Urban pluvial flooding caused by extreme rainfall has been increasing globally, thus exacerbating loss of life and damage to property. Accurate and updated real-time forecasts are critical needs for urban flooding response and defense. Big data has opened new avenues for inundation depth prediction in complex urban settings. In this study, a new method for pluvial flood classification was proposed for an inundation depth change index (IDCI) by dividing floods into three types: pluvial persistent floods (PPFs), pluvial normal floods (PNFs) and pluvial flash floods (PFFs). Prediction models are identified for the three flood types using data from a network of sensors (109 rainfall stations and 80 flood water depth stations) in Shenzhen, China. The results show that backpropagation neural networks (BPNNs) and long short-term memory (LSTM) exhibit good performance in depth prediction but are not significantly different from one another. In addition, PPFs require a longer rainfall sequence to obtain a better forecast. The Nash-Sutcliffe efficiency (NSE) of the depth prediction results at all stations is 0.86. The prospects for generalizing this approach and its usage are discussed.