Session: 02-02: AI for Energy Sustainability II

Paper Number: 156600

156600 - Clustering Buildings Energy Consumption for Personalized Lightweight Forecasting 

Abstract: 

Optimizing energy consumption in buildings is crucial for enhancing sustainability and energy efficiency. The ability to predict a building's energy consumption facilitates the identification of high usage, compliance with energy policies, and the planning of efficiency measures. However, a key challenge is the heterogeneity of the physical characteristics of buildings, which makes consumption patterns not easily transferable between different structures. Factors such as occupancy, location, internal activities, and equipment used significantly influence consumption patterns, supporting the need for personalized prediction models that adapt to the unique characteristics of each building.

This study addresses this challenge by using clustering techniques to improve energy consumption forecasting and segmenting consumption patterns within a single building. Instead of applying a single model for the entire building, we propose that each building can host multiple consumption patterns, each responding to specific internal and external dynamics, such as usage schedules, occupancy variability, or seasonal changes. Cluster segmentation allows for identifying recurring behaviors, improving the accuracy of predictive models. Additionally, these models are lightweight and quick to train, facilitating continuous updates without significant computational overhead.

Two main clustering approaches are explored: feature-based clustering, which transforms consumption time series into a lower-dimensional space by extracting relevant features with high interpretability, and distance-based clustering, which segments time series based on similarity metrics that can handle pattern variations and capture complex consumption behavior. We will evaluate the combination of both approaches to determine which provides the highest accuracy in consumption forecasts, enabling a more precise selection of model features.

For energy consumption prediction, we will use ensembles of decission trees due to their simplicity, ease of training, and ability to capture nonlinear relationships in the data that emerge from the nature of energy consumption. Instead of procuring linearized models that are limited in their representation capacity, these models integrate well with the clustering results, allowing for accurate predictions based on the identified patterns. By using lightweight approaches for clustering and forecasting, we ensure the effectiveness and applicability of the proposed solution. Traditional statistical models will be used as a baseline to compare the proposed approaches. Although tree ensembles do not automatically adapt to changes in the data, their retraining is feasible and efficient when significant changes in consumption patterns occur.

The analysis will be carried out using a diverse set of energy consumption data from buildings with different functions, such as offices, conference rooms, lecture halls, computer labs, and cooling units of various capacities. These scenarios showcase the variability and complexity of the application context. These data will allow us to evaluate how clustering and forecasting techniques impact prediction accuracy and provide a basis for designing more efficient solutions for energy management in buildings. We expect this study not only to demonstrate the feasibility of the proposed approaches but also to provide a practical framework for the scalable implementation of predictive models in building energy management, enabling continuous adjustments and real-time improvements as new data are incorporated.

Presenting Author: Jose Cordova-Garcia ESPOL University

Presenting Author Biography: Jose Cordova-Garcia is an Associate Professor with the Faculty of Electrical Engineering and Computer Science at ESPOL University. He received his PhD degree from the State University of New York at Stony Brook. Currently he leads the Artificial Intelligence and Analytics for Sustainability initiative. His research interests include differentiable optimization, cyber-physical systems, machine learning for optimization, and data issues in remote sensing. He was a recipient of a Fulbright Scholarship from the Department of State during his M.S. Studies.