[Paper Review] Neural Network Ensembles to Real-time Identification of Plug-level Appliance Measurements

Karim Said Barsim, Lukas Mauch+1 more|arXiv (Cornell University)|Feb 20, 2018

Smart Grid Energy Management13 references21 citations

TL;DR

This paper proposes a neural network ensemble model that directly classifies plug-level appliance types using raw high-resolution current and voltage waveforms, eliminating the need for engineered features. By normalizing and segmenting periodic signals and augmenting training data via phase shifts, the model achieves 89.7% accuracy on the PLAID dataset, demonstrating robustness to dataset size, sampling frequency, and signal phase variations.

ABSTRACT

The problem of identifying end-use electrical appliances from their individual consumption profiles, known as the appliance identification problem, is a primary stage in both Non-Intrusive Load Monitoring (NILM) and automated plug-wise metering. Therefore, appliance identification has received dedicated studies with various electric appliance signatures, classification models, and evaluation datasets. In this paper, we propose a neural network ensembles approach to address this problem using high resolution measurements. The models are trained on the raw current and voltage waveforms, and thus, eliminating the need for well engineered appliance signatures. We evaluate the proposed model on a publicly available appliance dataset from 55 residential buildings, 11 appliance categories, and over 1000 measurements. We further study the stability of the trained models with respect to training dataset, sampling frequency, and variations in the steady-state operation of appliances.

Motivation & Objective

• Address the appliance identification problem in non-intrusive load monitoring and plug-wise metering using raw electrical waveforms.
• Overcome limitations of prior methods that rely on handcrafted features or engineered signatures.
• Evaluate model robustness to variations in training data size, sampling frequency, and signal phase shifts.
• Demonstrate the effectiveness of neural network ensembles on real-world high-resolution power measurements.
• Provide a scalable, end-to-end solution for real-time appliance identification without user intervention.

Proposed method

• Extract one full period of current and voltage waveforms at a given sampling frequency, with d = fs / fg samples per cycle.
• Apply segment-based normalization to remove amplitude variations, mapping each signal segment to [-1, 1]^d.
• Construct input vectors as concatenated normalized current and voltage segments, forming [−1, 1]^{2d} inputs.
• Use algorithmic data expansion by sliding a window over each measurement with step ε, generating multiple training samples per appliance.
• Train an ensemble of fully connected feedforward neural networks on the expanded dataset to improve generalization and robustness.
• Apply majority voting across ensemble members for final classification predictions.

Experimental results

Research questions

• RQ1 Can neural network ensembles effectively identify appliance types directly from raw current and voltage waveforms without engineered features?
• RQ2 How does model performance vary with the size of the training dataset in a real-world plug-level appliance dataset?
• RQ3 To what extent is the model robust to variations in sampling frequency of the input measurements?
• RQ4 How sensitive is the model to phase shifts in the extracted signal segments during steady-state operation?
• RQ5 Can prior knowledge about household-specific appliance availability improve classification performance?

Key findings

• The ensemble model achieved an unweighted accuracy of 89.7% and Cohen’s kappa of 0.882 on the full PLAID dataset, with the best performance using voting across ensemble members.
• Performance degraded with smaller training sets, indicating strong dependence on data quantity, with accuracy dropping notably as r decreased below 0.5.
• The model remained robust to reduced sampling frequencies, maintaining over 80% accuracy even at 2.5 kHz, showing resilience to lower-resolution inputs.
• The model demonstrated high robustness to phase shifts in signal extraction, maintaining high performance across different starting points in the steady-state period.
• When prior knowledge about household-specific appliance availability was used to reduce the label space, accuracy improved to 94%, highlighting the impact of domain-specific constraints.
• Microwaves, CFLs, vacuum cleaners, and laptops were the most identifiable appliances, while temperature control devices (e.g., air conditioners, heaters, fridges) were the most challenging to classify.