Introduction
Energy Disaggregation, or Non-Intrusive Load Monitoring (NILM), aims to decompose a single aggregate power stream from a building's main breaker into individual appliance consumption profiles. This task is traditionally modeled using Factorial Hidden Markov Models (FHMMs). However, FHMMs face a significant computational hurdle: as the number of appliances ($C$) increases, the state space grows exponentially ($2^C$), making standard learning algorithms like Baum-Welch computationally prohibitive ($O(2^{2C}T)$).
The Variational BOLT Framework
To overcome the exponential complexity of FHMMs, this paper introduces Variational BOLT, an unsupervised learning framework that combines deep recurrent neural networks with variational inference. The core of Variational BOLT lies in several key innovations:
- Variational Approximation via Stacked LSTM: Instead of exact inference, the model uses a Stacked Long Short-Term Memory (LSTM) network to approximate the filtering distribution—the probability of appliance states given the observed power data.
- Complexity Reduction: By assuming posterior independence between appliances and enforcing sparsity (assuming only a few devices are active simultaneously), the computational complexity is reduced from exponential to quasilinear ($O(C^{\epsilon+1}T)$).
- Direct Expectation Approximation: The authors address the non-differentiability of discrete "on/off" states by designing a loss function that penalizes the neural network for deviating from an approximation of the true filtering distribution, bypassing the need for the standard re-parameterization trick.
- Graphical Model Integration: The framework maintains the underlying structure of an FHMM while using the neural network for efficient inference.
Experimental Evaluation
The model was evaluated on the Reference Energy Disaggregation Data set (REDD), a standard benchmark containing high-frequency waveforms and low-frequency circuit-level ground truth.
Key Results
- Unsupervised Performance: Variational BOLT achieved a General Signal Processing Accuracy (GSPA) of 0.63, significantly outperforming the previous state-of-the-art unsupervised method, NFHMM, which scored 0.25.
- Competitive Accuracy: Despite being unsupervised, the model's precision on specific circuits (e.g., 97.3% for Washer/Dryer) approached that of supervised methods like AFAMAP (98.8%).
- Efficiency: The use of a change-point detector allows the model to process data only when significant events occur, potentially reducing the temporal processing load by three orders of magnitude.
Discussion and Future Work
While Variational BOLT represents a major step forward in unsupervised NILM, it currently assumes binary (on/off) states for appliances. This limitation makes it difficult to model complex multi-state devices like dishwashers. Additionally, the model sometimes "over-disaggregates" complex appliances into multiple virtual components. Future research may explore more flexible sparsity constraints and alternative similarity measures for electrical waveforms to further refine appliance "signatures."
Summary updated