Source-linked AI summary
An Empirical Investigation of V-I Trajectory based Load Signatures for Non-Intrusive Load Monitoring
Taha Hassan, Fahad Javed, Naveed Arshad
TL;DR
NILM depends on choosing load signatures that can distinguish appliance states from aggregate household demand, but evidence for voltage-current trajectory features in NILM was limited. The paper evaluates wave-shape features against power and harmonic signatures on REDD, while using differential evolution for model-parameter selection. Across tested algorithms and conditions, WS provides better or generally comparable prediction precision and robustness, including with dynamic, noisy, and highly similar signatures.
Problem
The paper addresses the limited prior evaluation of wave-shape features for NILM appliance classification and prediction, where load-signature choice is fundamental.
Method
The study evaluates voltage-current trajectory-based wave-shape features against power and harmonic signatures on REDD and uses enhanced differential evolution to select learning-algorithm parameters.
Results
WS offers better or generally comparable prediction accuracy and robustness to PQ and HAR across tested classification algorithms and dynamic, noisy, and highly similar load conditions.
Takeaways & Limitations
Wave-shape features are presented as an effective basis for appliance classification and prediction in semi-automated energy disaggregation and monitoring.
Abstract
from arXiv · showhide
Choice of load signature or feature space is one of the most fundamental design choices for non-intrusive load monitoring or energy disaggregation problem. Electrical power quantities, harmonic load characteristics, canonical transient and steady-state waveforms are some of the typical choices of load signature or load signature basis for current research addressing appliance classification and prediction. This paper expands and evaluates appliance load signatures based on V-I trajectory - the mutual locus of instantaneous voltage and current waveforms - for precision and robustness of prediction in classification algorithms used to disaggregate residential overall energy use and predict constituent appliance profiles. We also demonstrate the use of variants of differential evolution as a novel strategy for selection of optimal load models in context of energy disaggregation. A publicly available benchmark dataset REDD is employed for evaluation purposes. Our experimental evaluations indicate that these load signatures, in conjunction with a number of popular classification algorithms, offer better or generally comparable overall precision of prediction, robustness and reliability against dynamic, noisy and highly similar load signatures with reference to electrical power quantities and harmonic content. Herein, wave-shape features are found to be an effective new basis of classification and prediction for semi-automated energy disaggregation and monitoring.
I. INTRODUCTION
The paper evaluates wave-shape features derived from instantaneous voltage-current trajectories as load signatures for event-based NILM. Using REDD, it compares their prediction precision and robustness with established signatures and applies differential evolution to model selection.
- Motivation and contribution: Wave-shape features use the mutual trajectory of instantaneous voltage and current waveforms to characterize residential appliance loads.The study evaluates these features for appliance profiling and prediction in multi-class NILM.
- Motivation and contribution: The study compares WS with traditional power quantities and harmonic characteristics under dynamic, noisy, and highly similar load signatures.The experiments assess both disaggregation capacity and robustness across classification algorithms.
- Experimental scope: REDD provides the publicly available benchmark dataset for evaluating the employed learning algorithms and disaggregation conditions.The paper reports an overview of algorithmic disaggregation capacity as experimental conditions vary.
- Research gap: A prior examination of WS for NILM was absent to the authors’ knowledge, motivating this empirical evaluation.Earlier work introduced WS for load taxonomies, but this paper studies their use in NILM prediction.
- Experimental scope: The experiments account for ambient electrical noise and interference in household aggregate demand profiles.These conditions are treated as part of the robustness evaluation.
- Methodological contribution: Enhanced differential evolution selects adjustable learning-algorithm parameters through population-based global search.The optimization is presented as a novel technique in the NILM context.
III. DESCRIPTION OF DESIGN METHODOLOGY
The design uses manual calibration, event-based delta-form signatures, high-frequency voltage-current acquisition, and preprocessing to create appliance-labeled switching-event examples. These examples support classification using waveform-derived and benchmark load metrics.
- Manual-setup NILM: Manual-setup NILM learns appliance signatures during a one-time calibration period and stores them for identifying appliances at later switching events.The system distinguishes a training and optimization path from the subsequent NILM query process.
- Signature representation: The study considers delta-form signatures, which capture brief windows containing a single switching event rather than aggregate multi-event snapshots.Snapshot-form signatures represent aggregate power consumption, whereas delta-form signatures are used for this paper.
- Data acquisition: High-frequency NILM acquires instantaneous voltage and current waveforms, typically at more than 100 samples per cycle.These raw waveforms are used to derive wave-shape information from their mutual trajectory.
