Example Configuration Files

The ~/flood-forecasting/example-configs directory contains reference configuration files in YAML format. These files define the experimental setups for different model architectures and datasets used in Google’s production forecasting systems.

Overview

The configurations provided here serve two primary purposes:

  • Operational Replication: Replicating the exact or near-exact settings used by the Google FloodHub production models.

  • Benchmarking: Providing stable baselines for model development and evaluation on standard datasets like CAMELS-US.

Configuration Files

File Name

Model Architecture

Description

floodhub-settings-config.yml

mean_embedding_forecast_lstm

Near-identical replication of the current (2025) production FloodHub model.

handoff-forecast-lstm-config.yml

handoff_forecast_lstm

Configuration for the former production model and methodology described in the Nature (2024) paper.

camels-multimet-mean-embedding-forecast-lstm-config.yml

mean_embedding_forecast_lstm

Optimized for benchmarking on the CAMELS-US (531 basins) dataset. Used as a core development reference.

camels-multimet-handoff-forecast-lstm-config.yml

handoff_forecast_lstm

Benchmarking configuration for the State Handoff model tailored for the CAMELS-US (531 basins) dataset.

Operational Models

FloodHub 2025 (Mean Embeddings)

The floodhub-settings-config.yml file defines the current operational standard. It utilizes the mean_embedding_forecast_lstm model, which handles multiple meteorological products by averaging their latent representations.

FloodHub 2024 (State Handoff)

The handoff-forecast-lstm-config.yml file provides the settings for the previous operational generation. It focuses on transferring internal LSTM states (hidden and cell states) from a hindcast period to a forecast period via the handoff_forecast_lstm architecture.

Benchmarking & Development

The configurations prefixed with camels-multimet- are specifically designed for testing on the CAMELS-US dataset. These CAMELS experiments are used internally by the Google team to verify that system updates do not degrade model performance.