Grid Loss Time Series Forecasting: Three Machine Learning Approaches

Description

This project is inspired by the kaggle competition and paper linked below.

data source: https://www.kaggle.com/trnderenergikraft/grid-loss-time-series-dataset

reference paper: https://aaai.org/ojs/index.php/AAAI/article/view/7018

Project setup

This project manages virtual environments using pipenv. This can be downloaded using pip install pipenv.

When the projects is cloned open the directory in a terminal and run pipenv shell, and then pipenv install, to install dependencies. Then everything should be good to go.

File structure overview

📦TronderEnergiLossForecast
 ┣ 📂Data
 ┃ ┗ 📂raw (Test Data Location)
 ┃ ┃ ┣ 📜test.csv
 ┃ ┃ ┣ 📜test_backfilled_missing_data.csv
 ┃ ┃ ┗ 📜train.csv
 ┣ 📂Notebooks
 ┃ ┗ 📜LSTMgridloss.ipynb (Notebook containing LSTM Model)
 ┣ 📂Source
 ┃ ┣ 📜Baseline.py (The baseline prediction function)
 ┃ ┣ 📜DataLoading.py (Used for loading data to models)
 ┃ ┣ 📜DataVisualization.py (Plotting utility)
 ┃ ┣ 📜Evaluation.py (Used to run evaluation metrics)
 ┃ ┣ 📜ExponentialSmoothing.py (Run to see Holt-Winters results)
 ┃ ┣ 📜ExponentialSmoothingOptimization.py (Fitting method Holt-Winters)
 ┃ ┣ 📜ExponentialSmoothingPlotting.py (Plotting for Holt-Winters)
 ┃ ┣ 📜Metrics.py (Evaluation metrics)
 ┃ ┗ 📜Linear.py (Linear Regression model)
 ┣ 📜Pipfile (packages in pyenv)
 ┣ 📜Pipfile.lock
 ┗ 📜README.md

Guide to run the models

Holt-Winters

The files regarding the Holt-Winters method are

The implementations of the Holt-Winters functions are found in ExponentialSmoothing.py. For this file it is enough to run it. The main method will run through the dataset using, the offline method, evaluating it and plotting it. Then it will run the online method, evaluate it and plot it twice, firstly with the whole dataset, and the second time only with the test set.

The model parameters that are set as default in the holt-winters methods are found running the optimalization method in ExponentialSmoothingOptimization.py. This file can also be run as it is, however might take some time to complete. The output from running the file will be the found optimal values for α, β and γ.

Linear Regression

Everything related to the linear regression model is found in Linear.py.

To run the model on the test set with online learning: python Source/Linear.py

Parameters are found by doing a grid search with cross validation, before the model is trained on the entire training set and then is run on the test set. It takes about 1 hour to run.

LSTM

There are 2 options to run the LSTM model

The notebook is available in Kaggle: https://www.kaggle.com/potetsos/lstmgridloss

This notebook is already computed, and the output values and results are therefore available in the notebook.
The notebook can also be run in Kaggle to reproduce the results.

Download and run LSTMgridloss.ipynb
In order to run the notebook locally, the input locations of the .csv files need to be changed to point at the local .csv files.

Kobrestad/TronderEnergiLossForecast