Short-term forecasting of solar irradiance using decision tree-based models and non-parametric quantile regression

Area of Science:

• Renewable Energy Systems
• Statistical Modeling
• Meteorological Forecasting

Background:

• Accurate solar irradiance (SI) forecasting is vital for managing renewable energy generation and supply.
• Intermittent nature of solar power necessitates advanced forecasting techniques.
• Existing SI studies often focus on conditional mean distribution, limiting uncertainty representation.

Purpose of the Study:

• To evaluate and compare the forecasting performance of quantile regression random forest (QRRF) and quantile generalised additive model (QGAM) for solar irradiance.
• To introduce QRRF and QGAM as novel forecasting frameworks for SI studies.
• To assess the models' ability to represent forecast uncertainty beyond prediction intervals.

Main Methods:

• A simulation study using multivariate data-generating processes to compare forecasting accuracy.
• Evaluation of models using pinball loss scores and mean absolute scaled errors.
• Application of models to real-life solar irradiance data.

Main Results:

• QRRF showed comparable performance to QGAM in predicting the forecast distribution.
• QGAM demonstrated clear superiority over QRRF in terms of pinball loss and mean absolute scaled errors.
• Both QRRF and QGAM provided complete information on forecast uncertainty via quantile estimation.

Conclusions:

• QGAM is recommended over decision tree-based models like QRRF for solar irradiance prediction due to its superior accuracy.
• QRRF can serve as an alternative for predicting the forecast distribution.
• The QRRF and QGAM frameworks can be extended to model other meteorological-dependent renewable energy sources.