Forecasting Electricity Consumption and Classifying Weekends

Authors: Adam Topolski (12600468), Adrianna Bartoszek (12601333), Wojciech Jurewicz (12600946), Katarzyna Kordala (12600809)

This repository contains the implementation of a deep learning approach (PatchTST) and a classical baseline (Prophet) to forecast electricity consumption for 370 clients.

1. Multivariate Forecasting: Given 256 past 15-minute intervals (64 hours), forecast the next 96 intervals (24 hours).
2. Weekend Classification: Classify whether the consumption data falls on a weekend using a shared encoder approach.

• eda_deep.ipynb - Deep exploratory data analysis, stationarity tests, and feature discovery.
• patchtst_final_training.ipynb - Final architecture, training loop, and evaluation for the PatchTST model.
• patchtst_patch_size_experiments.ipynb - Hyperparameter sweep and patch-size experiments.
• prophet_all_normalized_256_96.ipynb - Prophet forecasting baseline (matched to the 256 lookback/96 horizon).
• xgboost.ipynb - XGBoost classification baseline.

1. Dataset Setup: Download the Electricity Load Diagrams 2011-2014 dataset from the UCI Machine Learning Repository and place the extracted LD2011_2014.txt file directly into the dataset/ folder.
2. Dependencies: This project uses uv. Install the required packages via pyproject.toml and uv.lock.
3. Execution: You can reproduce the results by running the notebooks, starting with the EDA and proceeding to the model trainings.