This repository contains a Python script and configuration file to train, evaluate, and visualize a model that predicts future values of hydrological time-series at the Geneve, Halle de l'Ile station using a multilayer perceptron (MLP).

The model learns to forecast a fixed number of hours into the future based on:

• Historical station measurements
• Time-of-day, day-of-month, and month features

The script performs the following steps:

1. Reads hyperparameters and paths from config.toml
2. Loads time-series data from an HDF5 file
3. Builds a MLP model in PyTorch
4. Trains the model (or loads an existing one)
5. Predicts values n_hours_prediction into the future
6. Evaluates performance on validation/test data
7. Saves:
   • The trained model
   • Loss curves
   • Prediction vs truth plots
   • Diagnostic statistics

The model predicts a vector of future values (e.g., the next 24 hours) for multiple hydrological stations simultaneously.

[model]
dim_hidden = 1024
n_layers = 8
activation = "LeakyReLU"

[training]
epochs = 1000
batch_size = 512
learning_rate = 0.001
patience = 50
min_delta = 0.001
loss = "HuberLoss"
weight_decay = 0

[run]
seed = 2234
gpu = 1
data_fraction = 1
train_split = 0.8
valid_split = 0.1
n_hours_prediction = 24
read_model = 0
station_features = ["2009", "2606", "2174", "2170", "2027"]
time_features = ["Hour", "Day", "Month"]

[input_output]
input_data = "/path/to/data.hdf5"
output_folder = "/path/to/results"
model_name = "best_model"
overwrite = 1
n_check_plots = 100

The easiest way to get the code to work is to install a conda environment via

conda env create -f environment.yml

This will create a torch environment you can access by running

conda activate torch

Open the config.toml file and set:

input_data = "/absolute/path/to/your/data.hdf5" output_folder = "/absolute/path/where/results/should/be/saved"

If you created the conda environment, access it. Then, from the project directory, run:

python train_model.py config.toml

Once the code finishes, inside the output_folder the script will create the following nested structure

output_folder/
+-- config_used.toml
+-- best_model.pth
+-- feature_scaler.joblib
+-- target_scaler.joblib
|
+-- plots/
|   +-- Performance.pdf
|   +-- Target_vs_prediction.pdf
|
+-- std.txt
+-- mean_relative_difference_over_time.txt
+-- loss_values.txt