Explainability-Informed Feature Selection and Performance Prediction for Nonintrusive Load Monitoring.

Area of Science:

• Energy Systems
• Artificial Intelligence
• Machine Learning

Background:

• Smart metering enables high-resolution energy data, crucial for accurate billing and demand response.
• Nonintrusive load monitoring (NILM) uses this data to identify individual appliance energy consumption.
• Existing NILM models often lack trustworthiness and explainability, hindering user adoption and model improvement.

Purpose of the Study:

• To develop a trustworthy and explainable NILM approach using interpretable machine learning models.
• To enhance NILM model performance by leveraging feature importance for appliance classification.
• To predict model performance on unseen data and minimize testing time.

Main Methods:

• Utilized a naturally interpretable decision tree (DT)-based approach for multiclass NILM classification.
• Employed explainability tools to determine local and global feature importance for each appliance.
• Designed a feature selection methodology informed by explainability to optimize prediction on unseen data.

Main Results:

• Explainability-informed feature selection improved toaster classification from 65% to 80%.
• Optimized classifier configurations (e.g., three-classifier for kettle, microwave, dishwasher) significantly boosted performance.
• Dishwasher classification improved from 72% to 94%, and washing machine from 56% to 80%.

Conclusions:

• Interpretable decision trees and explainability tools enhance NILM trustworthiness and performance.
• Feature selection based on explainability is effective for improving appliance classification accuracy.
• Tailored classifier configurations improve NILM performance for specific appliance sets.