Analysis of Machine Learning Methodologies for Hydrodynamic Modeling at a Local Scale

This project presents the Master's Thesis by Mirko Rupani, a student in the Master in Data Science program (2023/2024) at the Universidad de Cantabria and the Universidad Internacional Menéndez Pelayo. The thesis was co-directed by Professors Ana Julia Abascal Santillana and Rodrigo García Manzanas.

📖 Overview

The project involves high-resolution hydrodynamic modeling of the Bay of Santander using various machine learning techniques. The focus is on currents, sea level, temperature, and salinity.

Code Structure

The project adheres to object-oriented programming (OOP) principles, allowing for a modular and reusable structure. LowerCamelCase is used for naming variables, scripts, and other elements, except when calling external libraries that do not follow this convention.

Execution

The project structure is designed for flexibility and ease of use. All code is executed from a single main script (main.py), which takes parameters from a configuration file (config.json). This file contains all adjustable variables and parameters, allowing users to modify it for different case studies without changing the code directly.

Annex

The annex folder contains the obtained results, including metrics, figures, optimal hyperparameters, and other relevant outputs.

🛠 Configuration File (config.json)

The configuration file contains various parameters essential for the project's execution. Below is a detailed explanation of each section and its possible options:

randomState

• Description: Seed for random number generators to ensure reproducibility.
• Possible Values: Any integer (e.g., 42).

predictands

• predictandsFolder: Folder path for predictand data.
• station: Station ID number.
• hisFile: List of file paths for historical data.
• removeTimesteps: Number of timesteps to remove from the start.
• variables: List of variables to model (e.g., ["u_x", "u_y", "waterlevel"]).
• sigmaLayer: Sigma layer for vertical level.
• resample: Method for resampling data (e.g., "mean").

predictors

• predictorsFolder: Folder path for predictor data.
• wind: Source of wind data (e.g., "meteogalicia").
• hydro:
  • dataset_id: List of CMEMS dataset IDs for hydrodynamic data.
  • point: Coordinates [latitude, longitude]. The closest available data from the specified point will be retrieved.
  • variables: List of lists of variables (2D and 3D) to use as predictors. Different models will be trained for each list (e.g., [["ubar", "vbar", "zos"], ["thetao", "so"]] will train two models), except in the case of AdaBoost (always one model per variable).
• discharge: File path for .mat discharge data.
• tidalRange: File path for .mat tidal range data.

preprocess

• trainTestSplit:
  • method: Method for splitting data (e.g., "temporal").
  • testSize: Proportion of data to use for testing.
• scale:
  • method: Method for scaling data (e.g., "standard").
• dimReduction:
  • method: Method for dimensionality reduction (e.g., "pca").
  • nComponents: Number of components to keep, or explained variance in the case of "pca", if using dimensionality reduction.

model

• method: Machine learning method to use ("analogues", "adaboost", or "lstm").

analogues

• clustering: Clustering method (e.g., "spectral").
• nAnalogues: Number of analogues.
• regressor: Regressor type ("knn" or "krr") to reconstruct the time-series.
• knn:
  • n_neighbors: Number of neighbors.
  • weights: Weight function (e.g., "distance").
  • metric: Distance metric (e.g., "minkowski").
• krr:
  • kernel: Kernel type (e.g., "rbf").
  • alpha: Regularization parameter.
  • gamma: Kernel coefficient.

adaBoost

• nSplits: Number of splits for cross-validation.
• estimator:
  • maxDepth: List of maximum tree depths.
  • criterion: List of criteria for splitting (e.g., ["squared_error"]).
  • splitter: List of split strategies (e.g., ["best"]).
  • minSamplesSplit: List of minimum samples required to split.
  • minSamplesLeaf: List of minimum samples required at a leaf node.
• nEstimators: List of estimator counts.
• learningRate: List of learning rates.
• loss: List of loss functions (e.g., ["square", "exponential", "linear"]).
• scoring: Scoring metric (e.g., "neg_mean_squared_error").
• nJobs: Number of parallel jobs.

lstm

• differentNetworks: List of lists with variable groups for different networks (e.g., [["u_x", "u_y", "waterlevel"]]).
• nTimesteps: Number of timesteps for input sequences.
• stepSize: Step size for sequences.
• lstmLayers:
  • minLstmLayers: Minimum number of LSTM layers.
  • maxLstmLayers: Maximum number of LSTM layers.
  • minLstmUnits: Minimum units per LSTM layer.
  • maxLstmUnits: Maximum units per LSTM layer.
  • stepLstmUnits: Step size for units.
• dropout:
  • minDropout: Minimum dropout rate.
  • maxDropout: Maximum dropout rate.
  • stepDropout: Step size for dropout rate.
• denseLayers:
  • minDenseLayers: Minimum number of dense layers.
  • maxDenseLayers: Maximum number of dense layers.
  • minDenseUnits: Minimum units per dense layer.
  • maxDenseUnits: Maximum units per dense layer.
  • stepDenseUnits: Step size for units.
• train:
  • optimizer: Optimizer (e.g., "adam").
  • loss: Loss function (e.g., "mean_squared_error").
  • metrics: List of metrics (e.g., "mean_squared_error", "mean_absolute_error").
  • learningRates: List of learning rates.
  • earlyStopping:
  • monitor: Metric to monitor (e.g., "val_loss").
  • patience: Number of epochs with no improvement for early stopping.
  • batch:
  • minBatchSize: Minimum batch size.
  • maxBatchSize: Maximum batch size.
  • stepBatchSize: Step size for batch size.
• hyperband:
  • objective: Objective to optimize (e.g., "val_loss").
  • direction: Optimization direction (e.g., "min").
  • maxEpochs: Maximum number of epochs.
  • factor: Reduction factor for Hyperband.
  • overwrite: Whether to overwrite previous results.
  • directory: Directory for Hyperband results.
  • projectName: Project name for Hyperband results.

🌱 Environment Setup

To set up the environment, use the provided environment.yml file. This file includes all dependencies required for the project.

🚀 Running the Project

1. Clone the repository: sh git clone <repository-url>

2. Navigate to the project directory: sh cd <repository-directory>

3. Create and activate the conda environment: sh conda env create -f environment.yml conda activate hydrodynamic-modeling

4. Run the main script: sh python main.py

5. Adjust the config.json file as needed for different scenarios and re-run the main script.

🤝 Contributions

Contributions are welcome! Please follow the standard GitHub flow for contributions: fork the repository, create a feature branch, make your changes, and submit a pull request.

📜 License

This project is licensed under the MIT License. See the LICENSE file for details.

📧 Contact

For any questions or issues, please contact Mirko Rupani at [mirko.rupani@tecnalia.com].