Conceptual weather environmental forecasting system for identifying potential failure of under-construction structures during typhoons

- CWEFS is developed for predicting the failure of under-construction structures during typhoons.
- Data-driven models, SVRs, regression, C5.0 and CART, and Holland model are presented.
- Structure reference load analysis is performed to evaluate the reference load.
- SVRs provides excellent prediction accuracy compared with regression, C5.0, and CART in all structure scenarios.

In this study, we developed a conceptual weather environmental forecasting system (CWEFS) for predicting the failure of under-construction structures during typhoons. Major functions of the developed system include 1) forecasting hourly typhoon wind velocity, 2) analyzing structure reference load during the construction stage, 3) identifying potential failure of under-construction structures, and 4) evaluating the weather in future hours to determine whether the conditions are suitable for work. Data-driven models, namely support vector machines for regression (SVRs), regression, and two decision trees (namely C5.0 and CART) were employed in this study as forecasting techniques to predict the wind velocity on Orchid Island, Taiwan, the study site. Structure reference load analysis was performed using a finite element model to evaluate the reference load on an experimental tank under construction. Typhoons Nanmadol (2011) and Saola (2012) were selected for real-time simulation by using the proposed CWEFS. This study identified potential collapses by using 1- to 6-h-ahead wind speed predictions. However, prediction errors inevitably occur. The results showed that the SVRs provided excellent prediction accuracy compared with regression, C5.0, and CART regarding the average time error between the observed and predicted values in all structure scenarios. A high forecast time error might result in increased construction costs and delays in construction schedules. Thus, we suggest that shorter prediction windows (e.g., 1 and 2 h) and models with higher prediction accuracy (e.g., SVR and C5.0) be employed to create a reasonable warning system.