PSI - Issue 73

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

140

3

Table 1. Key features from the Oslo climate dataset (hourly resolution) used for model training. Feature Unit Source

minimum 25%

50% fractile

75% fractile

maximum

fractile

10 m Wind Speed

m/s

ERA5

0

1.6

2.4

3.4

10.3

10 m Wind Direction

degrees

ERA5

0

83

185

259

359

10min Wind Speed at 10m height 2 m Relative Humidity

m/s

ERA5

0

2.8

4.4

6.2

16.8

%

ERA5 ERA5 ERA5

0.14 -27.2

0.68

0.84

0.92 12.6 11.1

1

2 m Temperature

°C °C

0

5.7 4.8

32.1 18.1

Soil Temperature Level 3 (at 0.3 m depth) Soil Temperature Level 4

-3.2

0.6

°C

ERA5

0.1

1.7

5.3

9.9

13.9

Surface Pressure

Pa

ERA5 ERA5

92000

97400

98200

99000

102000

Surface Solar Radiation

W/m²

0

0

4

154

841

Surface Thermal Radiation

W/m²

ERA5

146

256

296

327

417

Total Cloud Cover

(%)

ERA5

0

0.32

0.91

1

1

Total Precipitation

mm

ERA5 ERA5

0 0

0 0

0 0

0

7.9

Direct Solar Radiation

W/m²

54

749

Snow Depth (m)

mm

ERA5

0

0

0

11.3

401.5

Solar Azimuth Angle

radians

Derived from ISO52010 Derived from ERA5

-3.14

-1.51

0

1.65

3.14

Wind Power Capacity Factor (100 m) Diffuse Horizontal Irradiance

0 -1(fraction)

0

0

0.04

0.15

1

W/m²

ERA5

0

0

4

0.8

392

Ground Snow Load

mm SWE (converted to kN/m²)

SIMELT model

0

0

0

2.3

9.5

Roof Snow Load (U = 1.0, no sliding)

mm SWE (converted to kN/m²)

SIMELT model

0

0

0

0.1

2.7

• ERA5 data from Copernicus Climate Change Service were used as input (Hersbach et al., 2020). SIMELT is a snow accumulation/melt model used to estimate the roof snow load.

3. Machine Learning Implementation 3.1. Models selected and feature engineering

We used four ensemble machine learning models—RF, GBM, XGBoost, and CatBoost—to predict roof snow load from meteorological features. All are tree-based and well-suited for structured regression tasks with nonlinear relationships. Recent studies (Madirisha, Simwanda, and Mtei, 2025; Simwanda et al., 2024) have shown these models excel other models in terms of performance. RF reduces variance via bagging, while GBM builds trees sequentially to minimize prediction error. XGBoost extends gradient boosting with optimized algorithms for accuracy and efficiency, and CatBoost handles ordinal features effectively through the use of symmetric trees, offering greater stability (Chen and Guestrin, 2016). Models were implemented in Python using scikit-learn and official APIs. All models were initially trained on the full set of features from Table 1 (excluding the target – roof snow load). Key predictors included current weather and ERA5 ground snow depth, which provided useful memory of past accumulation and melt. Instead of explicitly engineering features like recent snowfall or melt hours, we relied on the