PSI - Issue 73

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

142

5

Fig. 1 Predicted vs. actual roof snow load (kN/m²) for the training (blue) and testing (red) sets across all four models: (a) Random Forest, (b) GBM, (c) CatBoost, (d) XGBoost. The black line indicates the 1:1 perfect prediction reference. All models exhibit tight clustering around the diagonal, confirming excellent predictive alignment. 4.2. Feature Importance and Interpretation Interpretability of the ML model is achieved through SHAP (Shapley Additive Explanations) analysis, which provides insight into how each input influences the roof snow load predictions. The SHAP results not only rank the most important features driving the model but also reveal nonlinear effects and interactions between meteorological variables. This section presents a concise interpretation of the SHAP findings, emphasizing physical consistency. Key features such as ground snow load, snow depth, temperature, radiation, and wind speed show characteristic SHAP patterns that align with known snow accumulation and melt behavior. Such interpretability enhances trust in the model by demonstrating that its predictions are based on real-world weather relationships rather than opaque correlations. Fig. 2 shows the ranking of the features according to SHAP. The model identifies Ground Snow Load and Snow Depth as the top influences by a large margin (mean |SHAP| of 11.3 and 5.6, respectively), followed by temperature related features like subsurface Soil Temperature (Level 3 at 0.3 m depth) and 2 m Air Tempera ture (0.86). Radiative and wind-related features have smaller average contributions (mostly 0.04–0.35). This ranking confirms that the amount of available snow (on the ground) and ambient thermal conditions are the primary drivers of roof snow load predictions, aligning with physical expectations (i.e. more snow and colder weather lead to higher loads). Fig. 3 illustrates nonlinear effects and feature interactions in the ML model using SHAP dependence plots. Fig. 3a shows that air temperature has a sharp nonlinear effect on snow load predictions. SHAP values are highly positive below 0 °C (e.g. +4 to +5 at −20 °C), indicating snow accumulation under cold conditions. Above freezing, SHAP values become negative, reflec ting rapid melting. The steep transition at 0 °C captures the physical shift from accumulation to ablation, with high ground snow loads (red points) amplifying the effect in subzero temperatures. Fig. 3b reveals that 10 m wind speed has a non -monotonic influence. Calm winds (~0– 2 m/s) have little effect, while moderate winds (~3– 5 m/s) can slightly increase snow load due to drifting, especially when ground snow is abundant. At higher wind speeds (>6 m/s), SHAP values turn negative, suggesting wind-induced snow removal dominates. More investigations in this regard will follow. Fig. 3c highlights surface pressure as a proxy for storm activity. Low-pressure systems (~94– 96 kPa) strongly increase predicted snow load (SHAP ≈ +1.5), consistent with snowfall events, whereas high- pressure systems (>101 kPa) yield negative SHAP values (≈ −2.0) , indicating snow