Source-linked AI summary
Deep Sparse Coding for Non-Intrusive Load Monitoring
Shikha Singh, Angshul Majumdar
TL;DR
Energy disaggregation separates a building’s aggregate meter signal into appliance-level consumption, while prior dictionary-learning methods use one dictionary layer per device. This paper learns multiple dictionary layers for each device and reports better performance than state-of-the-art methods on two benchmark datasets.
Problem
Prior dictionary-learning approaches learn a single dictionary layer per device for the single-channel energy-disaggregation problem.
Source: §V · p. 8Method
The method learns multiple dictionary layers for each appliance and uses the resulting multi-layer basis for signal disaggregation.
Source: B. Sparse Coding · p. 3, §III · p. 4Results
Takeaways & Limitations
The authors conclude that deep sparse coding improves on state-of-the-art methods for energy disaggregation on the tested benchmarks.
Source: §V · p. 8Sparse-coding and dictionary-learning methods, including this work, cannot be used for real-time disaggregation; the authors say HMM-based techniques are more suitable for that need.
Source: §V · p. 8
Abstract
from arXiv · showhide
Energy disaggregation is the task of segregating the aggregate energy of the entire building (as logged by the smartmeter) into the energy consumed by individual appliances. This is a single channel (the only channel being the smart-meter) blind source (different electrical appliances) separation problem. The traditional way to address this is via stochastic finite state machines (e.g. Factorial Hidden Markov Model). In recent times dictionary learning based approaches have shown promise in addressing the disaggregation problem. The usual technique is to learn a dictionary for every device and use the learnt dictionaries as basis for blind source separation during disaggregation. Prior studies in this area are shallow learning techniques, i.e. they learn a single layer of dictionary for every device. In this work, we propose a deep learning approach, instead of learning one level of dictionary, we learn multiple layers of dictionaries for each device. These multi-level dictionaries are used as a basis for source separation during disaggregation. Results on two benchmark datasets and one actual implementation show that our method outperforms state-of-the-art techniques.
I. INTRODUCTION
Energy disaggregation separates a building’s combined energy signal into appliance-level consumption. The paper extends dictionary learning from one layer per appliance to multiple layers and reports better performance than state-of-the-art shallow architectures.
I. INTRODUCTION
- Residential and commercial buildings account for 40% of total energy consumption, and studies estimate that improved user behavior could avoid 20% of this consumption.Source: §I · p. 1
- Earlier approaches used finite-state models or learned appliance bases; finite-state methods performed poorly on time-varying appliances.Source: §I · p. 1
- The method extends sparse coding by learning multiple dictionary layers and using the concatenated multi-layer basis for disaggregation.Source: §I · p. 1
- The authors distinguish their deep approach from hierarchical dictionary learning, which they describe as shallow, single-level learning with hierarchically structured atoms.Source: §I · p. 1
- The proposed extension outperforms state-of-the-art shallow architectures on the REDD and Pecan Street benchmark datasets.Source: §I · p. 1
A. Deep Learning
The section surveys ways to learn a neural network’s representation layer, then describes stacking representations to form deeper architectures.
A. Deep Learning
- For a single representation layer, learning the input-to-representation weights and the representation is the challenge, while mapping the representation to the target is straightforward.The passage frames this as representation learning.Source: A. Deep Learning · p. 2
- A restricted Boltzmann machine learns network weights and a representation, while a deep Boltzmann machine stacks hidden layers above an RBM.The RBM is described as restrictive for binary data; the Gaussian-Bernoulli RBM handles values between 0 and 1 but not arbitrary real or complex inputs.Source: A. Deep Learning · p. 2
- An autoencoder learns encoder and decoder weights by minimizing reconstruction error, then removes the decoder and attaches the target after the representation layer.Its activation function is usually tanh or sigmoid, which squashes inputs to normalized values; the representation is intended to retain information for reconstruction.Source: A. Deep Learning · p. 2
- A stacked autoencoder nests autoencoders to learn multiple representation layers, but its optimization problem is complicated.The described greedy workaround learns the outer layers first, then uses their features as input to learn the inner layer; deep-network training removes the decoder and fine-tunes with backpropagation.Source: A. Deep Learning · pp. 2–3, Figure 6 · p. 2
B. Sparse Coding
Sparse coding learns a dictionary and sparse loading coefficients for each appliance, then estimates device contributions from aggregate power under a summed-load model.
