Multi-Type Building Energy Consumption Prediction Using Machine Learning and IoT Sensor Data

Raparthi, Bharath (2025) Multi-Type Building Energy Consumption Prediction Using Machine Learning and IoT Sensor Data. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (627kB) | Preview
Preview	PDF (Configuration Manual) Download (781kB) | Preview

Abstract

Energy consumption in buildings constitutes about 40% of total energy consumption. Accurate prediction of energy consumption therefore becomes a critical prerequisite for sustainability and operational efficiency. This article presents a machine learning model for multi-type energy consumption prediction based on IoT sensor data from commercial buildings. This study bridges gaps existing in multi-meter energy prediction; uncertainty problems associated with energy prediction; and model interpretability based on building types. Using the ASHRAE Great Energy Predictor III dataset, the research employed 2 million sampled training records and 1 million test records from the original 20.2 million hourly readings across 1,449 buildings and 16 sites. A three-phase pipeline was implemented incorporating advanced feature engineering, ensemble modeling, and systematic evaluation. The model creates 70+ features for predictions based on temporal variables, weather variables, building variables, and autoregressive variables for electricity consumption, chilled water consumption, steam consumption, and hot water consumption using XGBoost for quantile regression and LightGBM for predictions. Analysis shows that rolled means of previous values with a 7-day time span are more prominent with 50-60% importance for predictions with surprisingly low importance for weather variables. LightGBM gave better results for predictions for three out of four types of energy with 14-20% improvement in RMSE values compared to other methods. However, large values of MAPE (357-1,050%) indicate major issues with zero-inflated distributions and irregular consumption. The uncertainty quantification produced coverage estimates of 64-78%, failing to achieve confidence level targets. This work clearly shows that existing ensemble approaches are insufficient for multi-type energy prediction problems and thus necessitate a special architecture that can handle complexities of distributions in a manner that is interpretable for realistic implementation in building management systems.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Nolan, Eamon UNSPECIFIED
Subjects:	G Geography. Anthropology. Recreation > GF Human ecology. Anthropogeography > Human settlements > Built environment > Buildings H Social Sciences > HC Economic History and Conditions > Natural resources > Power resources > Energy consumption T Technology > TK Electrical engineering. Electronics. Nuclear engineering > Telecommunications > Computer networks > Internet of things Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Machine learning
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Ciara O'Brien
Date Deposited:	08 Sep 2026 11:54
Last Modified:	08 Sep 2026 11:54
URI:	https://norma.ncirl.ie/id/eprint/9898

Actions (login required)

View Item