Weather forecasts can't see what a plant did an hour ago. This model can.
Renewcast's forecasts come from weather models, so they miss what a plant is doing right now. Soiling, a tripped inverter or a cloud bank 20 km off course all show up in the readings first. I trained a small recurrent network that reads the latest readings next to the forecast customers already received and corrects the next four hours. One model covers the solar fleet and another covers the wind fleet.
- Role
- Model design, training and evaluation
- Stack
- PyTorch, Python, MLflow, Databricks
21.9%
Median cut in solar error over the first two hours
26.6%
Median cut in wind error over the first two hours
How it works
- 1
Delivered forecast
- 2
Latest plant readings
- 3
Fleet model
solar or wind
- 4
Four-hour correction
Results
Solar forecast error over the next four hours
Average error at each step ahead. Persistence is slightly better for the first 30 minutes and the fleet model is better from 45 minutes. Both hand back to the delivered forecast at four hours.
- Fleet GRU
- Persistence
- Delivered forecast
Backtest of an earlier version of the fleet model on Renewcast's solar fleet.
Show dataHide data
| Delivered forecast | Persistence | Fleet GRU | |
|---|---|---|---|
| 15 min | 6.1% | 2.7% | 3.1% |
| 30 min | 6.1% | 3.7% | 3.8% |
| 45 min | 6.0% | 4.3% | 4.1% |
| 60 min | 6.1% | 4.6% | 4.4% |
| 75 min | 6.1% | 5.0% | 4.7% |
| 90 min | 6.1% | 5.2% | 4.9% |
| 105 min | 6.0% | 5.3% | 5.1% |
| 120 min | 6.1% | 5.6% | 5.3% |
| 135 min | 6.1% | 5.5% | 5.3% |
| 150 min | 6.1% | 5.7% | 5.5% |
| 165 min | 6.1% | 5.8% | 5.6% |
| 180 min | 6.2% | 5.8% | 5.7% |
| 195 min | 6.2% | 6.0% | 5.7% |
| 210 min | 6.2% | 6.0% | 5.7% |
| 225 min | 6.1% | 6.0% | 6.0% |
| 240 min | 6.2% | 6.2% | 6.2% |
In a backtest that held out each client, the median plant-month's error over the first two hours fell 21.9% for solar and 26.6% for wind. Carrying the latest error forward managed 16.3% and 25.3%.
Key decisions
- 01
One model for the whole fleet
A single GRU learns how forecast errors develop across plants. It beat per-plant models, and it means maintaining two models instead of 224.
- 02
Test it the way it will run
Training uses forecasts that were actually delivered and readings with realistic delays. I held out each client in turn and scored each month with a model trained only on earlier months. A check rejects any feature that leaks the future.
- 03
Never correct its own output
The model always corrects the original forecast, never one it already corrected. Output stays within plant capacity. If the newest reading is more than two hours old, the original forecast goes out unchanged.