard
• Plotly — visualisations
• Pandas / NumPy — data processing
• CodeCarbon — emissions measurement
• Custom RL Agent — Q-learning MDP
• UK Carbon Intensity API — api.carbonintensity.org.uk

---

🌍 Environmental Impact

CO₂ Saved	Real-World Equivalent
1 kg	🌳 ~0.045 trees planted
10 kg	🚗 ~40 km of driving avoided
100 kg	✈️ One short-haul flight offset

Example: A team running 10 GPU training jobs per week, each saving 28–34% emissions, saves hundreds of kg of CO₂ annually — with zero infrastructure changes.

---

🇬🇧 UK Net Zero Relevance

This system is directly aligned with the UK's Net Zero 2050 commitment. It uses the National Grid's own free public carbon intensity API — infrastructure the UK government has already built.

Data centres and AI infrastructure are among the fastest-growing electricity consumers in the UK. CAML-TC provides the scheduling intelligence layer that was missing — making carbon-aware ML accessible to any researcher or engineer, not just those inside Google or Microsoft.

---

🚀 Roadmap

• PyTorch training callback (on_epoch_end auto-scheduling)
• HuggingFace Trainer integration
• AWS / Azure / GCP cloud scheduler hooks
• Transformer-based CI forecasting (replacing LSTM baseline)
• Continuous-action RL — DQN, PPO
• Multi-region geographic shifting (EU / US grids)
• Carbon-aware CI/CD for ML training pipelines
• Kubernetes-based orchestration

---

📄 Research & Citation

Published at IEEE EEEIC 2026 — International Conference on Environment and Electrical Engineering (Q1, Scopus-indexed, Web of Science)

If you use CAML-TC in your work, please cite:

@inproceedings{rehman2026camltc,
  title     = {Adaptive Carbon-Aware Machine Learning Training under Uncertainty:
               A Unified Scheduling Framework},
  author    = {Rehman, Sufiyan Ul},
  booktitle = {2026 IEEE International Conference on Environment and
               Electrical Engineering (EEEIC)},
  year      = {2026},
  publisher = {IEEE}
}

---

👨‍💻 Author

Sufiyan Ul Rehman
AI/ML Researcher · Lecturer, Ulster University & Solent University (via QA Higher Education), London
Building intelligent, carbon-aware AI systems for sustainable machine learning infrastructure.

🔗 Live App · 📦 PyPI · 🔬 IEEE Paper

---

📌 Closing Statement

The future of machine learning is not only about accuracy and performance,
but also about when and how sustainably computation is executed.

The carbon cost of AI is real. The fix is simpler than anyone realises.
It is just a matter of timing.

---

📜 License

MIT © 2026 Sufiyan Ul Rehman — see LICENSE