Time-series forecasting with Machine Learning

This folder contains notebooks to perform perliminary data analysis and time-series forecasting with machine learning using Python.

Installation

To work with the repository, one can simply clone it in the local machine:

git clone "https://github.com/andreabellome/time_series_forecasting"

If one wants to specify a specific target directory (should be empty):

git clone "https://github.com/andreabellome/time_series_forecasting" /path/to/your/target/directory

where /path/to/your/target/directory should be replaced with the desired local taregt directory.

The toolbox requires Python version 3.10 or above. The toolbox uses common Python libraries: numpy, scipy, matplotlib and plotly. If not already downloaded, please use pip to do so:

pip install numpy, scipy, matplotlib, plotly

Moreover, machine-learning-specific libraries are required: lightgbm for the regression models, skforecast for the Bayesian search and shap for interpreting the results. These can be installed using pip as well:

pip install lightgbm, skforecast, shap

Any version of Jupyter is also required.

License

The work is under license CC BY-NC-SA 4.0, that is an Attribution Non-Commercial license. One can find the specifics in the LICENSE file.

Only invited developers can contribute to the repository.

Usage

Two main files can be used to run the forecast analysis:

• A Jupiter notebook that describes step-by-step the data analysis, preparation and forecast performed.
• A small web-app that can be run in local machine that does the same analysis of the notebook but for a more user-friendly interface.

Notebook

A notebook is provided that is notebook_power_load_forecasting.ipynb. This is self-explanatory, and the interested developer is encouraged to explore the code.

To run the notebook, one needs to have the Jupyter.

pip install jupyter

Web-app

The script to run the web-app is app.py. To run this script, one needs to have streamlit installed:

pip install streamlit

Then, one should open a terminal window inside the time_series_forecasting folder. To start the web-app, just write to terminal:

streamlit run app.py

The web-app uses pre-trained parameters saved in the variable forecaster_params.pkl. One can generate new ones using desired lags and steps using the notebook.

Happy coding!