Predicting energy performances of buildings' envelope wall materials via the random forest algorithm

Aseel Hussien, Wasiq Khan, Abir Hussain, Panos Liatsis, A. I. Al-Shamma’a, Dhiya Al‐JumeilyJournal of Building Engineering 2023 · DOI

TL;DR

This paper demonstrates that the Random Forest (RF) algorithm can accurately predict the energy performance of building envelope materials, significantly outperforming traditional linear regression models. By training on data from high-fidelity thermal simulations, the researchers created a predictive tool that helps architects optimize wall thickness, orientation, and material choice for hot climates like the UAE without the computational overhead of full-scale thermodynamic modeling.

What Problem Does This Paper Address?

Designing energy-efficient buildings in extreme climates (like the UAE’s heat and humidity) requires a precise understanding of how building envelopes—the "skin" of the building—interact with the environment. Traditionally, engineers rely on complex simulation software like EnergyPlus or IES-VE. While accurate, these tools are data-hungry, computationally expensive, and often require thermodynamic inputs that aren't available during early-stage design.

The authors identify a gap in the literature: while machine learning (ML) is used for general energy forecasting, there is limited research specifically applying it to the selection of wall materials and construction configurations. The goal is to provide a faster, "intelligent" way to predict yearly energy consumption based on architectural parameters.

Methods

The researchers employed a two-stage approach:

  1. Data Generation: Because real-world longitudinal data on diverse wall constructions is scarce, they used IES-VE (Integrated Environmental Solutions Virtual Environment) software to simulate 1,119 different scenarios.
  2. Predictive Modeling: They compared two primary ML approaches:
  3. Multivariate Linear Regression (MLR): A standard statistical baseline to check for linear relationships between variables.
  4. Random Forest (RF): A non-linear ensemble learning method that uses multiple decision trees and "bagging" (bootstrap aggregating) to improve accuracy and prevent overfitting.

The team used one-hot encoding to handle categorical data (like material types and construction shapes) and performed 10-fold cross-validation to ensure the models generalized well to unseen data.

Datasets

The study utilized a custom-generated dataset of 1,119 simulated data points. The input features included:

  • Materials: Heavy-weight Concrete Block (HCB), Aerated Concrete Block (ACB), Common Fried Brick (CFB), and Unfired Brick (UFB).
  • Wall Thickness: Ranging from 222mm to 300mm.
  • Orientation: North (0°), 45°, 90°, and 135°.
  • Construction Types: 13 distinct wall layer configurations (W1–W13).
  • Shape Factors: Three different building geometries (Shapes A, B, and C).

The target variable was Yearly Energy Consumption (MWh).

Key Insights and Ideas

  • Non-Linearity is King: The building envelope's thermal behavior is highly non-linear. The MLR model failed (R² = 0.42) because it couldn't capture the complex interactions between variables like thermal mass and orientation. Random Forest, by contrast, thrived on these complexities.
  • The "Big Four" Predictors: Through feature importance analysis, the paper identified four critical variables that drive energy prediction: Thermal Mass, Decrement Factor (DF) summer day, U-value (thermal transmittance), and R-value (thermal resistance).
  • Encoding Matters: Interestingly, the RF model performed better when the data was "encoded" (transformed into numerical vectors) compared to the original raw categorical data, pushing the R² from 0.92 to 0.95.
  • Early-Stage Utility: The model allows professionals to "rank" construction types. For instance, the study found that specific construction types (like W1) contributed more significantly to model predictions than others, providing a roadmap for which designs to prioritize in hot climates.

Key Results

  • Random Forest Superiority: The RF model achieved an R² of 0.95 and a Mean Squared Error (MSE) of 0.06 on the testing set.
  • Baseline Failure: The Multivariate Linear Regression model only managed an R² of 0.42, indicating it could not reliably track extreme energy consumption values.
  • Statistical Significance: While factors like wall thickness and orientation showed low relative standard error, they weren't always statistically significant on their own in the linear model. However, in the RF model, the combination of these features led to highly robust predictions.

Limitations and Future Work

The primary limitation acknowledged is the reliance on simulated data. While IES-VE is a gold standard in industry, simulations may not capture the "noise" of real-world occupancy behavior or construction defects.

Future directions include: * Integrating larger, more diverse datasets that include real-world sensor data. * Exploring more advanced deep learning architectures, such as Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTM) networks, to account for time-series variations in climate and usage. * Expanding the model to include more complex variables like window-to-wall ratios and HVAC system efficiencies.

Summary updated

BibTeX
@article{hussien2023predicting,
  author = {Aseel Hussien and Wasiq Khan and Abir Hussain and Panos Liatsis and A. I. Al-Shamma’a and Dhiya Al‐Jumeily},
  title = {{Predicting energy performances of buildings' envelope wall materials via the random forest algorithm}},
  journal = {Journal of Building Engineering},
  year = {2023},
  doi = {10.1016/j.jobe.2023.106263},
}

Loading related content...

Figure