Factorial HMM for Non-Intrusive Load Monitoring State Identification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing non-intrusive load monitoring (NILM) techniques face challenges in accurately identifying and classifying the operational states of modern electric appliances without requiring an advance database, as they struggle with parameter adjustment, handling new appliances, and managing an explosion in the number of states, leading to impractical parameter calculation.
Innovation Solution
A data processing device and method utilizing a factorial Hidden Markov Model (HMM) to model the operational states of multiple electric appliances based on time series current data, eliminating the need for an advance database by obtaining model parameters and adjusting them dynamically, while constraining the observation probability parameter to be non-negative and fixing the number of states to two (on and off states).
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If discriminative models (support vector machine, Adaboost) are used for NILM, then discrimination accuracy is improved, but it requires advance database preparation and learning completion, making it unable to handle unknown electric apparatuses
Solution Approach 1:
The system performs self-learning by automatically acquiring operation data from electric apparatuses and generating learning data through pattern recognition, eliminating the need for manual advance database preparation. The system serves itself by continuously improving its discrimination capability through automated data collection and processing.
Solution Approach 2:
The system performs preliminary learning by acquiring operation data and generating learning data in advance through automated processes, so that when new electric apparatuses are introduced, the system already has the capability to handle them without requiring manual database updates.
2Adaptability or versatility
If factorial HMM is used to model operational states of multiple electric apparatuses, then the number of states explodes, making parameter calculation impractical
Solution Approach 1:
The system segments the complex multi-apparatus modeling problem into individual apparatus-level models. Each electric apparatus is modeled separately with its own HMM, and the overall system state is represented as a combination of individual apparatus states, avoiding the exponential state explosion of a unified factorial HMM.
Solution Approach 2:
The system transitions from a unified high-dimensional state space to a distributed lower-dimensional representation where each apparatus has its own state space. This dimensional decomposition allows the system to handle multiple apparatuses without the computational complexity of a factorial state space.
3Measurement precision
If change point detection is used in NILM, then on and off states can be identified, but it requires precise threshold adjustment and fails when the change point detector misses or erroneously detects changes
Solution Approach 1:
The system uses feedback mechanisms where the HMM model continuously refines its state estimates based on observed current patterns. The model adjusts its parameters automatically through learning from operation data, eliminating the need for manual threshold adjustment and reducing sensitivity to detection errors.
Solution Approach 2:
The system dynamically changes parameters through automated learning processes, where the HMM adapts its transition probabilities and observation models based on acquired operation data. This replaces static threshold-based detection with adaptive parameter-based modeling that is more robust to variations in operating conditions.
Data Source
AI summary
A data processing device includes: a data obtaining section obtaining time series data on a total value of current consumed by a plurality of electric apparatuses; and a parameter estimating section obtaining a model parameter when states of operation of the plurality of electric apparatuses are modeled by a factorial HMM on a basis of the obtained time series data.