Predicting Building Zone Air Temperatures using XGBoost and Feature Engineering: A Smart Buildings Challenge Submission

This paper details a submission for the Smart Buildings Challenge, aiming to predict withheld air temperature sensor measurements within Building SB1 using the Smart Buildings Control Test Suite dataset.The approach involved training a separate XGBoost model for each target temperature sensor, utilizing a tailored set of features for each.Features included measurements from devices located in the same zone as the target sensor.Recognizing that building states depend on recent history, lagged versions of these zonebased features were also incorporated.Additionally, features encoding temporal patterns, such as the hour of the day and day of the week, were included due to their significant impact on building dynamics.Results indicate that the XGBoost models achieved lower average Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE) compared to a baseline model that simply predicted the mean temperature for each sensor.However, performance varied significantly across individual devices; some models performed well, while others struggled to capture meaningful patterns.This variance underscores the need for further, more rigorous investigation.Notably, this submission was developed entirely within a single day as part of the challenge's beta testing phase.All the code used to obtain the results shown is available at https://github.com/gabriel-trigo/smart_buildings_challenge