Machine learning-based forecasting of campus building energy consumption under climate change scenarios: Toward sustainable energy management
Journal
Engineering Applications of Artificial Intelligence
Journal Volume
170
Start Page
114240
ISSN
09521976
Date Issued
2026-04-15
Author(s)
Abstract
Global climate change has become increasingly severe, with building energy consumption accounting for 40% of the worldwide total. Machine learning offers promising potential for predicting and controlling building energy usage. This study analyzes campus electricity consumption data through feature engineering and machine learning, considering building energy consumption trends under various climate change scenarios. Weather and historical building data are collected and employed for data preprocessing and feature selection. Three types of machine learning models were constructed and optimized with the Tree-structured Parzen Estimator (TPE) algorithm. The Long Short-Term Memory (LSTM) model achieved a mean absolute percentage error (MAPE) of 4.24% for short-term predictions, which was 0.36% to 1.06% better compared to the Random Forest (RF) and Multi-Layer Perceptron (MLP) models. The study further evaluates long-term forecasting, revealing through Shapley Additive Explanations (SHAP) analysis that factors such as the previous day's trend, one-week prior trend, campus day classification, and tomorrow's campus day category are critical determinants. Under the high-emission scenario Shared Socioeconomic Pathways (SSPs) 5-8.5, campus building energy consumption is projected to increase by 18%-23% between 2028 and 2048. This study proposes a comprehensive building energy management framework that integrates energy load forecasting with climate change scenarios, providing significant insights for sustainable schools and net-zero pathways.
Subjects
Campus buildings
Climate change
Energy consumption forecasting
Feature engineering
Machine learning
Publisher
Elsevier Ltd
Type
journal article