A Multiple Linear Regression Model Based on Spatial Temperature and Humidity Clustering for Building Energy Use Intensity Forecasting

Received: 24 February 2026 | Revised: 24 March 2026 and 3 April 2026 | Accepted: 4 April 2026 | Online: 24 April 2026

Corresponding author: Iwa Garniwa

Abstract

Accurate building Energy Use Intensity (EUI) is crucial for improving building energy efficiency, particularly in countries with significant climatic diversity such as Indonesia. However, many current regression-based energy prediction models rely on national-scale climate data, implicitly assuming a uniform relationship between climate variables and energy consumption across regions. This assumption may reduce prediction accuracy in geographically heterogeneous environments. To address this limitation, this study focuses on office buildings, whose operational cooling demand is strongly influenced by local climatic conditions, necessitating the integration of spatially clustered temperature and humidity data with Multiple Linear Regression (MLR) modeling. Provincial climate data from 38 Indonesian provinces are first classified into homogeneous climate zones using the Fuzzy C-Means (FCM) clustering method. FCM clustering was chosen because it can produce smoother transitions in differences across climate data and handle overlapping data through membership degrees in spatial clustering, compared to k-means clustering. Subsequently, cluster-specific MLR models are developed to predict EUI within each identified climate group. The performance of the cluster-scale model is compared and validated using the coefficient of determination (R²), along with Mean Absolute Error (MAE) and Root Mean Square Error (RMSE), to assess the variance in EUI explained by climate, relative to a conventional national-scale regression model. The results show that the cluster-specific regression models consistently improve the forecasting accuracy of the national-scale models across all climate clusters. The MAE decreased from 0.19–0.34 at the national scale to 0.04–0.13 at the cluster scale, while the RMSE decreased from 0.25–0.43 to 0.05–0.16. Overall, the findings indicate that aligning regression modeling with climate-based clustering significantly improves accuracy. Furthermore, these findings provide a basis for policymakers and building managers to develop climate-specific energy efficiency strategies. Although climate variables were the primary predictors in this study, operational factors, such as occupancy, HVAC control settings, and building envelope characteristics, could also be integrated in future studies.