Mix-BEATS: Mixer-enhanced Basis Expansion Analysis for Load Forecasting

This repository contains the official code for the paper:

Mix-BEATS: Mixer-enhanced Basis Expansion Analysis for Load Forecasting
Anuj Kumar, Harish Kumar Saravanan, Shivam Dwivedi, Pandarasamy Arjunan
Robert Bosch Centre for Cyber Physical Systems, Indian Institute of Science, Bangalore
To appear in ACM e-Energy 2025, June 17, 2025, Rotterdam, Netherlands.

---

📄 Paper

Mix-BEATS: Mixer-enhanced Basis Expansion Analysis for Load Forecasting (E-Energy ’25)

---

🧠 Overview

Mix-BEATS is a lightweight, hybrid model for short-term load forecasting that combines the residual learning of N-BEATS with the MLP-based patch and time mixing of TSMixer. Designed for efficiency and generalization, it achieves superior performance across diverse buildings while being suitable for edge deployment.

---

Features

• Hybrid Architecture Combining N-BEATS and TSMixer
  Leverages residual learning with patch and time mixing for efficient and accurate forecasting.

• Mixer-based Temporal Modeling
  Applies MLP-based patch and time mixing operations inspired by vision transformers for effective time-series representation.

• Basis Expansion Analysis
  Uses N-BEATS-style basis functions to improve interpretability and feature extraction.

• Pretrained on Large-Scale, Real-World Smart Meter Data
  Trained on hourly consumption data from over 38,000 buildings to ensure robust generalization.

• Lightweight and Edge-Deployable
  Optimized for computational efficiency, making it suitable for deployment in resource-constrained environments.

• Comprehensive Benchmarking
  Evaluated against state-of-the-art time series foundation models and generic models in zero-shot, fine-tuned, and domain-specific scenarios.

• Open and Reproducible
  Publicly available codebase with training, evaluation, and benchmarking scripts for easy replication and extension.

---

📊 Real-World Building Datasets

This project uses 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.

---

📚 Generic Benchmark Datasets

TSPulse is extensively evaluated on popular time-series forecasting benchmarks from domains such as energy, economics, traffic, weather, and disease surveillance.

Dataset	Variates	Timesteps	Granularity
Weather	21	52,696	10 min
Traffic	862	17,544	1 hour
Electricity	321	26,304	1 hour
ETTh1	7	17,420	1 hour
ETTh2	7	17,420	1 hour
ETTm1	7	69,680	15 min
ETTm2	7	69,680	15 min
Illness (ILI)	7	966	1 week
Exchange-Rate	8	7,588	1 day

---

📈 Comparative Evaluation

We benchmark Mix-BEATS against state-of-the-art Time-Series Foundation Models (TSFMs) and generic data-specific baselines across two broad settings:

---

🔍 1. Comparison with Existing Time-Series Foundation Models (TSFMs)

Mix-BEATS was evaluated against top foundation models including Moirai, Chronos, Lag-Llama and Tiny Time Mixers (TTMs).

🏗️ Dataset and Evaluation Setup

• Training Data: 38,000 real-world building time-series from 11 datasets (see Real Building Datasets section).
• Split Strategy:
  • 80% for training
  • 20% for validation
• Test Set: 1,000 unseen buildings, sampled evenly from all datasets to ensure generalization.
• Evaluation:
  • Zero-shot: Directly tested without task-specific fine-tuning.
  • Fine-tuned: For each building in the test set, data from the first half of the year is used for fine-tuning, while data from the remaining half of the year is used for evaluation. We finetuned model for 10 epochs.

