ai-iot/Forecasting-W-LSTMix

W-LSTMix: A Hybrid Modular Forecasting Framework for Trend and Pattern Learning in Short-Term Load Forecasting

W-LSTMix is a lightweight, modular hybrid forecasting model designed for building-level load forecasting across diverse building types. With approximately 0.13 million parameters, W-LSTMix combines:

• Wavelet-based signal decomposition
• N-BEATS for ensemble forecasting
• LSTM for gated memory
• MLP-Mixer for efficient patch-wise mixing

This model achieves high forecasting accuracy with a minimal computational footprint.

🚀 Features

• Hybrid Architecture Combining N-BEATS, LSTM and MLP-Mixer
• Lightweight: ~0.13M parameters and Edge-Deployable
• Modular design for flexible adaptation
• Effective generalization across building types
• Zero-shot capabilities

📓 Colab Quickstart

Use the following steps to try W-LSTMix on Google Colab:

!git clone https://github.com/shivDwd/W-LSTMix.git
%cd W-LSTMix
!git clone https://huggingface.co/datasets/shivDwd/W_LSTMix_test_dataset
!pip install -r requirements.txt
!python test.py

📊 Real-World Building Datasets

This model is trained on large-scale real-world building energy datasets from commercial and residential domains, collected from multiple countries.

Dataset	Location	Type	# Buildings	# Observations	Years
IBlend	India	Commercial	9	296,357	2013–2017
Enernoc	USA	Commercial	100	877,728	2012
NEST	Switzerland	Residential	1	34,715	2019–2023
Ireland	Ireland	Residential	20	174,398	2020
MFRED	USA	Residential	26	227,622	2019
CEEW	India	Residential	84	923,897	2019–2021
SMART*	USA	Residential	114	958,998	2016
Prayas	India	Residential	116	1,536,409	2018–2020
NEEA	USA	Residential	192	2,922,289	2018–2020
SGSC	Australia	Residential	13,735	172,277,213	2011–2014
GoiEner	Spain	Residential	25,559	632,313,933	2014–2022

Total: 39,956 buildings and 812M+ hourly observations

⚠️ These datasets are used under their respective terms/licenses for academic research only.

📈 Comparative Evaluation

We benchmark W-LSTMix against state-of-the-art Time Series Foundation Models (TSFMs) and N-BEATS under two broad settings: zero-shot and fine-tuning. Please refer to the publication for a detailed summary of the results.

W-LSTMix: A Hybrid Modular Forecasting Framework for Trend and Pattern Learning in Short-Term Load Forecasting
Shivam Dwivedi, Anuj Kumar, Harish Kumar Saravanan, Pandarasamy Arjunan
In Proceedings of the 1st ICML Workshop on Foundation Models for Structured Data, Vancouver, Canada. 2025
https://openreview.net/pdf?id=bG04Z3Jioc

To know more about W-LSTMix, please regfer to the official Github repository.

📄 Citation

If you use W-LSTMix in your research or applications, please cite our paper:

@inproceedings{
dwivedi2025wlstmix,
title={W-{LSTM}ix: A Hybrid Modular Forecasting Framework for Trend and Pattern Learning in Short-Term Load Forecasting},
author={SHIVAM DWIVEDI and Anuj Kumar and Harish Kumar Saravanan and Pandarasamy Arjunan},
booktitle={1st ICML Workshop on Foundation Models for Structured Data},
year={2025},
url={https://openreview.net/forum?id=bG04Z3Jioc}
}