A Bayesian Optimization-Based AutoML Framework for Non-Intrusive Load Monitoring

Technical Summary: A Bayesian Optimization-Based AutoML Framework for Non-Intrusive Load Monitoring

Problem Statement
Non-Intrusive Load Monitoring (NILM), or energy disaggregation, aims to estimate the power consumption of individual appliances by analyzing a home's aggregate electricity usage. While Intrusive Load Monitoring (ILM) offers high accuracy through dedicated sensors, it is impractical due to high installation costs and privacy concerns. NILM provides a cost-effective alternative but relies heavily on complex machine learning (ML) pipelines. Traditionally, developing these pipelines requires significant domain expertise in data science to manually select models and tune hyperparameters. This manual process is time-consuming, often evaluates only one model at a time, and can lead to suboptimal configurations. The paper addresses the need for an automated, accessible framework that can systematically identify optimal ML models and hyperparameters for energy disaggregation without requiring advanced ML expertise from the practitioner.

Methodology
The authors propose AutoML4NILM, a flexible, open-source framework that integrates Automated Machine Learning (AutoML) with Bayesian Optimization to solve the NILM problem.

1. Framework Architecture: The system automates the design of ML pipelines by treating model selection and hyperparameter tuning as a search problem. It defines a comprehensive search space encompassing 11 distinct algorithms, including:

   • Traditional ML: Decision Trees (DT), Random Forests (RF).
   • Deep Learning: Fully-Connected Neural Networks, Gated Recurrent Units (GRU), Window-based GRU, Seq2Seq, Seq2Point, Long Short-Term Memory (LSTM), and Denoising Autoencoders (DAE).
   • Probabilistic/Combinatorial: Factorial Hidden Markov Models (FHMM) and Combinatorial Optimization (CO).

2. Hyperparameter Search Space: The framework explores a diverse set of hyperparameters for these models, including:

   • Tree-based: Criterion, minimum samples split, number of estimators.
   • Neural Networks: Optimizers (Adam, Nadam), learning rates, loss functions, number of layers, dropout probability, sequence length, and window size.

3. Optimization Strategy: The core engine utilizes Bayesian Optimization via the Hyperopt library. Unlike Grid or Random Search, which treat evaluations independently, Hyperopt employs Sequential Model-Based Optimization (SMBO) with Tree-structured Parzen Estimators (TPE). This approach uses information from previous evaluations to intelligently navigate the search space, optimizing the Expected Improvement (EI) criterion to find the best model configuration with fewer iterations.

4. Experimental Setup:

   • Dataset: The UK-DALE dataset, containing aggregate and appliance-level data from five UK homes (2013–2015).
   • Preprocessing: Data was split into training (25 days), validation (7 days), and test (remainder) sets. A 60 Hz sampling rate was selected for disaggregation to balance detail and computational cost.
   • Metrics: Performance was evaluated using Mean Absolute Error (MAE) for regression accuracy and Classification Accuracy for state detection (ON/OFF).
   • Constraints: Experiments were limited to 20 epochs and a maximum of 30 hyperparameter evaluations per algorithm due to computational resource constraints (server crashes and unresponsiveness occurred with higher evaluation counts).

Key Results
The study benchmarked the AutoML4NILM framework against the 11 supported algorithms:

• Baseline Performance: When algorithms were run individually with standard configurations, Seq2Point achieved the lowest MAE (8.31) and high accuracy (97.98%). Conversely, Combinatorial Optimization (CO) and FHMM performed poorly in terms of MAE (224 and 177.3, respectively), and Window-based GRU showed the lowest accuracy (80.9%).
• AutoML Optimization: When the Bayesian Optimization framework was applied to tune hyperparameters:
  • Seq2Point remained the superior model, achieving a best MAE of 7.12 with the configuration: window size 20, learning rate 0.01, and Nadam optimizer.
  • Other models showed varying degrees of improvement, but none surpassed the optimized Seq2Point performance.
  • The framework successfully identified optimal configurations within the limited 30-evaluation budget, demonstrating the efficiency of Bayesian Optimization over manual trial-and-error.

Significance and Claims
The paper positions AutoML4NILM as a significant step toward democratizing NILM research and deployment. Its primary contributions are:

1. Accessibility: It empowers domain practitioners to apply state-of-the-art ML techniques for energy disaggregation without requiring deep expertise in data science or manual hyperparameter tuning.
2. Efficiency: The framework demonstrates significantly improved efficiency and superior experimental performance compared to traditional manual approaches that evaluate models one by one.
3. Extensibility: The framework is designed with a modular architecture that supports the addition of new algorithms and hyperparameters, fostering further research and industry adoption.
4. Validation: By identifying Seq2Point as the optimal architecture for the UK-DALE dataset under the tested conditions, the work validates the efficacy of sequence-to-sequence models for precise, appliance-level energy disaggregation.

The authors conclude that while the current implementation is limited by computational constraints (preventing exhaustive search spaces), the framework establishes a structured reference architecture for both academic and industrial NILM applications, streamlining the path from raw data to optimized energy monitoring solutions.