[Paper Review] On the Feasibility of Generic Deep Disaggregation for Single-Load Extraction

Karim Said Barsim, Bin Yang|arXiv (Cornell University)|Feb 5, 2018

Smart Grid Energy Management30 references43 citations

TL;DR

The paper proposes a fixed-architecture, generic deep disaggregation model based on a fully convolutional neural network to extract single-load activation profiles across multiple loads, evaluated on UK-DALE data.

ABSTRACT

Recently, and with the growing development of big energy datasets, data-driven learning techniques began to represent a potential solution to the energy disaggregation problem outperforming engineered and hand-crafted models. However, most proposed deep disaggregation models are load-dependent in the sense that either expert knowledge or a hyper-parameter optimization stage is required prior to training and deployment (normally for each load category) even upon acquisition and cleansing of aggregate and sub-metered data. In this paper, we present a feasibility study on the development of a generic disaggregation model based on data-driven learning. Specifically, we present a generic deep disaggregation model capable of achieving state-of-art performance in load monitoring for a variety of load categories. The developed model is evaluated on the publicly available UK-DALE dataset with a moderately low sampling frequency and various domestic loads.

Motivation & Objective

• Motivate data-driven, load-agnostic NILM methods that avoid per-load hyper-parameter tuning.
• Propose a generic deep disaggregation model with fixed architecture and hyper-parameters.
• Assess performance across diverse load categories using public UK-DALE data.
• Benchmark against prior per-load models to demonstrate generalization across loads and buildings.

Proposed method

• Use a fully convolutional neural network with 46 layers (input, 40 encoding/decoding, 4 representation, output).
• Model outputs posterior probabilities for a K-length window of activation for a target load using binary cross-entropy loss.
• Train with non-overlapping windows and ground-truth activation profiles derived from thresholding real power signals.
• Employ dilated convolutions, batch normalization, LReLU activations, activation noise, and skip connections to learn representations.
• Optimize parameters with Adam (with Nesterov momentum) and initialize from a zero-mean uniform distribution.
• Ground truth activation profiles are defined by thresholds and minimum on/off durations per load.

Experimental results

Research questions

• RQ1 Can a single, fixed-architecture deep model achieve state-of-the-art-like disaggregation across multiple load types without load-specific hyper-parameter tuning?
• RQ2 How well does a generic model generalize to unseen loads and buildings when trained on public NILM data?
• RQ3 What are the performance and evaluation characteristics of a fully convolutional NILM approach compared to load-specific baselines?

Key findings

• The proposed generic model outperforms prior work across all loads on the UK-DALE dataset chosen for evaluation.
• Performance variability across loads is quantified using multiple metrics beyond simple accuracy, addressing class-imbalance concerns.
• Across loads, the model achieves high f-measure and MCC for several categories (e.g., WM, KT, TS, MC) and strong AUC-like indicators in several others.
• The model maintains competitive or superior performance when transferring from known loads to unseen loads in new buildings, relative to a regression-based baseline.
• Activation-profile based evaluation provides near real-time feedback potential, since activations can be inferred from partial segments of the aggregate data.