Real‐time load disaggregation algorithm using particle‐based distribution truncation with state occupancy model

Abstract

A new particle-based distribution truncation method with a durationdependent hidden semi-Markov model for non-intrusive appliance load monitoring is presented. Unlike earlier works, the approach keeps track of a set of states without prematurely pruning away intermediately ranked states. It also enables appliance state duration characteristics to be incorporated in a straightforward manner. Results show that the approach outperforms both the Viterbi algorithm and conventional particle-filtering methods.Introduction: Non-intrusive appliance load monitoring (NIALM) uses aggregate information to mathematically infer the power consumption of contributing appliances [1]. Past works in the field typically employed methods to find the best point estimates at each time step. Prime examples include the maximum a posteriori probability [2] and the Viterbi algorithm [3] which only keep the best previous state for each possible current state value. This is undesirable because the most probable previous state does not necessarily correspond to the true state. In addition, measurements obtained so far may not provide sufficient information for good decisions to be made. Very recently, particle-filtering (PF) with a factorial hidden Markov model (FHMM) for appliance state tracking was proposed [4]. However, PF is encumbered by the problem of low particle diversity [5]. In the case of NIALM, where a given observed aggregate signal can be explained by many different combinations of component signals, this problem is wasteful considering that particle slots are premium and identical offsprings (with similar state sequence) offer no benefits but take up particle slots that could otherwise be occupied by particles with different state sequences. In addition, the use of FHMM implicitly assumes that durations of all appliance states are geometrically distributed, which in reality is clearly not the case.In this Letter, we propose a new particle-based distribution truncation (PDT) method with a duration-dependent (DD) state transition model which does away with such limitations. Our model is inspired by Vaseghi [6], and unlike typical hidden semi-Markov models [7] our formulation allows appliance state duration characteristics to be easily incorporated into the state transition model. The resulting inference stage is computationally efficient with no explicit search over all possible state durations.Tests were performed on a public dataset of real homes [8]. Results show that our approach performs better than the Viterbi algorithm and the PF method as outlined in [4].

Automated Non-Intrusive Load Monitoring System Using Stacked Neural Networks and Numerical Integration

“…The impact of these wrong classifications is mitigated during the disaggregation process when impossible sequences of events are identified and discarded. Table 6 presents a comparison of the results of this study with studies that use other techniques to perform NILM tasks, like [32], with Integer Programming (IP) and Constraint Programming (CP) HMM, [28] with Particle-based Distribution Truncation (PDT) HMM, [40] with Graph Signal Processing (GSP) and [21] with a long short-term memory (LSTM) DNN, a denoising autoencoder (DAE) DNN and a DNN which regresses the start time, end time and average power demand of each appliance activation (RECT).…”

“…al. [28] proposed a particle-based distribution truncation (PDT) method with a duration dependent hidden semi-Markov model (HSMM) for NILM. The method allows appliance state duration characteristics to be incorporated into the state transition model, eliminating the need to search over all possible state durations during the inference stage.…”

Integration of legacy appliances into home energy management systems

“…The HMM is commonly used to model appliances, while variants are employed to incorporate time and other elements. For example [51] uses a semi-HMM to represent more realistic appliance usage models, whose transitions are not geometrically distributed.…”

Comprehensive NILM Framework: Device Type Classification and Device Activity Status Monitoring Using Capsule Network

“…Another effective NILM solution based on gated linear unit convolutional layers is proposed in [18]. However, a novel particle-based distribution truncation method with a duration dependent hidden semi-Markov model for NILM is described in [19]. Moreover, a number of intricate signal processing techniques based NILM frameworks are presented in [20] - [24].…”