[Paper Review] Complexity of Power Draws for Load Disaggregation

Dominik Egarter, Manfred Pöchacker+1 more|arXiv (Cornell University)|Jan 13, 2015

Smart Grid Energy Management32 references17 citations

TL;DR

This paper proposes two novel, algorithm-independent complexity measures for load disaggregation problems: appliance set complexity and time-series complexity, based on appliance power states and temporal usage patterns. Evaluated on real-world datasets using a state-of-the-art NILM algorithm, the measures successfully classify problem hardness by quantifying the difficulty of distinguishing appliances, independent of the disaggregation method used.

ABSTRACT

Non-Intrusive Load Monitoring (NILM) is a technology offering methods to identify appliances in homes based on their consumption characteristics and the total household demand. Recently, many different novel NILM approaches were introduced, tested on real-world data and evaluated with a common evaluation metric. However, the fair comparison between different NILM approaches even with the usage of the same evaluation metric is nearly impossible due to incomplete or missing problem definitions. Each NILM approach typically is evaluated under different test scenarios. Test results are thus influenced by the considered appliances, the number of used appliances, the device type representing the appliance and the pre-processing stages denoising the consumption data. This paper introduces a novel complexity measure of aggregated consumption data providing an assessment of the problem complexity affected by the used appliances, the appliance characteristics and the appliance usage over time. We test our load disaggregation complexity on different real-world datasets and with a state-of-the-art NILM approach. The introduced disaggregation complexity measure is able to classify the disaggregation problem based on the used appliance set and the considered measurement noise.

Motivation & Objective

• Address the lack of fair comparison between NILM algorithms due to inconsistent test conditions and incomplete problem definitions.
• Identify key factors influencing load disaggregation complexity, including appliance types, power states, usage patterns, and preprocessing stages.
• Develop a quantitative, algorithm-agnostic measure to assess the inherent complexity of a load disaggregation problem based on appliance characteristics and temporal behavior.
• Enable meaningful comparison of NILM performance across different datasets by decoupling algorithm performance from problem-specific difficulty.
• Demonstrate that complexity measures correlate with disaggregation difficulty but not with specific algorithm performance, validating their role as problem descriptors.

Proposed method

• Define appliance set complexity as a statistical measure of similarity between power states of different appliances, capturing the difficulty of distinguishing them based on their electrical signatures.
• Define time-series complexity as a measure of similarity between the observed noisy power signal and the set of possible appliance states, reflecting temporal variability and noise impact.
• Use real-world household power datasets (e.g., GREEND, UK-DALE) to extract power states via clustering or segmentation of active power traces.
• Apply a state-of-the-art NILM algorithm (e.g., SMD, NILMTK-based) to evaluate disaggregation performance on the same datasets used for complexity calculation.
• Compute both complexity measures for each dataset and time window, using entropy-based or distance-based similarity metrics to quantify state overlap and signal ambiguity.
• Validate the measures through case studies comparing complexity values with actual disaggregation accuracy, showing consistency in problem difficulty classification.

Experimental results

Research questions

• RQ1 To what extent does the number and similarity of appliance power states affect the inherent complexity of a load disaggregation problem?
• RQ2 How does temporal appliance usage behavior influence the complexity of the aggregated power signal, independent of the appliance set?
• RQ3 Can a complexity measure be constructed that is independent of the NILM algorithm used, enabling fair comparison across different datasets and methods?
• RQ4 How do preprocessing steps (e.g., noise filtering) affect the measured complexity of a disaggregation problem?
• RQ5 Does the proposed complexity measure correlate with actual disaggregation performance, or is it purely a problem descriptor?

Key findings

• The appliance set complexity measure effectively captures the difficulty of distinguishing between appliances based on their power state similarity, with higher complexity indicating greater ambiguity.
• The time-series complexity measure reflects the impact of appliance usage patterns and measurement noise, showing that even complex appliance sets can yield low complexity if usage is regular and predictable.
• The two complexity measures are independent: high appliance set complexity does not imply high time-series complexity, and vice versa, confirming their complementary nature.
• The proposed complexity measures are algorithm-agnostic and successfully classify the hardness of different disaggregation problems across multiple real-world datasets.
• The measures correlate with the difficulty of load disaggregation but do not predict the performance of a specific NILM algorithm, confirming their role as problem descriptors rather than performance predictors.
• Case studies show that the complexity measures remain consistent across different houses and datasets, validating their reliability as relative, not absolute, indicators of problem difficulty.