# Rainfall-Runoff Prediction at Multiple Timescales with a Single Long Short-Term Memory Network

**Authors:** Martin Gauch, Frederik Kratzert, Daniel Klotz, Grey Nearing, Jimmy Lin, Sepp Hochreiter  
**arXiv:** [2010.07921](https://arxiv.org/abs/2010.07921)  
**Date:** 2020-10-15  
**Summary page:** https://www.summarizepaper.com/en/paper/2010.07921v1/

## Key points

- Long Short-Term Memory Networks (LSTMs) are successful in predicting daily discharge
- Accurate forecasting of short but extreme flood peaks is required in many practical scenarios
- Naively training an LSTM on hourly data results in long input sequences that make learning difficult and computationally expensive
- Multi-Timescale LSTM (MTS-LSTM) architectures jointly predict multiple timescales within one model, processing long-past inputs at a single temporal resolution and branching out into each individual timescale for more recent input steps
- MTS-LSTM architectures were tested on 516 basins across the continental United States and benchmarked against the US National Water Model, showing computational efficiency without any loss in accuracy compared to naive prediction with a distinct LSTM per timescale
- 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.
- Authors of the study include Martin Gauch, Frederik Kratzert, Daniel Klotz, Grey Nearing, Jimmy Lin and Sepp Hochreiter.
- The paper titled "Rainfall-Runoff Prediction at Multiple Timescales with a Single Long Short-Term Memory Network" was published in Hydrology and Earth System Sciences journal in 2021.
- This research provides significant insights into improving flood forecasting accuracy by using MTS-LSTM models that can handle multiple timescales efficiently without compromising prediction quality or computational efficiency.

## Summary

The use of Long Short-Term Memory Networks (LSTMs) has been highly successful in predicting daily discharge. However, many practical scenarios require more granular predictions, such as accurate forecasting of short but extreme flood peaks that may escape the coarse temporal resolution of daily predictions. Naively training an LSTM on hourly data results in long input sequences that make learning difficult and computationally expensive. To address this issue, a study proposes two Multi-Timescale LSTM (MTS-LSTM) architectures that jointly predict multiple timescales within one model. These models process long-past inputs at a single temporal resolution and branch out into each individual timescale for more recent input steps. The MTS-LSTM architectures were tested on 516 basins across the continental United States and benchmarked against the US National Water Model. Compared to naive prediction with a distinct LSTM per timescale, the multi-timescale architectures are computationally more efficient without any loss in accuracy. Moreover, 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 authors of the study include Martin Gauch, Frederik Kratzert, Daniel Klotz, Grey Nearing, Jimmy Lin and Sepp Hochreiter. Their paper titled "Rainfall-Runoff Prediction at Multiple Timescales with a Single Long Short-Term Memory Network" was published in Hydrology and Earth System Sciences journal in 2021. This research provides significant insights into improving flood forecasting accuracy by using MTS-LSTM models that can handle multiple timescales efficiently without compromising prediction quality or computational efficiency.