Description:

The WEA dataset is derived from the WeatherBench repository and designed for Multivariable Time Series (MTS) forecasting, which is forecasting with exogenous variables, at five geographically diverse cities: London (UK), New York (US), Hong Kong (China), Cape Town (South Africa), and Singapore. It spans the period from 1979 to 2018, with a temporal resolution of 6 hours and a spatial resolution of 5.625° in both latitude and longitude. Each city is matched to its nearest grid point on the WeatherBench grid using minimal absolute distance in both axes.

Task:

Time series forecasting of the 850 hPa temperature (T850) — a widely used mid-tropospheric climate indicator — at each location.

Features:

Target Variable: T850 (850 hPa temperature in Kelvin) at the central grid point.
Exogenous Variables (Total: 44):
- Local: Z500 (500 hPa geopotential), t2m (2m temperature), u10 (10m zonal wind), v10 (10m meridional wind)
- Spatial context: The same 5 variables (T850 + 4 exogenous) from the surrounding 8 grid points (3×3 window).

Temporal coverage:

Test Set: Years 2016, 2017, 2018 — each containing ~1,460 time steps.
Validation Set: One year preceding each test year (2015, 2016, 2017).
Training Set: From Jan 1, 1980 to the end of the year preceding validation (e.g., up to Dec 31, 2014 for the 2015 validation year).

Data usage:

To load weather time series data for a specific city (e.g., London, 2016), use the following Python code:

from datasets import load_dataset
dataset = load_dataset("ClaudiaShu/wea_mts", name="london_2016")

# Access different splits
train_data = dataset["train"]
val_data = dataset["validation"] 
test_data = dataset["test"]

Example entry in the WEA dataset:

train_data[0]
>>
{'time': '1980-01-01T00:00:00', 'z500': 51901.9375, 't850': 265.2726745605469, ..., 'v10_SE': 0.07839111983776093}

Alternatively, you can also define your own train/val/test split

from datasets import load_dataset
london_data = load_dataset(
    "ClaudiaShu/wea_mts", 
    name="london_2016",
    split={"train": "train[:80%]", "validation": "train[80%:90%]", "test": "train[90%:]"}
)

# Access different splits
train_data = london_data["train"]
val_data = london_data["validation"]
test_data = london_data["test"]

TODO:

The current dataset provides only basic train/val/test splits. In our experiments (see our repository), training inputs begin earlier to ensure full coverage of the target forecasting horizon over the training period.
Additionally, while we use t850 as the target variable in the paper, it is not explicitly defined here.

Citation:

If you find these datasets useful for your research, please consider citing the associated paper:

@article{shu2025sonnet,
    author    = {Yuxuan Shu and Vasileios Lampos},
    title     = {{Sonnet: Spectral Operator Neural Network for Multivariable Time Series Forecasting}},
    journal   = {arXiv preprint arXiv:2505.15312},
    year      = {2025}
}

Alternatively, we also have a paper on multivariable time series forecasting for ILI rate. You can check it out here and code here. Please also consider citing it if you find it helpful:

@article{shu2025deformtime,
    author    = {Yuxuan Shu and Vasileios Lampos},
    title     = {{\textsc{DeformTime}: Capturing Variable Dependencies with
                  Deformable Attention for Time Series Forecasting}},
    journal   = {Transactions on Machine Learning Research},
    issn      = {2835-8856},
    year      = {2025},
    url       = {https://openreview.net/forum?id=M62P7iOT7d},
}