📊 Results (Zero-Shot and Fine-Tune)

Table: Performance Comparison of TSFMs (Zero-Shot & Fine-Tuned)
Mix-BEATS consistently outperforms larger models despite its small size (0.18M params)

NRMSE (Normalized Root Mean Square Error) — Lower is better.
Forecasting context length: 168 hours (7 days)
Forecast horizon: 24 hours (1 day)

Dataset	Zero-shot Moirai	Chronos	Lag-Llama	TTMs	Mix-BEATS	Fine-tuned Moirai	Lag-Llama	Mix-BEATS
Enernoc	29.93	26.72	52.41	23.94	23.27	25.41	28.71	21.72
Iblend	28.03	16.90	63.89	22.24	20.31	22.73	19.17	20.62
Avg. (Top)	28.98	21.81	58.15	23.09	21.79	24.07	23.94	21.17
Mathura	110.33	103.51	114.22	136.79	132.82	92.19	100.07	121.70
Bareilly	70.63	74.57	90.61	65.15	64.01	55.14	70.00	58.96
MFRED	27.83	26.12	62.70	25.47	25.12	21.53	31.54	24.92
NEEA	81.13	83.37	92.66	70.04	69.25	67.54	85.00	70.79
NEST	72.43	72.06	85.73	65.03	65.44	55.71	67.63	51.27
Prayas	91.13	88.66	101.20	102.70	100.90	57.56	78.76	71.64
Smart*	66.63	70.42	84.34	62.27	60.74	65.87	89.85	74.70
Ireland	87.23	93.05	116.36	82.25	81.75	70.74	83.57	77.36
GoiEner	112.35	115.32	131.96	118.42	119.00	100.07	111.80	97.08
SGSC	92.74	100.16	112.64	93.63	92.20	85.97	89.96	83.42
Avg. (All)	81.24	82.72	99.24	82.18	81.12	67.23	80.82	73.18

Note:

• Bold = Best result
• Italic/Underline = Second-best result
• NRMSE values are lower-is-better, indicating better forecast accuracy.

---

🌐 2. Comparison with Generic Data-Specific Models

We also compare Mix-BEATS with popular models like TSMixer, FedFormer, TimesNet, and Non-Stationary Transformers on standard benchmark datasets across energy, weather, traffic, economics, and epidemiology domains.

🧾 Datasets and Splits

• ETT (ETTh1, ETTh2, ETTm1, ETTm2):

  • 60% Training / 20% Validation / 20% Testing

• Electricity, Traffic, Weather, Illness, Exchange-Rate:

  • 70% Training / 10% Validation / 20% Testing

📊 Results on Forecasting Tasks

Metric: MSE (Mean Squared Error)
Context Length: 96
Forecast Horizon: 24
Lower is better — Bold = Best, Italic/Underline = Second-best

Table: Performance Comparison of Generic Models
Mix-BEATS consistently performs well compare to other generic models despite its small size (0.18M params)

Dataset	Mix-BEATS	N-BEATS	Autoformer	LightTS	TSMixer	Reformer	Transformer	TiDE	DLinear
ETTh1	0.040	0.063	0.062	0.057	0.057	0.031	0.057	0.043	0.033
ETTh2	0.074	0.075	0.090	0.086	0.075	0.079	0.090	0.083	0.076
ETTm1	0.014	0.017	0.020	0.018	0.016	0.011	0.016	0.015	0.012
ETTm2	0.020	0.021	0.039	0.027	0.023	0.024	0.021	0.035	0.029
Electricity	0.211	0.202	0.347	0.250	0.206	0.256	0.223	0.361	0.312
Traffic	0.173	0.179	0.274	0.323	0.239	0.183	0.150	0.341	0.330
Weather	0.001	0.001	0.003	0.002	0.001	0.001	0.001	0.001	0.002
Exchange	0.043	0.035	0.057	0.053	0.048	0.035	0.050	0.037	0.037

✅ Note: All values are MSE scores.

• Bold values indicate the best performance on a dataset.
• Italic/underlined values denote the second-best performance.
• Mix-BEATS demonstrates competitive or superior performance across nearly all datasets while maintaining efficiency.

---

🧠 Summary

TSPulse delivers:

• 📉 Superior zero-shot performance—even better than fine-tuned models in many cases.
• 🪶 Extreme parameter efficiency—outperforming models 10–100× larger.
• ⚙️ Deployment-ready architecture—ideal for edge, mobile, and CPU-constrained environments.

---

📣 Citation

If you use this code or ideas from our paper, please consider citing us. (BibTeX will be added after publication.)

---

📬 Contact

For any queries, please contact Pandarasamy Arjunan (samy@iisc.ac.in) or raise an issue in the repository.

---