Rainfall-Runoff Prediction at Multiple Timescales with a Single Long Short-Term Memory Network (2010.07921v1)

Abstract: Long Short-Term Memory Networks (LSTMs) have been applied to daily discharge prediction with remarkable success. Many practical scenarios, however, require predictions at more granular timescales. For instance, accurate prediction of short but extreme flood peaks can make a life-saving difference, yet such peaks may escape the coarse temporal resolution of daily predictions. Naively training an LSTM on hourly data, however, entails very long input sequences that make learning hard and computationally expensive. In this study, we propose two Multi-Timescale LSTM (MTS-LSTM) architectures that jointly predict multiple timescales within one model, as they process long-past inputs at a single temporal resolution and branch out into each individual timescale for more recent input steps. We test these models on 516 basins across the continental United States and benchmark against the US National Water Model. Compared to naive prediction with a distinct LSTM per timescale, the multi-timescale architectures are computationally more efficient with no loss in accuracy. Beyond prediction quality, the multi-timescale LSTM can process different input variables at different timescales, which is especially relevant to operational applications where the lead time of meteorological forcings depends on their temporal resolution.

• The paper proposes and evaluates two Multi-Timescale LSTM architectures (sMTS-LSTM and branched MTS-LSTM) designed for efficient rainfall-runoff prediction across different temporal scales.
• These MTS-LSTM models demonstrated superior performance over the US National Water Model and naive LSTM approaches in predicting discharge and peak timing across diverse hydrologic regimes.
• The research presents a viable pathway for operational hydrologic models by effectively integrating inputs of varying temporal resolutions and offers a flexible framework applicable to other multi-timescale prediction problems.

Analysis of Multi-Timescale LSTM Architectures for Rainfall-Runoff Prediction

The paper "Rainfall--Runoff Prediction at Multiple Timescales with a Single Long Short-Term Memory Network" presents a comprehensive examination of two distinct Multi-Timescale Long Short-Term Memory (MTS-LSTM) architectures for hydrological modeling, demonstrating their efficacy in predicting discharge at multiple temporal scales. The research addresses a critical need for hydrologic applications that require predictions on granular timescales, such as those necessary for flood forecasting.

Overview and Methodology

The paper develops two architectures: the shared MTS-LSTM (sMTS-LSTM) and a distinct branched MTS-LSTM. Both models aim to retain computational efficiency by utilizing past input data at a coarse resolution while refining predictions at finer timescales for more recent inputs. This approach addresses the computational inefficiency of training an LSTM on long sequences of high-resolution input data.

The research leverages data from 516 basins across the continental United States, using the CAMELS dataset augmented with hourly meteorological forcings from the NLDAS-2 dataset. It benchmarks these models against the US National Water Model (NWM) and a naive LSTM trained separately for each timescale. Through rigorous evaluation criteria, including the Nash--Sutcliffe efficiency (NSE) and peak-timing metrics, the paper highlights the models' performance across multiple hydrologic regimes.

Key Insights and Results

Significantly, the MTS-LSTM models outperformed the NWM across nearly all evaluation metrics, with the sMTS-LSTM often yielding the highest median NSE values. The models demonstrated the ability to produce consistent and accurate predictions at both hourly and daily scales, a notable improvement in computational efficiency compared to naive approaches. With median NSE ranges for the MTS-LSTM architectures showing a clear advantage over process-based models like the NWM, these LSTM variants promise enhanced accuracy in hydrologic predictions.

Moreover, the utilization of cross-timescale regularization reduced prediction inconsistencies significantly. This regularization not only improved consistency but also provided a slight enhancement in prediction accuracy, especially for the sMTS-LSTM.

Implications for Future Research

The practical implications of this research are manifold. Primarily, it proposes a viable pathway for operational hydrologic models by incorporating inputs of varying temporal resolutions effectively. The feature of ingesting different input sets for different timescales stands out as a valuable asset for real-world hydrologic forecasting applications, where data may be available at varying resolutions. Furthermore, this work lays the foundation for future extensions into multi-objective optimization within hydrological systems and suggests methodologies that could effectively manage the computational burden typically associated with high-resolution temporal predictions.

In terms of theoretical implications, the architecture of the two MTS-LSTM variants presents a flexible methodological framework that can be adapted across other domains requiring predictions at multiple temporal scales. This opens avenues for exploring LSTM capabilities further in related applications, potentially extending into other domains such as environmental monitoring and resource management.

Conclusion

By providing a robust analysis of MTS-LSTM architectures for multi-timescale hydrologic modeling, the paper presents a compelling case for the integration of advanced deep learning techniques in environmental applications. The findings underscore the advantage of such models over traditional process-based approaches, particularly in terms of computational efficiency and predictive accuracy across diverse timescales. Future research inspired by this work is likely to focus on refining these models further and expanding their applicability, both in hydrology and other fields requiring sophisticated temporal predictions.