Forecasting Model Results
ML Wind Forecasting Dashboard
This page shows historical holdout forecasts from the final Spark ML model. It compares predicted next-day wind potential against actual observed outcomes so the model can be evaluated honestly.
Final Model
final_tuned_gbt
Spark MLlib Gradient-Boosted Trees
RMSE
4.55%
Root mean squared forecast error. Lower is better.
MAE
2.75%
Average absolute forecast error. Lower is better.
Bias
0.02%
Average prediction minus actual. Near zero is best.
Forecast evaluation scope
The selectable years are 2023–2025 because this page shows historical holdout evaluation, not live future forecasting. Actual outcomes are already known for these dates, so RMSE, MAE, and bias can be measured. Future-date operational forecasting belongs in the next inference-service layer.
Evaluation Rows
535,961
Forecast rows joined with actual next-day outcomes.
Coverage
48 states
1995-01-31 to 2025-08-27
Target
Next-day capacity factor
next_day_daily_region_capacity_factor
Model interpretation
- • The model tracks normal wind-potential movement reasonably well.
- • The largest errors happen during sudden wind spikes, which are hard to predict from historical daily features.
- • The near-zero bias means the model is not consistently overpredicting or underpredicting.
- • Wind speed and rolling capacity-factor features dominate the signal.
Forecast vs Actual Explorer
Select a holdout state and year to inspect how closely the final tuned GBT model followed actual next-day wind potential.
TX 2023 Rows
365
State-day forecasts available for this selection.
Selection MAE
2.04%
Average absolute error for the selected state and year.
Forecast Type
Historical holdout
Used for evaluation because actual outcomes are known.
Forecast vs Actual — TX, 2023
Capacity factor is shown as a percentage. The closer the prediction line stays to the actual line, the better the model performed for that period. Sudden spikes are the hardest events to forecast.
Model Feature Importance
These features explain what the model relied on most. Wind speed, wind variability, and rolling historical capacity-factor features carry most of the signal.
Sample Predictions — TX, 2023
A row-level preview of actual next-day wind potential, model prediction, and absolute error.
| Date | Actual | Prediction | Absolute Error | Avg Wind Speed |
|---|---|---|---|---|
| 2023-01-01 | 9.60% | 4.64% | 4.96% | 3.35 m/s |
| 2023-01-02 | 5.07% | 8.54% | 3.47% | 5.11 m/s |
| 2023-01-03 | 2.68% | 5.11% | 2.44% | 3.98 m/s |
| 2023-01-04 | 0.35% | 5.61% | 5.25% | 2.75 m/s |
| 2023-01-05 | 5.10% | 2.25% | 2.86% | 1.77 m/s |
| 2023-01-06 | 2.64% | 8.37% | 5.73% | 3.45 m/s |
| 2023-01-07 | 2.34% | 4.62% | 2.28% | 3.45 m/s |
| 2023-01-08 | 1.38% | 4.76% | 3.38% | 3.32 m/s |
| 2023-01-09 | 2.94% | 4.16% | 1.22% | 2.53 m/s |
| 2023-01-10 | 8.09% | 5.91% | 2.19% | 3.10 m/s |
| 2023-01-11 | 15.20% | 7.58% | 7.62% | 4.64 m/s |
| 2023-01-12 | 3.69% | 7.63% | 3.93% | 6.16 m/s |