We propose a novel reduced-order model and examine its applicability to the complex three-dimensional turbulent wake of a generic square-backed bluff body called the Ahmed body at Reynolds number Re H = U∞H/ ν =9.2×104 (where U∞ is free-stream velocity, H the height of the body and ν viscosity). Training datasets are obtained by large eddy simulation (LES). The model reduction method consists of two components, a VGG-based hierarchical autoencoder (H-VGG-AE) and a temporal convolutional neural network (TCN). The first step is to map the high-dimensional flow attributes into low-dimensional features, namely latent modes, which are employed as the input for the second step. The TCN is then trained to predict the low-dimensional features in a time series. We compare this method with a TCN based on proper orthogonal decomposition (POD), which utilizes time coefficients as the input in the second part. It turns out that the H-VGG-AE has a lower reconstruction error than POD when the number of latent modes is relatively small in the first part. As the number of latent modes increases, POD exceeds in the performance of model reduction. However, the H-VGG-AE-based TCN is still more effective in terms of spatio-temporal predictions because it has a lower prediction error and costs much less time.