Prediction and Ranking of Moisture Severity of Climate Years Using Supervised Projection to Latent Structures Applied to Tall Wood Buildings

Maurice Defo, Michael Lacasse, Terrence J. Moore2023 · DOI

Abstract

Hygrothermal simulation tools are commonly used to assess the moisture performance of building envelope components. Owing to the computational costs required to complete simulations over the long-term, one approach to reduce simulation time when undertaking hygrothermal design analysis is to select representative year(s) amongst sets of long-term climate data. To properly select these moisture reference year(s), a method is required to rank or predict the moisture severity of climate years for sets of long-term climate data. Several methods are used in the literature for this purpose, but none seems to be sufficiently accurate. In this study, the supervised projection to latent structures, also known as partial least squares regression, was trained and validated on data obtained from hygrothermal simulations of tall wood building wall assemblies for several cities across Canada. Models developed at the city level, for a given greenhouse gas emission scenario or time period, or encompassing historical and future time periods, showed comparable scores for ranking. In respect to prediction of the moisture severity of climate year sets, models developed at the city level were shown to be more accurate.

Introduction

Designing durable tall wood buildings requires a deep understanding of how building envelopes manage heat and moisture over decades—a field known as hygrothermal analysis. However, running high-fidelity 2D or 3D hygrothermal simulations for 30 consecutive years is computationally intensive, particularly when accounting for the uncertainties of climate change.

To streamline this process, engineers often select a "Moisture Reference Year" (MRY)—a single representative year that imposes severe moisture stress—to test their designs. The challenge lies in accurately identifying that year without performing the very simulations the MRY is intended to replace. This research explores whether machine learning—specifically Supervised Projection to Latent Structures (PLS)—can bridge this gap by predicting moisture severity from climate data.

Methods and Datasets

The study utilized a large ensemble of modeled climate data for 12 Canadian cities representing diverse climate zones, ranging from Vancouver’s mild, rainy climate to Whitehorse’s subarctic cold. Data covered a historical period (1986–2016) and a future period (2062–2092) based on the RCP8.5 high-emission scenario.

Annual rain distribution for 12 Canadian cities across historical (H) and future (F) climate periods.

The authors employed Partial Least Squares (PLS) regression, a dimension-reduction technique well-suited for datasets with numerous highly correlated predictor variables, such as temperature and solar radiation. The model aimed to predict the average Moisture Content (MC) of the outer layer of a Cross-Laminated Timber (CLT) panel. Ground truth data was generated using 1,448 one-dimensional hygrothermal simulations in DELPHIN v5.9.

Three modeling approaches were compared:

  1. City-level: Trained on data from a single city.
  2. Time-period: Trained on data from all cities for a specific era.
  3. Global: A single model encompassing all cities and all time periods.

Key Results

The study found a vital distinction between predicting exact moisture content and ranking years by severity. While local factors are crucial for absolute values, the underlying physics of moisture severity are consistent enough that a general model can rank years effectively.

Performance metrics for period-level and city-level models under future climate scenarios.

As shown in the results for future scenarios, all models demonstrated excellent ranking ability. Spearman’s rank correlation ($\rho$) values were typically above 0.90. City-level models were the most precise for absolute values, achieving a median Root Mean Square Error (RMSE) of 0.26% moisture content, compared to 0.54% for broader period-level models. Furthermore, the models were highly successful at identifying the "Top 10" worst years, typically matching 8 or 9 out of 10 years identified by the full-scale simulations.

Discussion and Limitations

A "Global model" performed nearly as well as specialized models for ranking the most severe years. This suggests that a single robust model can serve as a universal "severity filter," significantly reducing the need for city-specific model development. By using PLS instead of standard multiple regression, the authors successfully managed the collinearity of climate variables, leading to stable and reliable predictions.

However, some limitations remain. The training data was based on 1D simulations, which may not capture complex moisture patterns in 2D or 3D joints. Additionally, the models were trained on a specific 13-story CLT wall assembly; their generalizability to other wall types requires further investigation. Future work could involve expanding the training set to include more wall configurations and integrating these PLS models into design software for real-time risk assessment.

Summary updated

Abstract

Hygrothermal simulation tools are commonly used to assess the moisture performance of building envelope components. Owing to the computational costs required to complete simulations over the long-term, one approach to reduce simulation time when undertaking hygrothermal design analysis is to select representative year(s) amongst sets of long-term climate data. To properly select these moisture reference year(s), a method is required to rank or predict the moisture severity of climate years for sets of long-term climate data. Several methods are used in the literature for this purpose, but none seems to be sufficiently accurate. In this study, the supervised projection to latent structures, also known as partial least squares regression, was trained and validated on data obtained from hygrothermal simulations of tall wood building wall assemblies for several cities across Canada. Models developed at the city level, for a given greenhouse gas emission scenario or time period, or encompassing historical and future time periods, showed comparable scores for ranking. In respect to prediction of the moisture severity of climate year sets, models developed at the city level were shown to be more accurate.
BibTeX
@misc{defo2023prediction,
  author = {Maurice Defo and Michael Lacasse and Terrence J. Moore},
  title = {{Prediction and Ranking of Moisture Severity of Climate Years Using Supervised Projection to Latent Structures Applied to Tall Wood Buildings}},
  year = {2023},
  doi = {10.23967/c.dbmc.2023.069},
}
Figure