codenhenhe/chronos2-volta-solar-daily

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_year
• month
• ALLSKY_SFC_SW_DWN
• T2M
• WS2M
• RH2M

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:

from autogluon.timeseries import TimeSeriesPredictor

predictor = 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.