Deep Temporal Neural Networks for Water Level Predictions of Watershed Systems

Rainfall-runoff systems are complex hydrological environments that play a critical role in flood prevention. Currently, physics-based, process-driven computational models are often used to forecast future flooding events. However, these physics-based models are computationally expensive and require intensive physical measurements of hydrological environments beyond remote data collection. There is a growing body of literature that applies deep neural networks to time-series data for computationally efficient, real-time flooding predictions without the need for the complete virtual modeling of the hydrological system. However, these deep-learning networks’ robustness at forecasting far into the future remains an open question. In this study, we examine the capabilities of Long Short-Term Memory (LSTM) networks and Temporal Convolutional Networks (TCN), state-of-the-art temporal deep neural networks, to forecast rainfall-runoff system depths. Specifically, this study leverages primary, multi-modal, time-series data collected by remote sensors in the watershed system of Conner Creek, a tributary of the Clinch River in eastern Tennessee. These data were collected in 5-minute intervals over a course of 5 months. Notably, the Conner Creek watershed system consists of four interconnected reservoir basins. We forecast the water level of each reservoir basin independently for times ranging from five minutes to two hours into the future. Our results show that both the LSTM and TCN can effectively model and forecast future reservoir basin water levels. Specifically, when averaged across the four reservoir basins, the LSTM has an mean absolute error (MAE), with a 95% confidence interval, of 0.158 ± 0.049 ft and 0.490 ± 0.260 ft at 5 minutes and 120 minutes into the future, respectively. In comparison, the TCN has an MAE of 0.258 ± 0.160 ft and 0.375 ± 0.245 ft at 5 minutes and 120 minutes into the future, respectively. Our results show that the LSTM model outperforms the TCN for near lead time forecasting; however, the TCN retains a greater relative accuracy at larger lead time forecasting periods (two hours). Nevertheless, both models can be considered effective at capturing future trends of watershed systems, demonstrating them to be powerful tools for use in flood risk management systems.