Data-Driven Modeling for Hourly Building Energy Use Prediction with Corrupt Weather Dataset

The growing concern over the impact of energy use and carbon emissions on the environment has led to an increased interest in building energy consumption related topics. How to better predict hourly building energy consumption given the outdoor weather conditions without manually constructing complicated physics-based building energy model is remained to be explored. Accordingly, the hourly data on electricity, steam, and chilled water usage in a campus building are collected, which were further screened and processed before matching it with the corresponding hourly weather data. Unfortunately, the collected weather data was found to be incomplete and corrupt, with irregular and missing data points. In response, we developed a sliding-window-based method to obtain reliable and essential weather data for use in training data-driven models. Diverse machine learning algorithms were employed to construct data-driven models for chilled water and steam usage. To optimize the performance of the models, grid search approach was adopted to fine-tune the hyper-parameters. Subsequently, we compared and evaluated the predictive ability of these models using distinct feature selection tactics. The investigation demonstrated that the Adaboost trees and Random Forest algorithms exhibited the highest accuracy in forecasting the dataset.