Source-linked AI summary
Sequence-to-point learning with neural networks for nonintrusive load monitoring
Chaoyun Zhang, Mingjun Zhong, Zongzuo Wang, Nigel Goddard, Charles Sutton
TL;DR
Energy disaggregation is difficult because appliance contributions to aggregate mains readings are inherently unidentifiable, while long sequences complicate neural-network prediction. The paper proposes sequence-to-point learning, which predicts one appliance value from a mains window, and shows improved performance on real-world NILM datasets, alongside learned appliance signatures.
Problem
Energy disaggregation remains unidentifiable, and sliding-window neural networks create issues because predictions must be combined across windows.
Method
Sequence-to-point learning trains a neural network to predict the target appliance's midpoint value from a mains window.
Results
84% lower MAE and 92% lower SAE were achieved by seq2point versus seq2seq(Kelly) on UK-DALE, while seq2point improved over seq2seq on REDD.
Takeaways & Limitations
Neural networks learn meaningful appliance signatures, including change points, usage durations, and power levels, that support energy disaggregation.
Abstract
from arXiv · showhide
Energy disaggregation (a.k.a nonintrusive load monitoring, NILM), a single-channel blind source separation problem, aims to decompose the mains which records the whole house electricity consumption into appliance-wise readings. This problem is difficult because it is inherently unidentifiable. Recent approaches have shown that the identifiability problem could be reduced by introducing domain knowledge into the model. Deep neural networks have been shown to be a promising approach for these problems, but sliding windows are necessary to handle the long sequences which arise in signal processing problems, which raises issues about how to combine predictions from different sliding windows. In this paper, we propose sequence-to-point learning, where the input is a window of the mains and the output is a single point of the target appliance. We use convolutional neural networks to train the model. Interestingly, we systematically show that the convolutional neural networks can inherently learn the signatures of the target appliances, which are automatically added into the model to reduce the identifiability problem. We applied the proposed neural network approaches to real-world household energy data, and show that the methods achieve state-of-the-art performance, improving two standard error measures by 84% and 92%.
Energy disaggregation
Energy disaggregation recovers individual appliance consumption from aggregate mains readings, which are modeled as the sum of appliance signals, unknown usage, and noise. The problem remains unidentifiable, motivating models that incorporate appliance-specific and household-level domain knowledge.
- Energy disaggregation aims to recover individual appliance signals from mains measurements of total household power.
- The mains signal is modeled as the sum of selected appliance readings, unknown usage, and Gaussian noise.
- The problem remains unidentifiable even when inference algorithms estimate appliance signals from the additive factorial hidden Markov model.
- Prior approaches reduce identifiability by incorporating local appliance information and global household information into the model.
Sequence-to-sequence learning
Sequence-to-sequence learning maps sliding windows of mains readings to corresponding windows of appliance readings. Because long signal sequences create computational and optimization difficulties, overlapping predictions are combined by averaging them.
- Sequence-to-sequence NILM models map sliding mains windows to appliance-power windows with matching timestamps.
- Training can use a subset of possible windows to reduce computational complexity.
- For interior time steps, the model averages predictions from every sliding window containing that time step.
Sequence-to-point learning
Sequence-to-point learning predicts the midpoint appliance value from each mains window, simplifying sliding-window inference while targeting the midpoint distribution directly. The paper analyzes this design and shows that, with the same architecture, it provides a tighter approximation than sequence-to-sequence learning.
- Sequence-to-point learning: Sequence-to-point learning maps each mains window to only the midpoint value of the corresponding appliance window.The midpoint is x_τ, where τ = t + floor(W/2), and is modeled as a nonlinear regression of the mains window.
- Sequence-to-point learning: A single prediction is produced for every appliance reading, avoiding the need to average predictions from overlapping windows.The full input sequence is padded at both endpoints before applying the sliding-window model.
- Sequence-to-point learning: Both methods can use the same convolutional architecture, while the paper trains them using all sliding windows rather than heuristic sample selection.This separates the sequence-to-point comparison from differences in network architecture or training-window selection.
- Posterior distribution estimators: The method directly optimizes the midpoint value rather than jointly fitting all outputs in an appliance window.In the posterior-estimation analysis, sequence-to-point focuses optimization on p(x_τ|Y, θ), whereas sequence-to-sequence models the full window distribution.
- Posterior distribution estimators: Theorem 1 states that, under the same architecture, sequence-to-point has a tighter approximation to the target midpoint distribution than sequence-to-sequence learning.The result is expressed as φ_τ(θ*) ≤ φ_τ(eθ), comparing the optimized midpoint divergences of the two methods.
Experiments
Experiments compare AFHMM, seq2seq(Kelly), seq2seq, and seq2point on UK-DALE and REDD, using appliance-level MAE and SAE. Seq2point achieves the strongest reported overall performance, while feature-map analyses indicate that CNNs learn appliance signatures such as amplitude and duration.
- Experimental setup: The experiments compare AFHMM, seq2seq(Kelly), seq2seq, and seq2point on UK-DALE and REDD household energy data.The data sets contain appliance-level and whole-house measurements from UK and US homes, with separate training and test houses.
- Evaluation metrics: MAE measures per-timestep appliance-power error, whereas SAE measures total daily energy error and is less affected by isolated inaccurate predictions.SAE can therefore reflect daily-usage accuracy even when individual timestep predictions are less accurate.
- UK-DALE results: 84% lower MAE and 92% lower SAE: seq2point improves over seq2seq(Kelly) overall on UK-DALE.Seq2seq also outperforms the comparison methods, while seq2point wins over seq2seq for three of four appliances and overall.
- REDD results: 11% lower MAE and 24% lower SAE: seq2point improves over seq2seq overall on REDD.Seq2point outperforms seq2seq for most appliances and overall on this data set.
- Visualization of latent features: CNN feature maps indicate that the networks learn appliance signatures involving signal amplitude, usage duration, and characteristic activation patterns.For kettles, manually changing amplitude or extending usage beyond 8 minutes alters the network’s inferred suitability for that appliance.
Conclusions
The paper proposes sequence-to-point neural-network learning for energy disaggregation and evaluates it on real-world data. It reports better performance than sequence-to-sequence learning and shows that the networks learn meaningful appliance signatures.
- Conclusions: The paper proposes sequence-to-point learning with neural networks for energy disaggregation.The proposed schemes are applied to real-world data sets.
- Conclusions: Sequence-to-point learning outperforms previous work using sequence-to-sequence learning.The conclusion summarizes the reported comparative result without restricting it to a single data set.
- Conclusions: Visualized feature maps show that the neural networks learn meaningful appliance signatures crucial for energy disaggregation.These learned features connect the model’s predictions to appliance-specific signal structure.