Solar Generation Forecasting
Forecasting monthly U.S. solar generation, utility-scale and small-scale, from U.S. Energy Information Administration data for January 2022 to October 2024. The series is dominated by a strong 12-month cycle riding on fast capacity growth, so the question was which models capture both without overfitting 34 months of data.
Approach
Six models, from simplest to most flexible: two sinusoidal regressions (a seasonal sine around an exponential or logistic midline), SARIMAX, a plain feed-forward neural network, a time-lagged neural network, and a seasonal neural network (SANN) that sees the same month in prior years. The series was supplemented with a growth rate from a regression on installed solar capacity and with monthly total solar irradiance.
Results
| Model | RMSE | MAE |
|---|---|---|
| Sinusoidal exponential-midline regression | 889.8 | 748.7 |
| Sinusoidal logistic regression | 3354.0 | 2628.7 |
| SARIMAX | 1021.9 | 880.2 |
| Artificial neural network | 3294.4 | 2828.0 |
| Time-lagged neural network | 1360.3 | 1141.3 |
| Seasonal artificial neural network (SANN) | 817.2 | 637.8 |
The SANN had the lowest RMSE and MAE and held up across training runs, so it was the recommended model. The exponential-midline sinusoid came close with a far simpler model, which makes it the better choice where simplicity matters, though it can't use exogenous inputs.
Limitations
- Only 34 monthly observations, too few for the flexible models to learn long-run trend and seasonality reliably.
- Weather, economic and policy variables were left out; the forecasts assume history repeats without a structural break.
- Trained on a laptop, which limited features, epochs and cross-validation folds.