A dynamic classification-based long short-term memory network model for daily streamflow forecasting in different climate regions

Abstract

Daily streamflow forecasting is a major determinant of ecological processes in running waters, healthy stream ecology and surrounding environment, and accurate streamflow forecasting provides a powerful foundation for ecological assessment, management, and decision-making. Recently, data-driven models for different flow regimes have shown excellent potential in streamflow forecasting. However, the boundaries between different flow regimes were selected arbitrarily without considering the changes in boundaries that often occur over time in the real world. Therefore, in this paper, an integrated modelling approach that couples a dynamic classification method with a long short-term memory networks (LSTM) model without data transformation (the DC-LSTM model) and an LSTM with Box-Cox data transformation (the DC-B-LSTM model) is developed to improve the performance of streamflow forecasting considering different flow regimes. The boundaries of dynamic classification are dynamic changing interval values of related hydrological variables improved from traditional classification method just using static single-variable threshold, so dynamic classification can more fully explore the relationship and information of hydrological data. The performance of both the DC-LSTM and DC-B-LSTM models is compared to that of the LSTM model without data classification (the traditional LSTM model) and with data classification using a traditional static method (the C-LSTM model) based on data from 8 stations within 4 river basins in different climate regions in the United States. The results show that both the DC-LSTM and DC-B-LSTM models out-perform the traditional LSTM models (with or without static data classification) for all river basins considered. Furthermore, the DC-B-LSTM model displays better performance than the DC-LSTM model in arid areas.