Improving Daily Streamflow Forecasting Systems in Data-scarce Regions with A Long Short-term Memory Model
Abstract
In addition to gathering valuable information about flood and drought control, streamflow forecasting is of high importance for activities related to hydro power production and water supply management. Accessibility to such forecasts for short-term and long-term is critical, especially in northern environments where average annual precipitation is not abundant and a large portion of water supply becomes available during the spring freshet. Recently, deep learning models, and especially long short-term memory (LSTM) networks, have been acknowledged as powerful methods for learning long term dependencies along with capturing temporal features of hydrological processes such as runoff. In this study, a LSTM network modeling approach is proposed for streamflow simulation in Canada. Inputs to LSTM models, which includes meteorological data and physiographic characteristics of basins, were collected from two sources: (i) a modified version of ERA5 reanalysis data through the HYSETS database for Canada and (ii) the CaSPAr dataset of Environment and Climate Change Canada. Meanwhile, for model development, observed data from various sources were put together. At first, the pre-trained LSTM streamflow models with CAMELS dataset over the conterminous United States was utilized to transfer pre-trained knowledge to selected basins in Canada. Then, the LSTM streamflow models were trained locally with available data for the targeted basins. During this presentation, we will present the progress of this study. The overarching goal of this study is to compare the performance of the LSTM modeling approach with a recently implemented hydrological forecasting system based on an Ensemble Kalman Filtering (EnKF) data assimilation scheme. Results of this study will contribute to develop forecasting systems in regions with limited hydrometeorological data.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.H23A..01K