Repository for our published foundation model, MixForecast, designed for efficient and accurate short-term load forecasting (STLF) in smart buildings.
📢 Our paper on MixForecast has been accepted at FMSys 2025!
Check out the publication below:
MixForecast: Mixer-Enhanced Foundation Model for Load Forecasting
Anuj Kumar, Harish Kumar Saravanan, Shivam Dwivedi, Pandarasamy Arjunan
In Proceedings of the 2nd International Workshop on Foundation Models for Cyber-Physical Systems & Internet of Things (FMSys 2025)
DOI: 10.1145/3722565.3727193
MixForecast is a novel Time Series Foundation Model (TSFM) that combines N-BEATS and TSMixer blocks to deliver lightweight, generalizable, and high-performing short-term load forecasting (STLF) for smart buildings. Designed for edge-device deployment, it uses approximately 0.19M parameters, offering significant efficiency without compromising accuracy.
- 🧠 Hybrid architecture combining strengths of N-BEATS and TSMixer
- 💡 Lightweight model with only ~0.19M parameters
- 🏠 Trained on 63K smart meter time series from global buildings
- 🌍 Tested on 1,000 commercial and residential buildings worldwide
- 🔄 Supports both zero-shot and fine-tuned load forecasting
- 📈 Outperforms Lag-Llama, Tiny Time Mixers, Moirai, Chronos, and traditional models
- Details of Real Building Datasets Used for this Study (Hourly Data).
- These datasets will be publicly available on Hugging Face soon.
| Dataset Name | Location | No. of Buildings | Years |
|---|---|---|---|
| Commercial Buildings | |||
| Enernoc | USA | 100 | 2012 |
| IBlend | India | 9 | 2013-2017 |
| DGS | USA | 322 | 2016-2018 |
| EWELD | China | 386 | 2016-2022 |
| Residential Buildings | |||
| CEEW | India | 84 | 2019-2021 |
| Ireland | Ireland | 20 | 2020 |
| MFRED | USA | 26 | 2019 |
| NEEA | USA | 192 | 2018-2020 |
| NEST | Switzerland | 1 | 2019-2023 |
| Prayas | India | 116 | 2018-2020 |
| SMART* | USA | 114 | 2016 |
| SGSC | Australia | 13,735 | 2011-2014 |
| GoiEner | Spain | 25,559 | 2014-2022 |
| UK SMART Trials | UK | 14,319 | 2008-2010 |
| SAVE | UK | 4,691 | 2016-2018 |
| iFlex | Norway | 4,429 | 2020-2021 |
| Total | 64,103 |
Note: The dataset will be made available on HuggingFace shortly!
Results are shown Below.
The table below compares the forecasting performance of MixForecast with NBEATS and TTMS models across multiple datasets. Metrics include Normalized Root Mean Squared Error (NRMSE):
| Model | Dataset | average NRMSE |
|---|---|---|
| MixForecast | Commercial | 21.06 |
| NBEATS | Commercial | 25.80 |
| TTMs | Commercial | 22.36 |
| MixForecast | Residential | 80.49 |
| NBEATS | Residential | 84.26 |
| TTMs | Residential | 81.45 |
The table below compares the forecasting performance of MixForecast with TSFMs in both scenario zero-shot and finetuned across multiple datasets. Metrics include Normalized Root Mean Squared Error (NRMSE):
| Model | average NRMSE for Commercial Buildings | average NRMSE for Residential Buildings |
|---|---|---|
| MixForecast - zeroshot | 21.06 | 80.49 |
| Moirai - zeroshot | 28.25 | 80.51 |
| Chronos - zeroshot | 21.08 | 81.99 |
| Lag-llama - zeroshot | 57.42 | 98.51 |
| MixForecast - finetuned | 20.44 | 72.45 |
| Moirai - finetuned | 23.34 | 66.50 |
| Lag-llama - finetuned | 23.21 | 80.09 |
These Plots are for three consecutive windows of sequence length 168 (7 days) and prediction length 24 (1 day).
- MixForecast outperforms NBEATS and TTMS on NRMSE metrics, demonstrating its robustness and accuracy for STLF tasks.
- The architecture's focus on task-specific enhancements ensures adaptability to diverse building profiles.
- MixForecast outperforms NBEATS and TTMS on NRMSE metrics, demonstrating its robustness and accuracy for STLF tasks.
- The architecture's focus on task-specific enhancements ensures adaptability to diverse building profiles.
If you use MixForecast in your research or applications, please cite our paper:
@inproceedings{10.1145/3722565.3727193,
author = {Kumar, Anuj and Saravanan, Harish Kumar and Dwivedi, Shivam and Arjunan, Pandarasamy},
title = {MixForecast: Mixer-Enhanced Foundation Model for Load Forecasting},
year = {2025},
isbn = {9798400716089},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3722565.3727193},
doi = {10.1145/3722565.3727193},
booktitle = {Proceedings of the 2nd International Workshop on Foundation Models for Cyber-Physical Systems & Internet of Things},
pages = {25--30},
numpages = {6},
keywords = {Demand-side Load management, Energy Forecasting, Energy Informatics, Machine Learning, Short-term Load Forecasting (STLF), Smart Grid, Time Series Foundation Models (TSFM)},
location = {Irvine, CA, USA},
series = {FMSys}
}We would like to express our sincere gratitude to the authors of the following works for their valuable contributions to the field of short-term load forecasting (STLF) and energy prediction, which have influenced and inspired our work on MixForecast:
-
N-BEATS: Neural Basis Expansion Analysis for Time Series (Oreshkin et al., 2019)
- This work introduced N-BEATS, a powerful and scalable model for time series forecasting, which served as a key foundation for the development of our model architecture.
- Citation:
Oreshkin, B. N., Carpov, D., Chapados, N., & Bengio, Y. (2019). N-BEATS: Neural Basis Expansion Analysis for Interpretable Time Series Forecasting. arXiv preprint arXiv:1905.10437.
-
TSMixer: Lightweight MLP-Mixer Model for Multivariate Time Series Forecasting (Ekambaram et al., 2023)
- The TSMixer model's innovative approach to temporal sequence modeling provided significant insights for designing the temporal mixing layers in MixForecast.
- Citation:
Ekambaram, V., Jati, A., Nguyen, N., Sinthong, P., & Kalagnanam, J. (2023). TSMixer: Lightweight MLP-Mixer Model for Multivariate Time Series Forecasting. Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 459–469.
Additionally, we acknowledge the publicly available datasets and the authors of these datasets for making their data accessible, which enabled us to benchmark and evaluate our model on real-world scenarios.
We also thank the contributors, researchers, and reviewers who provided invaluable feedback and suggestions that helped refine this work.