B. Sparse Coding
- Each appliance is represented by a learned dictionary and sparse loading coefficients, with positivity imposed to conform to physics.Dictionary learning is non-convex and is solved by alternating between coefficient updates and dictionary updates; negative sparse-code values are set to zero.Source: B. Sparse Coding · p. 3
- During operation, aggregate power is modeled as the sum of appliance powers, assuming reactive loads only.Training data is described as readings collected over time when the smart meter logs only one device at a time.Source: B. Sparse Coding · p. 3
- Given fixed appliance bases, nonnegative loading coefficients are estimated by sparse recovery and used to compute each device’s power consumption.The coefficients are estimated from the aggregate-load model, after which per-device consumption follows from the corresponding dictionary and coefficients.Source: B. Sparse Coding · p. 3
- The authors claim that learning deeper dictionary levels improves upon the state of the art without complicated penalty terms.The claim follows their description of sparse-coding-based energy disaggregation.Source: B. Sparse Coding · p. 3
III. PROPOSED DEEP SPARSE CODING
The paper extends synthesis dictionary learning from a shallow basis to multiple layers, defining deep sparse coding. It also distinguishes this formulation from a collapsed single dictionary and outlines alternative ways to solve it.
III. PROPOSED DEEP SPARSE CODING
- Figures 8–9 frame dictionary learning as a network: D's columns connect representation Z to input X, rather than serving only as basis atoms.In this synthesis view, features and the dictionary generate the data.Source: §III · pp. 3–4, Figure 8 · p. 3, Figure 9 · p. 3
- Deep sparse coding extends shallow synthesis dictionary learning by composing multiple dictionary layers.The paper motivates deeper representations as potentially yielding more compact, abstract features, while noting the increased training-data demand and overfitting risk.Source: §III · p. 4
- The two-layer product D1D2 cannot be collapsed to a single dictionary: its tri-linear formulation yields different features from the shallow bi-linear model.This remains true even when the dimensions match.Source: §III · p. 4
- The paper considers greedy layerwise learning and an exact solution using variable splitting followed by alternating minimization.The greedy approach has no feedback between layers; the exact solution addresses the multilayer problem.Source: §III · p. 4
A. Greedy Solution
The greedy solution builds the deep model layer by layer, using alternating minimization and sparse coding steps. Its main limitation is one-way information flow without feedback to earlier layers.
A. Greedy Solution
- Greedy learning substitutes the remaining dictionary layers into each layer's problem and solves it by alternating minimization, continuing through the final layer.The first-layer coefficient is not sparse; later-layer coefficients are constrained to be sparse, making that step equivalent to sparse coding.Source: A. Greedy Solution · p. 4
- The greedy approach is easy to build from shallow dictionary-learning blocks, but information flows only forward, with no feedback from later layers.Because the task is unsupervised and has no targets, the paper says fine-tuning by back-propagation is not available.Source: A. Greedy Solution · p. 4
B. Exact Solution
The exact solution applies Split Bregman variable splitting to the multilayer objective, then uses alternating directions to create simpler subproblems. Most subproblems have closed-form solutions, while the final l1 minimization uses iterative soft thresholding.
B. Exact Solution
- Split Bregman variable splitting introduces relaxation variables to enforce equality at convergence, and alternating directions break the coupled objective into simpler subproblems.The paper illustrates the procedure with three levels, stating this entails no loss of generality for the exposition.Source: B. Exact Solution · p. 4
- Subproblems P1–P5 are linear least-squares problems with closed-form solutions; P6 is an l1 minimization solved by iterative soft thresholding.The Bregman relaxation variables are updated during each iteration.Source: B. Exact Solution · p. 5
- The Split Bregman iterations stop when the objective converges to a local minimum or when the maximum of 200 iterations is reached.The maximum-iteration limit is set to 200.Source: B. Exact Solution · p. 5
C. Energy Disaggregation
For disaggregation, the method replaces each appliance's shallow codebook with a deep, cascaded codebook. It retains the existing loading-coefficient procedure to estimate appliance energy.
