An end-to-end deep learning framework for multi-timescale, probabilistic sub-seasonal to seasonal streamflow prediction

Rainfall-runoff modeling has historically relied on conceptual models hindered by calibration dependencies and limited applicability in ungauged regions. While deep learning, specifically Long Short-Term Memory (LSTM) networks, offer a powerful alternative, sub-seasonal/seasonal streamflow predictions remain constrained by a forcing bottleneck, where long-term hydrological accuracy is fundamentally bounded by the rapid degradation of meteorological forecasts. Furthermore, handling structural uncertainty of precipitation, specifically its intermittent nature and highly skewed distribution, poses significant challenges. This thesis addresses these limitations by developing an end-to-end probabilistic forecasting framework. First, a Climatology-Guided Hierarchical State-Space LSTM utilizing a Zero-Inflated Log-Normal (ZILN) head is developed to generate ensemble meteorological forecasts by integrating global climate indices with local weather data. Second, these dynamic forecasts are coupled into a regional Multi-Timescale LSTM with Countable Mixture of Asymmetric Laplacians (MTS-LSTM-CMAL) to predict streamflow across 516 catchments in the United States. A rolling hindcast evaluation is utilized to assess the predictive limits of this framework. Results indicate that while the meteorological model successfully leverages short-term persistence, extended horizon forecasts converge to climatology, with global climate indices providing only marginal skill improvements. Hydrologically, continuous time-series evaluations reveal high predictive skill. However, performance degradation is heavily dictated by physical regimes. Snowmelt-driven basins maintain extended predictive skill due to catchment memory, whereas rain-driven catchments degrade rapidly alongside meteorological forcing. Furthermore, correlation analysis reveals the framework excels in dynamic, surface-driven catchments but struggles significantly in groundwater-dominated systems, highlighting a core limitation of the LSTM’s finite look-back window in capturing deep subsurface memory. These findings demonstrate vast potential of deep learning for regional, probabilistic streamflow forecasting while exposing critical architectural limitations regarding latent state representations and sub-seasonal/seasonal meteorological predictability.