PSI - Issue 73

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

145

8

5. Conclusion This paper presented a data-driven method to estimate snow loads on a heated roof (U = 1.0 W/m²K) using machine learning. A 31-year hourly dataset for Oslo, combining ERA5 climate data and a snow simulation model, was used to train four ensemble models: RF, GBM, XGBoost, and CatBoost. For the case of no sliding, all models achieved high accuracy (test R 2 ≈ 0.996 –0.997) and low error, closely matching the timing and magnitude of snow accumulation and melt. Key predictors included ground snow load, snow depth, air and soil temperatures, surface pressure and thermal radiation —aligning well with physical expectations. The impact of non-linear and interaction effects on the prediction of roof snow load was also revealed with local SHAP analysis. The results confirm that ML can effectively learn the complex drivers of roof snow load, offering a fast alternative to physical simulations. For instance, an XGBoost model can predict roof loads instantly from weather forecasts, enabling proactive snow management. Focusing on the U = 1.0 case demonstrated that heated roofs carry significantly less snow than the ground, and the ML models captured this reduction accurately. Future work will extend the approach to other roof insulation levels (e.g., U = 0.5–2.0 W/m²K), incorporate roof geometry and snow sliding dynamics, and potentially unify these into a general model. Integration with structural reliability analysis could enable probabilistic load assessment. Ultimately, validating the models with real-world measurements (e.g., from instrumented roofs) would enhance practical applicability and support updates to code based snow load coefficients. Acknowledgements This study has been supported by the Czech Science Foundation under Grants 23-06222S (roof snow modelling) and 24-10892S (ML modeling). The involvement of Dr Lenganji Simwanda in this research has been supported by the Global Postdoctoral Fellowship Program of the Czech Technical University in Prague. References Chen, T., Guestrin, C., 2016. XGBoost: A scalable tree boosting system, in “Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining”. ACM, San Francisco, pp. 785–794. https://doi.org/10.1145/2939672.2939785. Croce, P., Formichi, P., Landi, F., Mercogliano, P., Bucchignani, E., Dosio, A., Dimova, S., 2018. The snow load in Europe and the climate change. Climate Risk Management 20, 138–154. https://doi.org/10.1016/j.crm.2018.03.001. Croce, P., Formichi, P., Landi, F., 2021a. Extreme ground snow loads in Europe from 1951 to 2100. Climate 9, 1–20. https://doi.org/10.3390/cli9090133. Croce, P., Formichi, P., Landi, F., 2021b. Probabilistic assessment of roof snow load and the calibration of shape coefficients in the Eurocodes. Applied Sciences 11, 1–16. https://doi.org/10.3390/app11072984. He, Y., Yan, X., Li, X., 2022. Numerical simulation of snowdrift on an air-supported membrane structure and response analysis under snow loads. International Journal of Space Structures 38, 4–19. https://doi.org/10.1177/09560599221108624. Hersbach, H., et al., 2020. The ERA5 global reanalysis. Quarterly Journal of the Royal Meteorological Society 146, 1999–2049. Lundberg, S.M., Lee, S.-I., 2017. A unified approach to interpreting model predictions, in “Advances in Neural Information Processing Systems 30”. In: Guyon, I., Von Luxburg, U., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (Eds.), Curran Associates, Inc. Madirisha, M.M., Simwanda, L., Mtei, R.P., 2025. Predicting the hydrogen storage capacity of alumina pillared interlayer clays using interpretable ensemble machine learning. International Journal of Hydrogen Energy 120, 354–364. https://doi.org/10.1016/j.ijhydene.2025.03.216. Simwanda, L., Gatheeshgar, P., Ilunga, F.M., Ikotun, B.D., Mojtabaei, S.M., Onyari, E.K., 2024. Explainable machine learning models for predicting the ultimate bending capacity of slotted perforated cold-formed steel beams under distortional buckling. Thin-Walled Structures 205, 112587. Turner, R., Eriksson, D., McCourt, M., Kiili, J., Laaksonen, E., Xu, Z., Guyon, I., 2021. Bayesian optimization is superior to random search for machine learning hyperparameter tuning: Analysis of the black-box optimization challenge 2020, in “NeurIPS 2020 Competition and Demonstration Track”, pp. 3–26. Zhou, X., Chen, H., Wu, Y., et al., 2024. Simulation of roof snow loads based on a multi-layer snowmelt model: Impact of building heat transfer. Building Simulation 17, 907–932. https://doi.org/10.1007/s12273-024-1119-4.