v2.0.0

Summary of changes

Buildings-900K pretraining dataset

• Added the building simulation metadata files to the dataset, which contain attributes for the original EnergyPlus building energy models used to run the simulations that generated the timeseries. See Buildings-900K/end-use-load-profiles-for-us-building-stock/2021/resstock_amy2018_release_1/metadata/metadata.parquet for an example.
• Added the weather files that were used to run each building simulation, for both AMY2018 and TMY3. See Buildings-900K/end-use-load-profiles-for-us-building-stock/2021/resstock_amy2018_release_1/weather for an example.

BuildingsBench evaluation datasets

• Temperature timeseries data added for each of the 7 benchmark datasets. See docs for more details.

Library updates

• We primarily updated the dataset classes and training/evaluation scripts to add support for weather timeseries inputs
• Other minor usability improvements

New Transformer model configs for weather timeseries inputs

TransformerWithGaussian-t-* only uses temperature timeseries inputs, TransformerWithGaussian-th-* only uses temperature and humidity inputs, and TransformerWithGaussian-weather-* uses all available weather variables during pretraining.

v1.1.0

Summary of changes:

Buildings-900K Pretraining Dataset

• Load value scaling is fixed. By error, we averaged instead of summed when aggregating 15-min to hourly data

Outlier removal

• Updated paths to point to a filtered version of the time series with outliers removed
• Argument added to switch off outlier-removal (use the unfiltered data)
• Script with our implementation of outlier removal

New baselines

• DeepAutoregressiveRNN

Misc

• New script for splitting train/test for Buildings-900K
• Updates to zero-shot/transfer learning scripts for outlier removal
• Hyperparameter optimization arguments added to pretrain.py

Docs and tests updated accordingly.