TSFM_Building

🔥 News

• 🔥 v2.0 [11/20/26] - We have reorganized the repository to incorporate the extended journal version, Can Time-Series Foundation Models Perform Building Energy Management Tasks? Our extension preprint is available on ArXiv. Stay tuned for the final published version.
• v1.0 - BuildSys 2024 release: [BuildSys '24] Are Time Series Foundation Models Ready to Revolutionize Predictive Building Analytics?

🔥 We included classification evaluation using TSFM in v2.0.

Classification

# Create the environment from the YAML file
cd ./classification
conda env create -f nesl_environment.yml

# Activate it
conda activate nesl

• Follow the instruction in soft DTW to install the package.
• Download the WHITED dataset (WHITEDv1.1.zip)
• Copy and unzip the dataset:

mv WHITEDv1.1.zip ./classification/data/
cd ./classification
unzip ./data/WHITEDv1.1.zip

Note: We used the BTS dataset for our second classification task. At the time of the experiment, all of the data was not public due to an ongoing competition. Thus we used the data from the competition

# Download train_X_v0.1.0.zip and train_y_v0.1.0.csv from the competition website
# Place them in the data directory
mv train_X_v0.1.0.zip ./classification/data/
mv train_y_v0.1.0.csv ./classification/data/

# Navigate to the classification directory
cd ./classification

# Unzip the training data
unzip ./data/train_X_v0.1.0.zip -d ./data/

• Perform training using baselines resnet, dtw, and ts2vec:

python train_ts2vec_whited.py
python train_dtw_whited.py
python train_resnet_whited.py

• Perform evaluation using TSFM chronos and moment:

# MOMENT model - generates embeddings + trains SVM classifier
python train_tsfm_whited.py --model moment

# Chronos model - generates embeddings + trains SVM classifier
python train_tsfm_whited.py --model chronos

# To save embeddings to .npy files, add --save_embeddings flag:
python train_tsfm_whited.py --model moment --save_embeddings
# This saves: moment_train_embeddings_whited.npy, moment_test_embeddings_whited.npy, and labels

python train_tsfm_whited.py --model chronos --save_embeddings
# This saves: chronos_train_embeddings_whited.npy, chronos_test_embeddings_whited.npy, and labels

• Perform evaluation using TSFM chronos and moment on BTS dataset:

# MOMENT model - generates embeddings + trains SVM classifiers (multi-label)
python train_tsfm_bts.py --model moment

# Chronos model - generates embeddings + trains SVM classifiers (multi-label)
python train_tsfm_bts.py --model chronos

# To save embeddings to .npy files, add --save_embeddings flag:
python train_tsfm_bts.py --model moment --save_embeddings
# This saves: moment_train_embeddings_bts.npy, moment_test_embeddings_bts.npy, and labels

python train_tsfm_bts.py --model chronos --save_embeddings
# This saves: chronos_train_embeddings_bts.npy, chronos_test_embeddings_bts.npy, and labels

Forecasting with Covariates

For the forecasting with covariates task, we use a sampled version of the Ecobee dataset (linked below). The original R2C2 implementation is available at this codebase. To prepare the dataset, clone the R2C2 repository and run the sampling script:

# Clone the R2C2 repository
git clone https://github.com/ozanbarism/GBM4SingleZoneMultiNodeSystems.git
cd GBM4SingleZoneMultiNodeSystems

# Run the sampling script (requires the Ecobee dataset downloaded first)
python sample_ecobee_for_covariate.py

This will generate a folder called house_data_csvs with different noise levels (e.g., house_data_csvs_0, house_data_csvs_0.1, etc.) that models for forecasting with covariates will use.

We evaluate the forecasting performance of three TSFMs: Uni2TS, TimesFM, and TimeGPT, which are the only models in this study that explicitly support covariate-based predictions. For TimesFM, we use the Dec 30, 2024 release shown here with the Xreg + TSFM approach. Uni2TS (Moirai) and TimeGPT inherently support covariates so we use them as is.

