A Dataset and Baseline Approach for Identifying Usage States from Non-intrusive Power Sensing with MiDAS IoT-Based Sensors

Bharath Muppasani, Cheyyur Jaya Anand, Chinmayi Appajigowda, Biplav Srivastava, Lokesh Johri

Abstract

The state identification problem seeks to identify power usage patterns of any system, like buildings or factories, of interest. In this challenge paper, we make power usage dataset available from 8 institutions in manufacturing, education and medical institutions from the US and India, and an initial unsupervised machine learning based solution as a baseline for the community to accelerate research in this area.

Introduction: The State Identification Problem

In industrial and commercial settings, monitoring equipment activity is essential for energy efficiency and predictive maintenance. However, intrusive sensing—installing individual monitors on every motor, server, or machine—is often cost-prohibitive and logistically complex.

This research addresses the State Identification Problem (SIP) within the framework of Non-Intrusive Load Monitoring (NILM). The goal is to "disaggregate" a total power signal from a single entry point to understand individual device behaviors. The primary challenges in this field include the lack of labeled ground truth data and the high dimensionality of high-frequency electrical data.

Methods

The authors propose a multi-step unsupervised pipeline to tackle SIP:

  1. Feature Selection: The approach focuses on current harmonics (specifically odd harmonics), which serve as a "fingerprint" for different electrical loads, providing more detail than raw wattage alone.
  2. Clustering (State Discovery): Using K-Means clustering, the system identifies unknown operational states. The optimal number of states ($k$) is determined via the Elbow Method and Silhouette Scores.
  3. Dimensionality Reduction: Principal Component Analysis (PCA) is used to project high-dimensional harmonics data (45+ features) into a 2D space for interpretability.
  4. Classification: Once states are identified via clustering, a Random Forest classifier is trained on three weeks of data to predict these states in real-time for new observations.

Datasets: The MiDAS IoT-Based Sensor Dataset

The paper introduces a massive 27GB+ dataset following FAIR principles. It covers 15 days of data from 8 diverse locations, including manufacturing plants, a hospital, and data centers across the US and India.

Summary of the MiDAS dataset across eight diverse locations, including industry types and data characteristics.

The dataset includes:

  • Electricity Consumption Data (ECD): 28 features (Voltage, Current, Power Factor, etc.) sampled at 300ms intervals.
  • Harmonics Data: 193 features (Current and Voltage harmonics up to the 32nd level) sampled at 500ms intervals.

Key Results

The methodology was demonstrated on the "India-4" manufacturing site. Analysis of the load profile suggested the plant operates in six distinct states.

Active power load profile for the India-4 manufacturing plant, used as a primary case study for state identification.
  • State Validation: PCA plots confirmed that these states formed well-separated clusters, which were subsequently validated by domain experts on the ground.
  • Baseline Performance: The Random Forest classifier achieved F1 scores between 0.51 and 0.67. While modest, these scores provide a realistic baseline for future research using more advanced architectures like Transformers or LSTMs.

Discussion and Future Work

A key insight of this work is that electrical harmonics—distortions in the AC waveform—provide a much richer signature for identifying operational modes than raw power usage. Because manual labeling is unfeasible at scale, unsupervised clustering is presented as the most viable path forward.

However, limitations remain. The lack of granular ground truth makes it difficult to precisely map clusters to specific machine combinations (e.g., distinguishing "Lathe + HVAC" from "Lathe + CNC"). Future work will focus on improving model generalizability across industries and enhancing explainability for facility managers.

Summary updated

Abstract

The state identification problem seeks to identify power usage patterns of any system, like buildings or factories, of interest. In this challenge paper, we make power usage dataset available from 8 institutions in manufacturing, education and medical institutions from the US and India, and an initial unsupervised machine learning based solution as a baseline for the community to accelerate research in this area.
BibTeX
@inproceedings{muppasani2023dataset,
  author = {Bharath Muppasani and Cheyyur Jaya Anand and Chinmayi Appajigowda and Biplav Srivastava and Lokesh Johri},
  title = {{A Dataset and Baseline Approach for Identifying Usage States from Non-intrusive Power Sensing with MiDAS IoT-Based Sensors}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2023},
  url = {https://openreview.net/forum?id=In9jxq28L4},
}
Figure