Timeseries forecasting use case: level of water

Based on Open Data provided by Hub'Eau and largely inspired by https://www.tensorflow.org/tutorials/structured_data/time_series

Predicting the level of water in rivers can prove to be invaluable in areas with a high risk of flooding. While traditional methods are based primarily on physical model, another approach is emerging: artificial neural networks.

This experiment is using TensorFlow and Long Short-Term Memory Recurrent Neural Networks. We would also like to have a look to https://facebook.github.io/prophet/.

Background: research has shown that some weather parameters can be accurately predicted out to 72hrs using a 15 year data period of hourly measurements, why not water level ?

Data gathering

Simply run yarn install then node hubeau_data.js "River label", it will gather all required data for the experiment on the target river in the data directory:

• list of sites (ie measurement stations) of interested for the target river (eg "L'Aude")
• bydefault 1 month past data of measurements on each previously selected site with a time step of 30 minutes

You can also use the following additional options:

• --output output directory (defaults to 'data')
• --timestep time step in minutes (min 10 - max 60 - defaults 30)
• --timerange past time range in days (min 1 - max 30 - defaults 30)

All data zip files are to be extracted into the 📂 data/rawdirectory

Base data

Default data for training (train_data.zip) are provided from 07-01-2020 to 07-02-2020 for some rivers in Occitanie - France, which is an interesting period because a flood episode has occured from 22-01 to 25-01.

Some validation data (validation_data.zip) are also provided from 15-04-2020 to 15-05-2020 for some rivers in Occitanie - France, which is also an interesting period because smaller but intense precipitations episodes have occured.

Please note that data at some stations were missing during this period

Concerning the meteo station data (train_meteo_data.zip and validation_meteo_data.zip) we retrieved archived data on these periods from https://rp5.ru/ manually but data are really scrace.

Complementary data

Additional data (test_data.zip) are provided from 26-08-2020 to 26-09-2020 for some rivers in Occitanie - France, which is also an interesting period because a huge flood episode has occured from 19-09 to 20-09.

During this flood episode we gathered meteo station data (test_meteo_data.zip) from the dense RADOME network.

However, we did not make use of this data yet. Let us know if you have by opening an issue.

Data processing and model training

There are two different notebooks located in the 📂 notebooks folder:

• hubeau_univariate where a single variable, ie the water level given by the station, is used at a given station to make predictions
• hubeau_multivariate where multiple variables, ie the water level given by others stations or meteo stations, is used at a given station to make predictions - Work in progress

We use Anaconda distribution for simplicity and create a TensorFlow environment for our experiment:

conda create -n tf-gpu tensorflow-gpu
conda activate tf-gpu
conda install matplotlib
conda install pandas
conda install jupyter notebook
conda install papermil

Launch the notebooks using jupyter notebook hubeau_univariate.ipynb or jupyter notebook hubeau_multivariate.ipynb.

Execute the notebooks with parameters using papermill hubeau_univariate.ipynb test.ipynb -p timestep 60 -p model_type 'LSTM' -p stacked True -p epochs 20.

You might face this issue with papermill, resolve it with pip install nbformat --upgrade.

License

Copyright (c) 2017-20xx Kalisio

Licensed under the MIT license.