A Novel Attention-Based LSTM Cell Post-Processor Coupled with Bayesian Optimization for Streamow Prediction

Abstract

Abstract Streamow forecasting is critical for real-time water resources management and ood early warning. In this study, we introduce a novel attention-based Long-Short Term Memory (LSTM) cell deep learning (DL) model for postprocessing streamow simulations which will herein be referred to as Self-activated and Internal Attention LSTM, or SAINA-LSTM. In this model, we incorporate an improved self-attention mechanism into the inner structure of the LSTM cell to increase focus on the more important time points, thereby enhancing the information ow of the cell. Performance of the SAINA-LSTM is then compared against that of the current NationalWeather Service s operational streamow forecast ensemble postprocessor (EnsPost), a recently-developed multiscale alternative (MS-EnsPost), a robust machine learning algorithm (Gradient Boosting), and two other deep learning algorithms (LSTM and Gated Recurrent Unit (GRU)). Forecast performance in four basins in di_erent climatological regimes of the United States are compared. Several deterministic evaluation metrics are examined for one to seven-day-ahead predictions of daily ows. SAINA-LSTM reduce the biases in simulations of low, medium, and high daily ows. The other two deep learning models also outperform EnsPost and MS-EnsPost in most cases. The results highlight the relatively poor performance of EnsPost, particularly in low ow conditions. Generally, the novel SAINA-LSTM model outperforms other models in low, medium, and high ranges of ow and for 1- to 7-day ahead forecasts in all three highly nonlinear and non-snow-driven study basins. In snow-driven basins, due to low nonlinearity, the three deep learning models are relatively comparable and signi_cantly improved over statistical models. The results of the comparative evaluation demonstrate the capability of SAINA-LSTM to reduce the RMSE of daily-predicted ow by up to 20 percent compared to MSEnsPost. The promising result shown here is a motivation for extending this research to also cover ensemble forecasting using the novel model developed in this work.