shtolfensie / pp1-energy-forecasting

Residential home energy load forecasting

This repository contains datasets and models for energy forecasting experiments. Each model can be evaluated by running the following experiments:

1. univariate energy consumption forecasting
   • this can be evaluated on the test part of the given dataset, or on the test part of the other dataset, to check, if the model is transferable
2. multivariate energy consumption forecasting
   • A combination of the following exogenous variables can be tested: temperature, humidity, wind speed

Environment variables

Before running any experiment, make sure the appropriate environment variables are available:

variable name	value	purpose
KERAS_BACKEND	"torch"	Sets keras backend, required for any training or inference.
OPEN_WEATHER_MAP	YOUR_API_KEY	Downloading weather history data. Required when creating a new dataset.

Usage

This project uses two datasets with residential house energy consumption data. To get started, download the raw data into data/nist and data/frhouse respectively:

• NIST: https://pages.nist.gov/netzero/data.html
• IHEPC (frhouse): https://archive.ics.uci.edu/dataset/235/individual+household+electric+power+consumption

To setup an isolated environment, you can run:

python -m venv venv
source venv/bin/activate

Install dependencies:

pip install -r requirements.txt

Then transform the raw data into a dataset with added weather variables:

python src/construct_datasets.py

Experiments can be specified from the CLI by running:

python src/main.py -f history -f temperature --network tcn -e 1

To see all the various options, you can run:

python src/main.py --help

Weather data

Weather history data is downloaded for each datapoint.

Contains information from OpenWeather, which is made available here under the Open Database License (ODbL).