Abstract

Forecasting the discharge of large rivers is a critical challenge for water resource management, navigation, flood prevention, and the operation of hydraulic infrastructure. Yet, building a robust forecasting pipeline from real-world data requires far more than simply training a machine learning model: you first need to identify relevant data sources, clean and harmonize them, and engineer appropriate explanatory variables.

This tutorial offers a comprehensive approach to forecasting daily river discharge for the Mississippi River at St. Louis, Missouri, with a one-day ahead horizon, using open data from the USGS and Copernicus ERA5.

The progression is deliberately incremental in order to measure the contribution of each step, each of which constitutes a section of this tutorial:

1. Construction of the reference hydrological time series from USGS data;
2. Initial univariate experiments using only the local discharge history;
3. Integration of upstream hydrological stations located on the Mississippi and Missouri rivers;
4. Multivariate forecasts leveraging these different stations simultaneously;
5. Construction of a meteorological dataset from ERA5 data and spatial aggregation of precipitation;
6. Integration of meteorological variables into the forecasting pipeline.

Beyond the performance achieved, this tutorial emphasizes methodology: respecting data chronology, preventing time leakage, building sliding windows, feature engineering, and comparing several modeling approaches.

It is aimed at students, data scientists, and hydrologists who wish to build step by step a reproducible forecasting pipeline applied to a real-world time series problem.