PSI - Issue 73

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

143

6

melt or clear conditions.

Negligible in this case

Fig. 2 Mean absolute SHAP values for all features, illustrating their overall contribution magnitudes to the roof snow load model predictions.

Fig. 3d shows that lower thermal radiation fluxes (~200– 280 W/m²) increase snow load predictions, reflecting radiative cooling and snow preservation under clear skies. Higher values (>300 W/m²) are associated with melt conditions and negative SHAP values, especially when snow depth is low (blue points). Fig. 3e explores the interaction between air temperature and wind speed. Near the melting point (−5 °C to +5 °C), higher winds (red points) reduce SHAP values compared to low-wind conditions (blue), indicating enhanced snow loss via sublimation or scouring. Below −10 °C, wind has little impact as the cold preserves snow regardless. Fig. 3f flips this interaction: SHAP values for wind speed are coloured by air temperature. Cold-air conditions (blue) moderate the negative SHAP impact of high winds, while warm-air conditions (red) enhance snow removal effects. The model captures this joint dependency, aligning with physical behaviour of snowpack dynamics under different thermal-wind regimes. Together, these plots confirm that the ML model learns physically plausible relationships: snow accumulation is favoured by cold, calm, low-pressure conditions; melting and snow loss are driven by warmth, wind, and irradiance—modulated by interactions among variables.