Cooling towers are wind-sensitive structures whose flow physics remains unclear, and intensive measurement is lacking. Hybrid models combining proper orthogonal decomposition with neural networks are introduced to study the wind loading on a large cooling tower using data from limited pressure sensors in wind tunnel tests. Eighteen training taps on a single level of the tower are recommended, which has similar performance as the 24-tap layout and outperforms the 12-tap layout by 25%–46% reductions in overall root mean square error. Hybrid models well predict the overall wind loading, with total determination coefficients of 0.99 and mean drag biases of 2%. Flow at the bottom and top levels is more three-dimensional, where the prediction performances are slightly worse than at the mid-levels. The single-layer determination coefficient is beyond 0.98 except for the first three layers. At most vertical and circumferential locations, the hybrid model with long short-term memory slightly outperforms the other two models using backpropagation neural networks.
Dong et al. (Sun,) studied this question.