C. Energy Disaggregation
- Instead of the shallow codebook used in prior sparse-coding disaggregation, the proposed method learns a deep sparse codebook for every appliance.The appliance index is denoted by the superscript i.Source: C. Energy Disaggregation · p. 5
- Learning each appliance's codebook retains nonnegative sparse coefficients and normalized codebooks as constraints.Source: C. Energy Disaggregation · p. 5
- Disaggregation proceeds as before, but each appliance uses a cascade of dictionaries; multiplying that cascade by its loading coefficients yields the appliance's energy.Source: C. Energy Disaggregation · p. 5
IV. EXPERIMENTAL EVALUATION
The evaluation follows standard smart-meter data practice and uses the REDD and Pecan Street datasets.
IV. EXPERIMENTAL EVALUATION
- The study uses standard smart-meter data for load monitoring rather than alternate signatures.The passage notes research on current sensors, derivative power signatures, and V-I trajectories, but says most NILM studies depend on standard smart-meter data.Source: §IV · p. 5
- The evaluation covers two popular datasets: REDD and Pecan Street.Source: §IV · p. 5
- The passage frames alternate load signatures as promising while retaining standard smart-meter data for these experiments.Source: §IV · p. 5
A. REDD Dataset
On REDD, the exact deep sparse-coding method achieves the strongest aggregate disaggregation accuracy; the supplied passages also report Pecan Street comparisons and device-level normalized errors.
A. REDD Dataset
- The harder testing protocol trains on four households and predicts on the remaining household; REDD house 5 is omitted for insufficient data.The compared methods are FHMM, PED, SC, discSC, and the proposed methods.Source: A. REDD Dataset · p. 5
- 66.13 aggregate disaggregation accuracy is reported for Proposed (Exact), above 62.75 for Proposed (Greedy) and 59.29 for discSC on REDD.The authors report that the exact solution is best overall; the methods use 144-100-80 atoms across three layers, versus 144 atoms for SC and discSC.Source: A. REDD Dataset · pp. 5–6, Table II · p. 6
- Table IV reports normalized error for seven common devices, with Proposed (Exact) lower than the alternatives for each listed device.Source: Table IV · p. 7
B. REDD Dataset
Pecan Street evaluation uses a subset of Dataport, and the proposed method performs best across devices while producing close qualitative estimates in the illustrated examples.
B. REDD Dataset
- The experiment uses a subset of the Dataport dataset in NILMTK format, with one-minute circuit- and building-level electricity data from 240 houses, averaged over 10-minute periods for training and testing.About 80% of homes are assigned to training and the remaining 20% to testing. Training samples contain one day of power consumed by a particular device; testing samples contain one day of total power consumed in a particular house.Source: B. REDD Dataset · p. 7
- On the Dataport subset, the method outperforms other techniques by a wide margin, with larger gains on homes whose prior disaggregation accuracy was lower.Examples named by the authors include homes 6–8, 15, and 29.Source: B. REDD Dataset · p. 7
- The proposed method yields the best normalized error for every device, while FHMM and PED perform significantly worse.Sparse coding and discriminative sparse coding are also reported as worse than deep sparse coding.Source: B. REDD Dataset · p. 7
- In Fig. 10, the proposed greedy method’s estimates are close to actual consumption, whereas discSC estimates are considerably off.The figure caption identifies the proposed greedy method on the left and discSC on the right; red is actual energy and blue is predicted energy.Source: Figure 10 · p. 7, B. REDD Dataset · p. 7
V. CONCLUSION
The paper extends per-device dictionary learning from one layer to multiple layers and reports better results on two benchmark datasets. It also identifies a real-time-use limitation and proposes future extensions.
V. CONCLUSION
- Deep sparse coding replaces a single dictionary layer per device with multiple layers and is reported as better than state-of-the-art methods on two benchmark datasets.Source: §V · p. 8
- Sparse-coding and dictionary-learning methods, including this work, cannot be used for real-time disaggregation; the authors point to HMM-based techniques as more suitable for that need.Source: §V · p. 8
- The authors propose incorporating further device assumptions into their deep framework, noting prior reports of better shallow-method results when such assumptions are made.Source: §V · p. 8
- They also propose investigating whether the technique can disaggregate specific loads, such as heavy loads including AC.Source: §V · p. 8