Carbon-aware scheduling

A 30-line Python service that picks the cleanest hour in the next 24 to run a batch job, cutting its embedded emissions by 10–25% at zero cost.

Why shift load

India's grid carbon intensity swings between ~550 gCO₂/kWh at solar-heavy midday and ~800+ gCO₂/kWh in the late evening coal peak. Any workload that doesn't need to finish at a specific instant — a nightly ETL, model training, document indexing — can be scheduled against that curve for meaningful carbon savings with zero performance penalty.

The approach

1. Fetch the next 24 hours of forecasted carbon intensity from /forecast/carbon-intensity.
2. Find the contiguous window that minimises average intensity over your job's estimated duration.
3. Schedule the job to start at the window's start time.
4. (Optional) Poll /carbon-intensity/latest while running and log actual intensity for your green-ops dashboard.

Implementation

Minimal Python example. Assumes ATLAS_API_KEY is on a Pro tier or above (forecasts require Pro).

import os
import datetime as dt
import requests

API = "https://api.energymap.in/developer/v1"
HEADERS = {"X-API-Key": os.environ["ATLAS_API_KEY"]}


def best_start_time(job_hours: int) -> dt.datetime:
    """Return the UTC start time of the cleanest window of length `job_hours`."""
    r = requests.get(f"{API}/forecast/carbon-intensity", params={"horizon_h": 24}, headers=HEADERS, timeout=10)
    r.raise_for_status()
    forecast = r.json()["forecast"]

    best_idx, best_avg = 0, float("inf")
    for i in range(len(forecast) - job_hours + 1):
        window = forecast[i : i + job_hours]
        avg = sum(p["intensity_gco2_per_kwh"] for p in window) / job_hours
        if avg < best_avg:
            best_avg, best_idx = avg, i

    start = dt.datetime.fromisoformat(forecast[best_idx]["ts"].replace("Z", "+00:00"))
    print(f"cleanest {job_hours}h window: starts {start.isoformat()}, avg {best_avg:.0f} gCO2/kWh")
    return start


def current_intensity() -> int:
    r = requests.get(f"{API}/carbon-intensity/latest", headers=HEADERS, timeout=10)
    r.raise_for_status()
    return r.json()["intensity_gco2_per_kwh"]


if __name__ == "__main__":
    start_at = best_start_time(job_hours=6)
    # Wire up to systemd-run --on-calendar, Airflow, or your scheduler of choice.
    # For crontab: convert start_at to local time and `at <time>` the job.

For Airflow, wrap best_start_time in a short Python operator at the top of your DAG and pass its output as the schedule for the downstream task group. For systemd, systemd-run --on-calendar accepts an absolute timestamp.

Verifying the savings

Log the intensity at the moment each scheduled job starts and compare against a naive baseline (e.g. "ran at 22:00 IST every night"). After a month of runs you typically see 10–25% lower average intensity, higher on days with a clear midday solar peak, lower on overcast days where the curve is flatter.