Models and Predictions for "Rainfall-Runoff Prediction at Multiple Timescales with a Single Long Short-Term Memory Network"

Models and Predictions for the paper "Rainfall-Runoff Prediction at Multiple Timescales with a Single Long Short-Term Memory Network" GitHub: https://github.com/gauchm/mts-lstm 2026 update: The pickle files in this repository rely on outdated versions of pandas, xarray, and other libraries and have gotten difficult to read. We have therefore converted them to zarr files in `results-zarrs.tar.gz` which contain the same results. `example_zarr_loader.ipynb` describes how to use the zarr files. Thanks to Grey Nearing for the conversion! Results The file `results.tar.gz` contains: ensembled predictions for all models (generated from the models in `models/` using the `nh-results-ensemble` command). These predictions were used in the `results-analysis.ipynb` and `odelstm-analysis.ipynb` notebooks on the GitHub repository for the paper. the NWM predictions `nwm_chrt_v2_1h.p` contains hourly NWM predictions for the CAMELS basins between 1993 and 2007. The file is derived from the reanalysis on aws. `nwm_results.p` is derived from `nwm_chrt_v2_1h.p` and contains hourly and day-aggregated results and performance metrics for the test period of our paper. a file `signatures.p` with hydrologic signatures that were calculated from the models' predictions. These signatures were used in the `results-analysis.ipynb` notebook on the GitHub repository for the paper. Models The tar.gz files prefixed with `models-` contain the trained MTS-LSTM, sMTS-LSTM, and ODE-LSTM models from our experiments. For each experiment, there exist 10 model setups (one for each random seed).Besides the trained models, each model's tar.gz also contains the predictions on the test or validation perod and the configuration file used to train the model. MTS-LSTM `mtslstm_seed*` -- the MTS-LSTM from the benchmarking section of the paper (using one forcings product, trained on daily and hourly data) `mtslstm_multiforcing_seed*` -- the MTS-LSTM from the section on per-timescale input data, experiment "multi-forcing B" (using just NLDAS as hourly inputs) `mtslstm_multiforcing_dailyhourly_seed*` -- the MTS-LTSM from the section on per-timescale input data, experiment "multi-forcing A" (ingesting daily forcings into the hourly model) `mtsltsm_136H1D_seed*` -- the MTS-LTSM from the section on prediction at other timescales (1-, 3-, 6-hourly and daily predictions) sMTS-LSTM `smtslstm_seed*` -- the sMTS-LSTM from the benchmarking section of the paper (using one forcings product, trained on daily and hourly data) `smtslstm_noregularization_seed*` -- the sMTS-LSTM from the section on cross-timescale consistency (trained without regularization) Time-Continuous Experiments The file `models-timecontinuous.tar.gz` contains one sub-folder per basin on which we conducted our initial experiments.Each basin directory contains: Experiment A (trained on daily and 12-hourly, evaluated on hourly): `odelstm_a_seed*` -- the ODE-LSTM from experiment A `mtslstm_a_seed*` -- the MTS-LSTM from experiment A Experiment B (trained on hourly and 3-hourly, evaluated on daily) `odelstm_b_seed*` -- the ODE-LSTM from experiment B `mtslstm_b_seed*` -- the MTS-LSTM from experiment B Related Datasets: https://doi.org/10.5281/zenodo.4072700 contains the hourly NLDAS forcings and USGS streamflow required to use the models from this dataset.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.