# Create the environment from the YAML file
conda env create -f covariate_environment.yml

# Activate it
conda activate covariate_env

• Perform training using baseline forecasting models with covariates (R2C2, Ridge, Random Forest):

# Note: First prepare the dataset using sample_ecobee_for_covariate.py as described above

# Train all baseline models on a specific noise level
python train_forecasting_baselines.py --noise 0.1 --pred_hrz 64 --duration 448

# Train on all noise levels (0, 0.1, 0.2, 0.5, 1, 2, 5)
python train_forecasting_baselines.py --all_noise

Usage

The following code will run inference on each model and dataset:

python main.py --model YOUR_MODEL --real_data DATA

Here,

real_data \in {ecobee, electricity_uci, umass}

model \in {AutoARIMA,
              SeasonaARIMA,
              moment,
              chronos,
              TimeGPT,
              TimesFM,
              uni2ts
              ...}

Download dataset

Download electricity_uci dataset

Download the dataset from ElectricityLoadDiagrams20112014 Then put LD2011_2014.txt in the ./data/ folder.

Download ecobee dataset

Download the dataset Ecobee and extract it. Then run the sampling script for univariate forecasting:

# Run the sampling script (from the root directory of this repository)
python sample_ecobee_for_univariate.py

This will generate combined_thermostat_data.csv. Move it to the ./data/ folder:

mv combined_thermostat_data.csv ./data/

Download smart* dataset

The smart* (umass) dataset has been uploaded with the repo. We downloaded the apartment dataset called ' apartment-electrical.tar.gz' and merged the data for years 2015 and 2016. It uses CC BY 4.0 license, which allows copy and redistribute the material.

TS-foundation-model Setup

To reproduce the results, please first setup environments for each model:

Chronos installation

You can find it in this repo

Create an environment for Chronos

virtualenv chronos -p python3.10
source chronos/bin/activate

Install general packages

pip install -r requirements.txt

Install chronos

python3.10 -m pip install git+https://github.com/amazon-science/chronos-forecasting.git

Moment installation

You can find it in this repo

Create a environment for Moment. Please note that only python version >= 3.10 is supported.

virtualenv moment -p python3.10
source moment/bin/activate

Install general packages

pip install -r requirements.txt

Install moment

python3.10 -m pip install git+https://github.com/moment-timeseries-foundation-model/moment.git

Uni2ts installation

1. Clone repository:

git clone https://github.com/SalesforceAIResearch/uni2ts.git
cd uni2ts

2. Create virtual environment:

virtualenv uni2ts -p python3.10
source uni2ts/bin/activate

3. Build from source:

pip install -e '.[notebook]'

Install general packages

pip install -r requirements.txt

TimeGPT installation

Install general packages

pip install -r requirements.txt

pip install nixtla>=0.5.1

need to get api key from nixtla

import pandas as pd
from nixtla import NixtlaClient

# 1. Instantiate the NixtlaClient
nixtla_client = NixtlaClient(api_key = 'YOUR API KEY HERE')

TimesFM installation

View the original repo. Follow the instruction in the repo to install the model.

[Update on 09/30/24] TimesFM can be installed via pip install timesfm

LagLlama installation

Create a new conda env

# Alternative with conda:
# conda create -n LagLlama python=3.10
# conda activate LagLlama

# Using virtualenv:
virtualenv LagLlama -p python3.10
source LagLlama/bin/activate

Install general packages

pip install -r requirements.txt

clone the repo

git clone https://github.com/time-series-foundation-models/lag-llama/

Go to the current folder

cd <Current_folder>

Hyperlink your lag-llama folder to the current folder

ln -s <Your_lag-llama_folder> lag_src

Copy the model file and go to the lag-llama folder

cp lagllama_model.py lag_src
cd ./lag_src

install the requirements

pip install -r requirements.txt --quiet
# This could take some time

download pretrained model weights from HuggingFace 🤗

huggingface-cli download time-series-foundation-models/Lag-Llama lag-llama.ckpt --local-dir ./

and it will copy the model file to lag-llama

Contact Information

If you have any questions or feedback, feel free to reach out:

• Name: Ozan Baris Mulayim

• Email: [email protected]

• Name: Pengrui Quan

• Email: [email protected]

License

This dataset is released under the BSD 3-Clause License. See the LICENSE file for details.

Acknowledgement

Mario Bergés and Mani Srivastava hold concurrent appointments as Amazon Scholars, and as Professors at their respective universities, but work in this paper is not associated with Amazon. Dezhi Hong is also affiliated with Amazon but work in this paper is not associated with Amazon. This research was sponsored in part by AFOSR award #FA95502210193, DEVCOM ARL award #W911NF1720196, NSF award #CNS-2325956, NIH award #P41EB028242, and Sandia National Laboratories award #2169310.

We would like to extend our thanks to authors of ts2vec, from which the baseline is built on.