- Benchmark data: The evaluation uses REDD consumption data and compares traditional power metrics with waveform-based representations of switching events.The figure captions identify the plotted subset as consumption values from REDD switching events using traditional power metrics.
- Preprocessing: Preprocessing separates switching events and extracts cycle-by-cycle voltage and current snapshots as undifferentiated delta-form signatures.K-means clustering groups the signatures into cohesive appliance-labeled classes for multi-class classification.
B. Standard Benchmark Load Signatures
The study benchmarks traditional power metrics, harmonic content, and V-I trajectory wave-shape features as appliance load signatures. WS captures trajectory shape and phase-related behavior, while PQ and HAR characterize power and spectral properties.
- Power Metrics (PQ): PQ represents real and reactive appliance power consumption together with total odd and even current harmonic distortion.
- Harmonic Content (HAR): HAR represents spectral energy in contiguous Fourier-transform segments of appliance current waveforms.
- Wave-Shape Features (WS): V-I trajectories are mutual loci of instantaneous voltage and current waveforms, with distinct shapes for resistive, motor-driven, and power-electronic appliances.
- Wave-Shape Features (WS): Looping direction records clockwise or counter-clockwise trajectory curvature and corresponds to the sign of the voltage-current phase-angle difference.
- Correlation Analysis: WS metrics show strong or weak linear relationships with selected PQ quantities and harmonic content across combinations of about fifty thousand REDD switching events.
D. Disaggregation Algorithms
The evaluation applies four classification algorithms to feature vectors representing appliance load metrics and labels. The algorithms include neural, evolutionary-hybrid, kernel-based, and boosting approaches.
- Learning Setup: Each learning example is a feature vector x_i paired with an appliance label y, with dimensionality determined by the selected load signature.
- Algorithms: The study evaluates feed-forward ANN, ANN + EA, Gaussian-kernel SVM, and AdaBoost with decision stumps.
- ANN: ANN is trained with the Levenberg-Marquardt method.
- ANN + EA: ANN + EA performs evolutionary local search around the ANN solution to seek better prediction accuracy.
- Rationale: The four algorithms were selected because related NILM studies had used them separately, while their systematic examination with WS over benchmark datasets was not available.
E. Model Selection and Performance Optimization
The paper uses differential-evolution-based global search to select learning-model setup parameters before training. Enhanced differential evolution replaces a fixed recombination choice with fitness-informed search behavior.
- Model Selection: Enhanced differential evolution automatically searches setup parameters such as ANN hidden-layer neurons and EA momentum constants during model selection.
- Search Procedure: The search initializes populations of system variables, evaluates prediction-error objective functions, iterates population updates, and selects the individual with the least objective value.
- Differential Evolution: DE is a population-based global-search strategy for minimization over nonlinear continuous functions.
- Search Scope: The EDE procedure is applied to model selection for ANN, ANN + EA, and SVM.
- Enhanced Search: EDE replaces DE's empirical recombination rate with a fitness function that dynamically weighs mutant-population fitness against the original population.
- Multiclass Optimization: For multiclass classification, the study considers either parallel per-class optimization or a single objective function, while more sophisticated adaptive selective objectives remain outside scope.
IV. EXPERIMENTAL EVALUATION
The experimental evaluation measures appliance-label prediction for switching events extracted from REDD. It compares WS, PQ, and HAR across four learning algorithms using prediction precision under a perfect-recall assumption.
- Evaluation Scope: The evaluation measures precision of appliance-label prediction for switching events extracted from the REDD dataset.
- Compared Methods: WS, PQ, and HAR are evaluated with ANN, ANN + EA, SVM, and AdaBoost.
- Metric: Prediction precision is computed as correctly predicted classes weighted by the total number of extracted switching events.
- Assumption: The study assumes perfect recall, meaning all true switching events are detected and reported to the learning algorithms.
A. Description of Dataset and Initial Conditions
The evaluation uses high-frequency REDD data from two houses, with house #3 providing 22 voltage/current channels. Delta-form signatures are filtered, partitioned into training, validation, and test sets, and optimized before Monte-Carlo evaluation.
- Dataset: REDD provides detailed household energy data at low and high measurement granularities.The evaluation uses high-frequency waveform data from approximately twenty days across two houses.
- Dataset: House #3 contributes 22 voltage and current waveform channels for the evaluation.
- Initial filtering: Only appliances meeting the study’s inclusion threshold are retained before clustering delta-form signatures.
