See the cloud before the power drops.
মেঘ আসার আগেই দেখুন, ডিজেল পোড়ানোর আগেই ভাবুন।
MeghDrishti watches the sky with a low-cost camera and tells solar operators in Bangladesh what the next two hours will bring, so they stop running diesel gensets just in case.
- Monsoon clouds move fastA rooftop plant can lose half its output within minutes, then recover just as quickly.
- Diesel runs just in caseWith load shedding and costly fuel imports, factories keep gensets on standby because they cannot see a drop coming.
- Hourly forecasts can't see a cloudA sky camera can. In our research it gave the biggest gain on partly cloudy days.
Scanning the sky.
Rooftop fisheye camera, one image per minute. The dashed line marks a cloud on course for the sun.
Solar output, next 2 hours
Diesel kept off today
Compared with keeping the genset on standby whenever the sky is not clear.
Edit assumptions
CO₂ uses 2.68 kg per litre of diesel. MeghDrishti puts the genset on standby only during red alerts.
What the model looks at
Share of attention per source as the forecast looks further ahead.
Pattern from our paper (Fig. 6): the camera matters more for longer leads and on cloudy days. Weights are descriptive, not causal.
Operator alerts in Bangla
Sent by SMS or app when the advice changes.
Panel cleaning advisor
On clear periods, measured output should match the physics-based clear-sky limit. A growing gap means dust on the panels.
Dust loss rate is an assumption until measured at the Dhaka pilot site.
From one sky photo to one clear instruction
Every five minutes MeghDrishti runs the same loop on a small computer at the site. It keeps working on a weak internet connection because the model is small enough to run locally.
- CaptureA Raspberry Pi with a fisheye lens photographs the whole sky every minute, for under BDT 15,000.
- CorrectTimestamps are checked against the sun's path, and pvlib computes the clear-sky limit for the site.
- ForecastHAMF fuses the last three sky images, 60 minutes of power and the clear-sky prior into 19 quantiles for 5 to 120 minutes ahead.
- DecideRules turn the forecast range into one action: charge the battery, delay a load, keep diesel off, or prepare the genset.
- AlertThe operator gets the action in Bangla on the dashboard and by SMS, with how confident the forecast is.
Built on tested research, with honest limits
The forecasting model comes from our research on SKIPP'D, a public dataset of sky images and PV power (517 days). The paper is under double-blind review.
| Model (20 test days) | RMSE (kW) | Skill vs persistence | Probabilistic skill | 90% range width |
|---|---|---|---|---|
| HAMF (ours) | 3.38 | 17.9% | 34.8% | 5.6 kW |
| TiDE-style dense model | 3.47 | 15.8% | 31.3% | 6.6 kW |
| Probabilistic persistence | 4.12 | 0.1% | 0.0% | 9.8 kW |
| Smart persistence | 4.12 | 0.0% | −16.6% | – |
| SUNSET CNN (benchmark) | 4.25 | −3.1% | −80.9% | – |
HAMF beats persistence and the SUNSET benchmark with p < 0.001. Its 90% range is 43% narrower than probabilistic persistence while covering more observations (87.6% vs 83.4%).
What we still need to prove
- The model was trained at one site in California. Performance in Bangladesh is untested until our Dhaka camera pilot.
- The gain over a strong dense baseline (TiDE-style) is not statistically significant (p = 0.20).
- Packaged weather forecasts added no skill at these horizons. The camera does the heavy lifting.
- Diesel and CO₂ numbers in this demo come from the stated assumptions, not field measurements.
Plan through the Grand Finale
Each round moves one step from research result to a working system on a Bangladeshi rooftop.
- Idea and evidenceProposal, research results and this interactive prototype.
- Working prototypeHAMF served through an API, this dashboard on live inference, and a sky camera installed on a Dhaka rooftop.
- Live demo at AIUBForecasts from real Dhaka sky images and a first local comparison against persistence.
- PilotFine-tune on local data, SMS gateway, satellite fallback for sites without a camera, and trials with one factory and one mini-grid.