Neural Networks for Estimating Storm Surge Loads on Storage Tanks

Abstract

Failures of aboveground storage tanks (ASTs) during past storm surge events have highlighted the need to evaluate the reliability of these structures. To assess the reliability of ASTs, an adequate estimation of the loads acting on them is first required. Although finite element (FE) models are typically used to estimate storm surge loads on ASTs, the computational cost of such numerical models can prohibit their use for reliability analysis. This paper explores the use of computationally efficient surrogate models to estimate storm surge loads acting on ASTs. First, a FE model is presented to compute hydrodynamic pressure distributions on ASTs subjected to storm surge and wave loads. A statistical sampling method is then employed to generate samples of ASTs with different geometries and load conditions, and FE analyses are performed to obtain training, validation, and testing data. Using the data, an Artificial Neural Network (ANN) is developed and results indicate that the trained ANN yields accurate estimates of hydrodynamic pressure distributions around ASTs. More importantly, the ANN model requires less than 0.5 second to estimate the hydrodynamic pressure distribution compared to more than 30 CPU hours needed for the FE model, thereby greatly facilitating future sensitivity, fragility, and reliability studies across a broad range of AST and hazard conditions. To further highlight its predictive capability, the ANN is also compared to other surrogate models. Finally, a method to propagate the error associated with the ANN in fragility or reliability analyses of ASTs is presented.