PSI - Issue 73

Lenganji Simwanda et al. / Procedia Structural Integrity 73 (2025) 138–145 Simwanda et al. / Structural Integrity Procedia 00 (2025) 000–000

139

2

multiplying the ground snow load by a shape coefficient that accounts for roof geometry, exposure, and thermal properties (Croce, Formichi, and Landi, 2021b). Heated roofs (those with significant heat transfer through the roof) tend to sustain lower snow loads than unheated roofs, because heat transfer from the interior melts the snowpack. The ratio of roof snow load to ground snow load is sometimes expressed as a thermal coefficient in codes (Zhou et al., 2024). This coefficient is less than 1.0 for heated buildings (indicating reduced roof load) and depends on factors like indoor temperature and roof insulation level (Croce et al., 2021). In some cases, if a building is well insulated or kept cold, the roof snow load may approach or even equal the ground snow; where the internal temperature is intentionally kept below 0 °C such as for freezer buildings or ice-skating arenas, the roof load may even exceed the ground load. Accurately predicting roof snow loads is important for structural design and risk assessment (He et al., 2022). However, predicting the time-varying snow load on roofs is complex. It involves accumulation from snowfall and ablation through melting, sliding, and sublimation, all of which depend on meteorological conditions. Physically based models (e.g. multi-layer energy balance snow models) can simulate these processes, but they require detailed calibration and can be computationally intensive to run for long-term climates. With the increasing availability of high resolution climate data (e.g. the ERA5 reanalysis provides hourly weather variables globally (Hersbach et al., 2020) and advanced machine learning techniques, a data-driven approach offers a promising alternative. Machine learning (ML) models can learn the relationship between weather parameters and resulting roof snow loads from historical data, capturing complex nonlinear interactions automatically. Once trained, such models can rapidly predict snow load for new weather inputs, which is valuable for near-real-time forecasting or probabilistic analysis over many simulated winters. This paper explores the use of ML to predict snow load on a heated roof with thermal transmittance U = 1.0 W/m²K (a moderately insulated roof), using a long-term climate dataset for Oslo. The objective is to develop and compare ensemble-based ML models that take meteorological observations as inputs and output the roof snow load. We focus exclusively on the case U = 1.0 W/m²K to demonstrate the methodology. Four different ML algorithms – Random Forest (RF), Gradient Boosting Machine (GBM), Extreme Gradient Boosting (XGBoost), and Categorical Boosting (CatBoost) – are implemented and evaluated. Key questions addressed include: (i) how accurately data-driven models can reproduce the simulated roof snow load, (ii) which weather features are most influential for prediction, and (iii) how the models compare in performance and interpretability. The study is intended as a proof-of-concept for data driven modeling of roof snow loads. While only the U = 1.0 case is detailed here, the approach can be extended to other insulation levels (U = 1.5, 2.0, etc.), which we briefly discuss as future work. 2. Database Development The analysis uses a custom hourly dataset from 1990–2020 (31 years) representing Oslo’s climate, based on ERA5 reanalysis data (60°N, 10.75°E, 247 m elevation). The dataset contains a total of 271,753 data points. Key features (listed in Table 1) include temperature, humidity, wind speed and direction, pressure, cloud cover, radiation fluxes, and hourly precipitation. ERA5 also provides ground snow depth (SWE) as a proxy for snow load. The dataset captures full seasonal cycles, ensuring the ML models are exposed to diverse snow, melt, and weather conditions (Table 1). To generate ML targets, the physics-based SIMELT model simulated hourly ground and roof snow loads (SWE in mm) from weather data. Ground load was based on a flat, unheated surface, while the roof model included heat transfer (U = 1.0 W/m²K) and snow sliding. Two roof cases were modeled: (a) No Sliding – all snow accumulates (e.g. for flat roofs or where sliding is restricted); and (b) With Sliding (e.g. for sloped roofs). Simulating 31 years in Oslo, maximum SWE reached ~274 mm (2.7 kN/m²) on the ground and ~236 mm (2.3 kN/m²) on the non-sliding roof; sliding roofs showed lower loads. Roof snow load was zero ~75% of the time, showing non-winter seasons. The resulting dataset (weather- SWE) was used to train and validate ML models. Note that in the following, only the case of No sliding is analyzed. Before training, timestamps were standardized and non-informative features (e.g., raw date/time, wind direction) were removed or transformed. The target was roof snow load (U = 1.0, no-sliding), representing worst-case accumulation. SWE was converted to kN/m ² (1 mm ≈ 0.00981 kN/m ² ) for structural relevance. Features were standardized, and ERA5 data had no significant outliers. The final dataset included ~271,753 hourly samples of meteorological inputs and corresponding snow load levels.