Urban real-time rainfall-runoff prediction using adaptive SSA-decomposition with dual attention

Document Type

Article

Publication Date

6-2025

Department

Department of Computer Science

Abstract

Urban real-time rainfall-runoff prediction is a complex task in hydrological simulation due to the strong nonlinearity and fluctuation inherent in urban rainfall-runoff processes. Recently, decomposition based deep-learning (DD) frameworks have gained popularity for significantly improving runoff prediction accuracy. However, most of these DD frameworks treat the entire data sequence as a single decomposition unit, which brings the risk of data leakage, limiting their real-time practical applicability. Additionally, these DD frameworks also neglected careful consideration of how sub-sequences obtained from decomposition should be extracted in subsequent deep-learning (DL) modules. To address these limitations, this investigation proposes adaptive singular spectrum analysis (SSA) decomposition with dual attention (ASDA) model. This model first segments the entire rainfall-runoff sequence into many short decomposition units, thus effectively eliminating the risk of data leakage; then we design 3 filters based on adaptive SSA that can adaptively decompose each such short unit into trend, fluctuation and noise sub-sequences by dynamically calculating singular values of embedding matrix. As for DL-based feature extraction procedure, dual attention mechanism that consists of a back-query attention (BQA) and a trans self-attention (TSA) is proposed, BQA dynamically assess the contributions of trend and fluctuation sequences before LSTM-based main feature extractors, and TSA interactively fuse the two extracted features after extractors. Additionally, soft dynamic time warping (soft-DTW) is incorporated into the loss evaluation to better guide ASDA model training. Performance of the proposed ASDA is validated by our observed urban rainfall events from January 2018 to December 2019 in a complex terrain covering 3.52 km2 in Chongqing, China. Key experimental results outperform conventional DL and machine-learning models: LSTM, GRU, RNN, TCN, Transformer and LightGBM in terms of NSE, KGE, RMSE, MAE and Pbias coefficients, providing a potential URRP approach for water-resource management in urban areas.

Publication Title

Journal of Hydrology

Share

COinS