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 (SIP) seeks to identify power usage patterns of complex systems, such as industrial buildings or medical facilities. In this paper, we present a comprehensive power usage dataset collected from eight institutions across the US and India, spanning manufacturing, education, and healthcare sectors. We also provide an unsupervised machine learning-based baseline solution to accelerate research in non-intrusive load monitoring (NILM).

Introduction

In energy management, while forecasting total consumption is common, there is a critical need for State Identification (SIP)—the ability to automatically detect whether a system is operating normally, idling, or experiencing early-stage failure based solely on its electrical "fingerprint." Traditionally, this required intrusive sensors on every device. This work focuses on Non-Intrusive Load Monitoring (NILM), which disaggregates aggregate power data from a single point, such as a main breaker, to understand the operational states of downstream equipment.

Methodology

Data was collected using MiDAS IoT sensors, which capture electrical signals at high resolution (300ms to 500ms). Beyond standard wattage, these sensors record electrical harmonics up to the 32nd level. These harmonics serve as unique signatures for different machinery types.

Our baseline approach utilizes a four-step unsupervised pipeline:

  1. Feature Selection: Focusing on odd current harmonics that carry the most state information.
  2. State Discovery: Using K-Means clustering to identify distinct operational states. The optimal number of clusters (k) is determined via the Elbow Method and Silhouette Scores.
  3. Dimensionality Reduction: Principal Component Analysis (PCA) is employed to project 15-dimensional harmonic data into a 2D space for visualization.
  4. Classification: A Random Forest classifier is trained on the discovered clusters to predict states in real-time on new data.

Datasets

The released dataset comprises over 27GB of data collected over 15 days across eight diverse locations. It includes Electricity Consumption Data (ECD) with 28 features and a Harmonics Dataset with 193 features. The locations range from a hospital's main supply to specific manufacturing equipment like CNC and lathe machines.

Summary of the 8 monitored locations, including industry types, load illustrations, and dataset characteristics.

Key Results and Discussion

The "India-4" manufacturing site served as the primary case study for the baseline model. Analysis determined that six distinct states best described the plant's operations.

Active power (ActivePT) vs. time for the India-4 manufacturing site, illustrating the complex signal used for state identification.

Following state discovery, the Random Forest classifier achieved F1 scores between 0.51 and 0.67 on test data. While modest, these results establish a robust baseline for unsupervised SIP in noisy, real-world environments. PCA visualizations confirmed that discovered states formed well-separated groups, validating that harmonics effectively capture physical operational differences.

Limitations and Future Work

Real-world data challenges include missing data (up to 10% in some locations) and the lack of physical "ground truth" labels for discovered clusters. Future research will focus on more robust data interpolation, improving algorithm generalizability across diverse industries, and developing "human-in-the-loop" systems to allow facility managers to label discovered states for better actionability.

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 un-supervised 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},
}

Loading related content...

Figure