Chronos-2 Fine-tuned for Australian Solar Generation Forecasting
This repository contains a fine-tuned Chronos-2 model for daily solar generation forecasting in Australia. The model was fine-tuned using AutoGluon TimeSeriesPredictor and saved as an AutoGluon predictor directory.
Model Details
- Base model:
amazon/chronos-2
- Framework: AutoGluon TimeSeriesPredictor
- Fine-tuning method: LoRA
- Forecasting task: Daily solar generation forecasting
- Prediction length: 365 days
- Context length: 730 days
- Target variable: Daily solar generation
- Evaluation metric during training: MSE
Input Features
The model uses historical solar generation together with temporal and weather-related covariates.
Known covariates include:
day_of_yearmonthALLSKY_SFC_SW_DWNT2MWS2MRH2M
Initial Metrics
The following metrics were obtained from the initial evaluation after fine-tuning. The model was evaluated on the 2024 and 2025 test periods.
Point Forecast Accuracy
| Model | Period | MAE | MSE | RMSE | sMAPE (%) | R2 |
|---|
| Chronos Fine-Tuned | 2024 | 2.374061 | 11.042445 | 3.323017 | 10.768443 | 0.889873 |
| Chronos Fine-Tuned | 2025 | 2.468275 | 12.076611 | 3.475142 | 11.053124 | 0.883387 |
| Chronos Fine-Tuned | Overall | 2.421168 | 11.559528 | 3.399931 | 10.910783 | 0.886614 |
Probabilistic Metrics
| Model | Period | Q_0.1 | Q_0.9 | Avg_Q_Loss |
|---|
| Chronos Fine-Tuned | 2024 | 0.601763 | 0.500016 | 0.550889 |
| Chronos Fine-Tuned | 2025 | 0.635956 | 0.506356 | 0.571156 |
| Chronos Fine-Tuned | Overall | 0.618859 | 0.503186 | 0.561023 |
Baseline Comparison
The fine-tuned model was also compared against a seasonal naive baseline.
| Model | MAE | RMSE | MAPE (%) | R2 |
|---|
| Chronos2-Volta AI | 2.600748 | 3.640213 | 12.897249 | 0.870021 |
| Seasonal Naive Baseline | 6.080324 | 8.468857 | 29.693408 | 0.296490 |
Loading the Model
This model is stored as an AutoGluon TimeSeriesPredictor directory. After downloading the repository, it can be loaded as follows:
1from autogluon.timeseries import TimeSeriesPredictor
2
3predictor = TimeSeriesPredictor.load("path_to_downloaded_model")
Intended Use
This model is intended for postcode-based daily solar generation forecasting in Australia. It can be used as part of a pipeline that maps a user postcode to a representative solar reference point, retrieves or simulates weather covariates, and forecasts future solar generation.
Limitations
The model depends on the quality of postcode mapping, weather covariates, and historical solar generation data. Forecast accuracy may vary across regions, seasons, and unusual weather conditions.
