Exploring the best sequence LSTM modeling architecture for flood prediction

Accurate and efficient models for rainfall–runoff (RR) simulations are crucial for flood risk management. Recently, the success of the recurrent neural network (RNN) applied to sequential models has motivated groups to pursue RR modeling using RNN. Existing RNN based methods generally use either sequence input single output or unsynced sequence input and output architectures. In this paper, we propose a synced sequence input and output long short-term memory (LSTM) network architecture for hydrologic analysis and compare it to existing methods (sequence input single output LSTM). We expect the model will improve RR prediction in terms of accuracy, calibration training time, and computational cost. The key idea is to efficiently learn the long term dependency of runoff on past rainfall history. To be more specific, we use the indigenous ability of the LSTM network to preserve long term memory instead of artificially setting a time window for input data. In this way, we can avoid losing long term memory of the input, the calibration of the time window length, and excessive computation. The whole procedure mimics the traditional process-driven methods and is closer to the physics interpretation of the RR process. We conducted experiments on real-world hydrologic data from the Brays Bayou in Houston, Texas. Extensive experimental results clearly validate the effectiveness of our proposed method in terms of various statistical and hydrological related evaluation metrics. Notably, our experiment shows that some rainfall events could affect the runoff process in the test watershed for at least a week. For fine temporal resolution prediction, this long term effect needs to be carefully handled, and our proposed method is superior in this case.