Hands-on teaching of modern machine learning and deep learning techniques heavily relies on the use of well-suited datasets.
The "weather prediction dataset" is a novel tabular dataset that was specifically created for teaching machine learning and deep learning to an academic audience.
The dataset contains intuitively accessible weather observations from 18 locations in Europe. It was designed to be suitable for a large variety of different training goals, many of which are not easily giving way to unrealistically high prediction accuracy. Teachers or instructors thus can chose the difficulty of the training goals and thereby match it with the respective learner audience or lesson objective.
We provide examples of possible training tasks that could be covered with the dataset (see the Jupyter notebooks in \notebooks\ for the used Python code).
The compact size and complexity of the dataset makes it possible to quickly train common machine learning and deep learning models on a standard laptop so that it can used in live hands-on sessions.
-
The dataset can be found in the
\dataset\folder in this repository or can be downloaded from zenodo: https://doi.org/10.5281/zenodo.7525955 -
The dataset was presented at the "Teaching Machine Learning" workshop at ECML 2022: https://teaching-ml.github.io/2022/
-
The dataset is part of the Open Datasets for (Data Science) education collection
weather_prediction_dataset.csv- Main data file, tabular data, comma-separated CSV. Contains the data for different weather features (daily observations, see below for more details) for 18 European cities or places through the years 2000 to 2010.weather_prediction_picnic_labels.csv- Optional data to be used as potential labels for classification tasks. Contains booleans to characterize the daily weather conditions as suitable for picnic (True) or not (False) for all 18 locations in the dataset.weather_prediction_dataset_map.png- Simple map showing all 18 locations in Europe.metadata.txt- Further information on the dataset, the data processing and conversion, as well as the description and units of all weather features.
The initial meteorological data was retrieved from ECA&D [1] a project that makes available daily observations at meteorological stations throughout Europe and the Mediterranean. 18 different European cities or places were selected of which multiple daily observations were available through the years 2000 to 2010. Those where Basel (Switzerland), Budapest (Hungary), Dresden, Düsseldorf, Kassel, München (all Germany), De Bilt and Maastricht (the Netherlands), Heathrow (UK), Ljubljana (Slovenia), Malmo and Stockholm (Sweden), Montélimar, Perpignan and Tours (France), Oslo (Norway), Roma (Italy), and Sonnblick (Austria).
Recordings of daily meteorological observations for these 18 locations span different times, some contain collections that go back to the 19th century. Here, however, we only selected the time span from 2000 to 2010 resulting in 3654 daily observations. The dataset is then constructed from all data of those 18 locations.
The data in addition consists of different observations. While all selected locations provide data for the variables ‘mean temperature’, ‘max temperature’, and ‘min temperature’, we also included data for the variables 'cloud_cover', 'wind_speed', 'wind_gust', 'humidity', 'pressure', 'global_radiation', 'precipitation', 'sunshine' wherever those were available.
After collecting the data, very basic cleaning of the data was performed. Columns with > 5% invalid entries (“-9999”) were removed, columns with <= 5% invalid entries where kept but invalid entries were replaced by mean values (Carefull! This is a manipulation of the original data and was made because this dataset is intended to be used for training purposes!). This resulted in 165 variables (or features) over the course of 3654 days. Finally, we transformed several data units to achieve more similar ranges of the present values. This makes the data more suitable to be used for machine learning or deep learning even without additional processing. We deliberately did not chose to fully standardize the data because we wanted to keep the presented units and values as intuitively accessible as possible. Temperature are now given in degree Celsius, wind speed and gust in m/s, humidity in fraction of 100%, sea level pressure in 1000 hPa, global radiation in 100 W/m2, precipitation amounts in centimeter, sunshine in hours.
| Feature (type) | Column name | Description | Physical Unit |
|---|---|---|---|
| mean temperature | _temp_mean | mean daily temperature | in 1 °C |
| max temperature | _temp_max | max daily temperature | in 1 °C |
| min temperature | _temp_min | min daily temperature | in 1 °C |
| cloud_cover | _cloud_cover | cloud cover | oktas |
| global_radiation | _global_radiation | global radiation | in 100 W/m2 |
| humidity | _humidity | humidity | in 1 % |
| pressure | _pressure | pressure | in 1000 hPa |
| precipitation | _precipitation | daily precipitation | in 10 mm |
| sunshine | _sunshine | sunshine hours | in 0.1 hours |
| wind_speed | _wind_gust | wind gust | in 1 m/s |
| wind_gust | _wind_speed | wind speed | in 1 m/s |
If you make use of this dataset, in particular if this is in form of an academic contribution, then please cite the following two references:
[1] Klein Tank, A.M.G. and Coauthors, 2002. Daily dataset of 20th-century surface air temperature and precipitation series for the European Climate Assessment. Int. J. of Climatol., 22, 1441-1453. Data and metadata available at http://www.ecad.eu
- Florian Huber, Dafne van Kuppevelt, Peter Steinbach, Colin Sauze, Yang Liu, Berend Weel, "Will the sun shine? – An accessible dataset for teaching machine learning and deep learning", DOI TO BE ADDED!
