2FUTURE Application of XGBoost to Site-specific Weather Forecast in Australia

Site-specific weather forecasts are essential to the accurate prediction of power demand, and consequently are of great interest to energy operators. However, current numerical weather prediction (NWP) models lack the fine resolution needed for localised weather forecasts, and instead provide the averaged weather information within each model gridbox (usually of order kilometers in size). Even after post-processing and bias correction, area-averaged information is usually not optimal for sites. Prior work on site optimisation has focused on linear methods, weighted consensus averaging, time-series methods and others.

In this study, we are investigating the feasibility of optimising forecasts at sites using popular machine learning model: gradient boosting decision trees, powered by Python version of the package XGBoost. Regression trees were trained with historical NWP and site observation data to predict temperature and dew point at multiple site locations. A working ML framework, named 'Multi-SiteBoost' has been established and initial results show a significant improvement compared with gridded values from bias-corrected NWP models. A partial comparison against some alternative methods will also be presented.