XGBoost with physics-informed features and residual regressor for the SBCS benchmark

Vladimir Pyltsov

Keywords: XGBoost, Forecasting, Feature Engineering

TL;DR: This work proposes an XGBoost model with engineered features and residual regressor for the SBCS benchmark.

Abstract: This study proposes an enhanced XGBoost model with physics-informed features and a residual regressor. The model is used in the long-term forecasting setting of the Smart Buildings Control Suite (SBCS) benchmark, with a context of building energy dataset with a large exogenous matrix and similar lengths of training and test sets. The results show improvements over more than 10% across the majority of the selected horizons compared to the baseline XGBoost model without any modifications. A notable error improvement includes the horizon of the full test set. The proposed model can be used as an initial step towards further advancements in the capabilities of tree-based models in long-term forecasting and building energy setting.

Submission Number: 57