Rainfall-Runoff modeling using transductive long short-term memory and hyperparameter optimization

Abstract

Water resource management and disaster risk reduction depend on accurate rainfall-runoff modeling (RRM). Time-series data frequently exhibits fine details and local variations that are difficult for traditional LSTM models to capture. To overcome these problems, we introduce an improved RRM model that utilizes a spatial attention-enhanced transductive LSTM (TLSTM) network. By using transductive learning on data points that closely resemble the test set, this model improves performance and captures subtle temporal differences. With the fusion of a spatial attention mechanism, the model can focus on the most important parts of the input. It also includes a more sophisticated differential evolution (DE) algorithm to facilitate complex hyperparameter tuning. For the DE algorithm, we used a mutation strategy that finds a significant cluster using K-means clustering. We applied the Catchment Attributes and Meteorology for Large Sample Studies (CAMELS) dataset and used individual and regional RRM throughout our assessment of the model. There were very positive results for individual basins with an 8-day runoff prediction of 0.728 Nash-Sutcliffe efficiency (NSE). For regional assessment, the model had an NSE of 0.878. This method supports that combining TLSTM with spatial attention and sophisticated DE algorithms increases the accuracy and reliability of rainfall-runoff (RR) predictions, providing opportunities for improved planning and disaster management of water resources. The source code is publicly available at https://github.com/ZhaoChenchina/RRM.