[Paper Review] Data augmentation for dealing with low sampling rates in NILM

Tai Le Quy, Sergej Zerr+2 more|arXiv (Cornell University)|Mar 30, 2021

Advanced Adaptive Filtering Techniques12 references4 citations

TL;DR

This paper proposes data augmentation via interpolation techniques—such as Denton-Cholette, stepwise, cubic spline, and device-specific methods—to increase the sampling rate of low-frequency energy disaggregation data for Non-Intrusive Load Monitoring (NILM). The approach improves NILM algorithm performance, with device-specific interpolation yielding the highest F-score gains, especially on cooling, entertainment, and computer appliances, demonstrating that data-level augmentation enhances model robustness and accuracy on low-sampling datasets.

ABSTRACT

Data have an important role in evaluating the performance of NILM algorithms. The best performance of NILM algorithms is achieved with high-quality evaluation data. However, many existing real-world data sets come with a low sampling quality, and often with gaps, lacking data for some recording periods. As a result, in such data, NILM algorithms can hardly recognize devices and estimate their power consumption properly. An important step towards improving the performance of these energy disaggregation methods is to improve the quality of the data sets. In this paper, we carry out experiments using several methods to increase the sampling rate of low sampling rate data. Our results show that augmentation of low-frequency data can support the considered NILM algorithms in estimating appliances' consumption with a higher F-score measurement.

Motivation & Objective

• Address the challenge of poor NILM performance due to low sampling rate and incomplete data in real-world energy disaggregation datasets.
• Improve the quality of low-sampling-rate datasets (e.g., 1/60 Hz) to support more accurate appliance-level power consumption estimation.
• Investigate whether data-level augmentation can enhance the performance of existing NILM algorithms without modifying the algorithms themselves.
• Evaluate the effectiveness of various interpolation techniques in reconstructing high-resolution time series from sparse data.
• Demonstrate that data augmentation is a viable, method-agnostic approach to improve NILM performance, especially for data-hungry deep learning models.

Proposed method

• Apply multiple interpolation techniques—Denton-Cholette, stepwise, cubic spline, and device-specific interpolation—to reconstruct high-sampling-rate signals from low-sampling-rate power consumption data.
• Use two benchmark datasets (ECO and iAWE) with original 1 Hz sampling rates, downsampled to 1/60 Hz for augmentation experiments.
• Apply the augmentation techniques to generate synthetic high-frequency data while preserving the temporal and power characteristics of the original signals.
• Evaluate the augmented data using two established NILM algorithms: Weiss’s algorithm (supervised) and Parson’s algorithm (semi-supervised) with factorial HMMs.
• Assess performance using F-score as the primary metric, comparing original, undersampled, and augmented data across different appliance categories.
• Categorize appliances into functional groups (cooling, cooking, entertainment, computer, lighting) to analyze performance variation across device types.

Experimental results

Research questions

• RQ1 Can data augmentation via interpolation significantly improve the F-score of NILM algorithms on low-sampling-rate datasets?
• RQ2 Which interpolation method—Denton-Cholette, stepwise, cubic spline, or device-specific interpolation—yields the best performance in energy disaggregation?
• RQ3 How does data augmentation affect the performance of NILM algorithms across different appliance types, particularly those with low or variable power consumption?
• RQ4 Does the proposed data-level augmentation approach generalize across different NILM algorithms and datasets?
• RQ5 To what extent does augmentation mitigate the negative impact of data sparsity and low sampling rates on NILM performance?

Key findings

• Device-specific interpolation achieved the highest average F-score improvement, with Weiss’s algorithm reaching 0.744 on cooling devices and 0.300 on cooking devices in the ECO dataset.
• Parson’s algorithm showed strong performance with device-specific and stepwise interpolation on cooling and computer appliances, achieving 0.620 and 0.485 F-scores respectively.
• The Denton-Cholette method produced smoother time series, which reduced event detection accuracy and led to lower F-scores (e.g., 0.305 for Weiss’s algorithm on iAWE), indicating over-smoothing harms performance.
• Cubic spline and stepwise interpolation showed mixed results, with cubic performing poorly on ECO (F-score 0.176 for Weiss’s algorithm), likely due to overfitting to noise or gaps.
• Lighting devices consistently showed low F-scores across all methods due to their low power consumption, highlighting a persistent challenge for NILM even with augmentation.
• The original high-quality data (1 Hz) outperformed all augmented versions in some cases, but augmentation significantly improved performance over undersampled data, with Weiss’s algorithm on iAWE increasing from 0.206 to 0.416 F-score using stepwise interpolation.