---
license: cc
---
## Illinois building energy consumption

This repository contains two datasets of 592 Illinois buildings each, one being more heterogenous than the other. The data is sourced from the [NREL ComStock](https://comstock.nrel.gov/) model/dataset.

## Usage

**Multivariate dataset**

The file `custom_dataset.py` contains the function `get_data_and_generate_train_val_test_sets`, which takes in three arguments as follows:

- `data_array`: This takes in a `np.ndarray` of shape `(num_buildings, time_points, num_features)` (note that for our experiments, `num_features` is fixed to 8). You can load them from the `.npz` files provided in this repository as `data_array = np.load(./IllinoisHeterogenous.npz)['data']`.
- `split_ratios`: A list of positives that sum upto 1, denoting the split (along the time axis) into train-validation-test sets. For example, `split_ratios = [0.8,0.1,0.1]`. Must sum to 1.
- `dataset_kwargs`: Additional kwargs for configuring data. For example, `dataset_kwargs = { 'lookback':96, 'lookahead':4, 'normalize':True, 'transformer':True }`.
  - 'lookback` is the number of previous points fed as input. Also denoted by L.
  - `lookahead` is the number of points ahead to predict. Also denoted by T.
  - `normalize` (boolean): If set to `True`, data is normalized per-feature.
  - `transformer` (boolean): If set to `True` and `normalize` is also `True`, then categorical time features are not normalized. Useful for embedding said features in Transformers.

The outputs are as follows:

- `train`: `torch.utils.data.Dataset` containing the train data.
- `val`: `torch.utils.data.Dataset` containing the validation data.
- `test`: `torch.utils.data.Dataset` containing the test data.
- `mean`: `np.ndarray` containing the featurewise mean. If `normalization` is `False`, then it defaults to all `0`s.
- `std`: `np.ndarray` containing the featurewise standard deviation. If `normalization` is `False`, then it defaults to all `1`s.

**Univariate dataset**

The file `custom_dataset_univariate.py` is used in the same way as `custom_dataset.py`, except `dataset_kwargs` does not take in a key called `transformer` (it is deprecated) since there are no categorical features to normalize.

## Requirements and Example

This repository only requires `numpy` and `torch` packages to run. For details on elements of each dataset, plus its different configurations, one is encouraged to run `example_dataset.py`.

## ComStock Notice

This data includes information from the ComStock™ dataset developed by the National Renewable Energy Laboratory (NREL) with funding from the U.S. Department of Energy (DOE). This model was trained using ComStock release 2023.2. NREL regularly publishes updated datasets which generally improve the representation of building energy consumption. Users interested in training their own models should review the latest dataset releases to assess whether recent updates offer features relevant to their modeling objectives.

**Suggested Citation:**

Parker, Andrew, et al. 2023. ComStock Reference Documentation. Golden, CO: National Renewable Energy Laboratory. NREL/TP-5500-83819. https://www.nrel.gov/docs/fy23osti/83819.pdf