- Initial conditions: The base ANN setup uses 0.45, 0.1, and 0.45 proportions for training, cross-validation, and test sets.Early stopping is used for ANN and ANN+EA to limit over-fitting.
- Initial conditions: Enhanced differential evolution selects training parameters before subsequent Monte-Carlo simulations.
B. Monte-Carlo Simulations
Monte-Carlo simulations vary preprocessing and data-partition conditions while preserving the underlying switching-event sequence. These variations test algorithm robustness when training examples are limited and appliance signatures are similar or noisy.
- Simulation design: Each Monte-Carlo iteration evaluates learning precision repeatedly under unchanged computational capacity.For ANN, Levenberg-Marquardt searches the weights at every iteration using the selected hidden-layer sizes.
- Simulation design: The simulations vary training and test-set sizes, appliance-power threshold Pmin, and clustering purity.
- Robustness factors: Training/test proportions assess robustness when early event-based NILM systems have too few examples to model all load scenarios.
- Robustness factors: Pmin controls which appliances enter clustering, while similar, noisy, or non-sinusoidal signatures can increase the required number of appliance classes.
- Scenario generation: These adjustable criteria generate distinct load scenarios without changing the original switching-event sequence.
C. Numerical Results
Across the benchmark comparisons, wave-shape metrics generally match or exceed power-quantity and harmonic signatures in prediction accuracy. Performance remains broadly resilient to test-set changes, but declines when low-power, closely clustered loads are included.
- Prediction accuracy: WS outperforms or generally compares with PQ and HAR in median overall, training-set, and test-set prediction accuracy.WS uses 7 real numbers versus 77 for HAR, achieving comparable or better accuracy with fewer features.
- Algorithm comparison: Adding EA with momentum to ANN does not substantially improve performance over ANN.
- Robustness: AdaBoost and ANN show almost comparable resilience across PQ, HAR, and WS as the test-set proportion increases.
- Robustness: Lower Pmin reduces ANN prediction precision by admitting many low-power load signatures with close feature values.
V. EXTENDED SIMULATIONS
Extended simulations introduce dynamic appliance loading and electrical noise across thousands of REDD-derived load scenarios. Noise reduces precision, with WS showing the largest degradation at lower signal-to-noise ratios.
- Scenario construction: REDD feature vectors generate thousands of load scenarios containing switching-event sequences spanning roughly one day to one month.Scenarios are formed by sequentially adding or subtracting appliance signatures from the composite load.
- Extended conditions: The simulator incorporates dynamic loading from household motors and air conditioners together with electrical noise.
- Noise sensitivity: Zero-mean white Gaussian noise is added to the aggregate demand profile to simulate noise effects.
- Noise sensitivity: All employed algorithms classify more than 90% of training load-signature instances in the specified Pmin = 50W simulation.
- Noise sensitivity: Precision decreases as SNR falls, most sharply for WS compared with PQ and HAR.Adaptive Boost and Support Vector Method cases are illustrated in Figures 9 and 10.
- Noise sensitivity: Adaptive Boost shows about a 20% precision drop at Pmin = 50W under the illustrated low-SNR condition.The response is associated with distortion of appliance V-I trajectories at small SNR.
B. Sensitivity towards Dynamic Load Signatures
The study models dynamic appliance-load variability through waveform variation and evaluates how adjusted current waveforms affect prediction precision. WS and HAR generally show more robust precision than PQ in the tested simulations, although WS loses its advantage at lower SNR.
- Dynamic-load modeling: Dynamic appliance loads exhibit inherent current-waveform shape variability, so the study evaluates signature behavior under changing waveform conditions.Variability is represented through single-cycle current snapshots and their Fourier-domain variation.
- Dynamic-load modeling: Uniform sampling of database variations produces possible instantaneous-current waveforms for updated current-waveform Monte Carlo simulations.The sampling function selects available Fourier-domain components for an appliance with uniform probability.
- Simulation results: WS and HAR show relatively robust prediction precision, while PQ generally exhibits slightly greater disparity across the tested Monte Carlo scenarios.The simulations use Pmin = 25W, 20 days, SNR = 10, and 15 switching events per hour with normally distributed frequency.
- Simulation results: At smaller SNR values, the precision difference between WS and PQ diminishes and WS suffers an explicit disadvantage.The observed robustness is associated with greater separation between WS and HAR signature clusters in the tested setting.
- Overall implication: Across tested scenarios, WS offers superior or generally comparable load-disaggregation performance and corresponds directly to appliance operating characteristics in current wave-shape.The conclusion identifies unsupervised energy disaggregation and device fault monitoring as promising future applications.