NILM Neural Network Training for Sparse Household Energy Labels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Non-intrusive load monitoring (NILM) and disaggregation of energy usage in households are challenging due to the variety of devices and limited availability of labeled data sets, making it difficult to accurately predict energy usage of target devices like electric vehicles and large appliances from general household energy usage.
Innovation Solution
A novel learning scheme using a trained machine learning model, specifically a neural network, that predicts energy usage of a target device by leveraging labeled energy usage data from multiple source locations, incorporating energy usage data from other devices to improve prediction accuracy, even with limited training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional NILM methods are used to disaggregate device energy usage from household energy usage, then the system can provide energy monitoring capabilities, but the accuracy of predicting energy usage for target devices is insufficient due to the large variety of household devices and limited labeled data sets
Solution Approach 1:
The patent applies preliminary action by pre-training machine learning models using synthetic data generated from device catalogs and simulated energy usage patterns before deploying them in real-world NILM applications. This pre-training prepares the models to handle the limited availability of actual labeled data sets, enabling accurate energy usage prediction for target devices even when real labeled data is scarce.
Solution Approach 2:
The patent employs copying by creating synthetic copies of real device energy usage patterns through simulated data generation. The system generates artificial labeled data sets that mimic real household energy consumption, allowing the machine learning models to learn from these synthetic copies when actual labeled data is insufficient, thereby improving prediction accuracy without requiring extensive real-world labeled data collection.
2Productivity
If machine learning models are trained with limited labeled data sets, then the system can be deployed faster, but the prediction accuracy for target device energy usage remains insufficient
Solution Approach 1:
The system performs preliminary training actions by pre-training models on synthetic data before deployment. This allows the models to be deployed faster with limited real labeled data while maintaining prediction accuracy, as the pre-training on synthetic data provides a strong foundation that requires less actual labeled data for fine-tuning.
Solution Approach 2:
The patent applies dynamics by implementing a two-stage training approach where the model is first trained on synthetic data and then fine-tuned on limited real labeled data. This dynamic training strategy allows the system to achieve high prediction accuracy with minimal real labeled data, balancing deployment speed with model performance by adaptively using different data sources at different training stages.
3Adaptability or versatility
If the system incorporates energy usage data from multiple device types and source locations, then the generalization capability improves, but the complexity of data processing and model training increases
Solution Approach 1:
The patent applies segmentation by organizing the training process into distinct stages: first training on synthetic data from device catalogs, then fine-tuning on real labeled data from multiple source locations. This segmented approach allows the system to incorporate diverse data from multiple devices and locations while managing complexity through structured, phased training rather than attempting to process all data simultaneously.
Solution Approach 2:
The system employs universality by using a unified machine learning model architecture that can handle multiple device types and data sources through a single training framework. The model is designed to be multi-functional, processing energy usage data from various devices and locations using the same underlying structure, thereby improving generalization capability without proportionally increasing system complexity.
Data Source
AI summary
Embodiments implement non-intrusive load monitoring using a novel learning scheme. A trained machine learning model configured to disaggregate device energy usage from household energy usage can be stored, where the machine learning model is trained to predict energy usage for a target device from household energy usage. Household energy usage over a period of time can be received, where the household energy usage includes energy consumed by the target device and energy consumed by a plurality of other devices. Using the trained machine learning model, energy usage for the target device over the period of time can be predicted based on the received household energy usage.