Neural Network for Precise Power Usage Breakdown: WaveNILM
In the realm of Smart Grids, a new architecture named WaveNILM is making waves in the field of real-time disaggregation. This innovative system, designed for the analysis of waveform input signals, promises to revolutionize the way we monitor and manage power consumption.
What is WaveNILM?
WaveNILM is a causal neural network architecture, built on top of the established Convolutional Neural Network (CNN) architecture. Its primary focus is on the disaggregation of appliance-level power usage from aggregate signals in real-time, making it a valuable tool for grid operators and consumers alike.
Key Features
- Causality: WaveNILM is designed to work with real-time data, making it suitable for online deployment without the need for future data.
- Neural-network-based: WaveNILM utilises convolutional layers tailored for waveform data, allowing it to capture both transient and steady-state characteristics.
- Efficiency: The architecture is designed to be lightweight and computationally efficient, making it an ideal choice for embedded or edge deployment.
- Accuracy: With its rich waveform input and deep learning's pattern recognition capabilities, WaveNILM generally achieves higher disaggregation accuracy, particularly on transient and complex load profiles.
Comparing WaveNILM to SSHMM
Two approaches in Non-Intrusive Load Monitoring (NILM) are WaveNILM and SSHMM. While they share the common goal of disaggregating appliance-level power usage, they differ significantly in their methodologies.
Accuracy
- WaveNILM tends to outperform SSHMM in accuracy. This is largely due to its ability to leverage high-resolution waveform data and deep neural network learning, enabling it to better capture transient events and nonlinear load behaviours.
- Multiple studies show that WaveNILM achieves lower disaggregation errors on common datasets (such as REDD or UK-DALE) compared to SSHMM, especially in real-time settings.
Efficiency and Real-Time Capability
- WaveNILM’s causal architecture enables real-time disaggregation with lower latency, making it more suitable for online NILM.
- SSHMM models can be more computationally intensive, often requiring offline batch processing or more complex inference, which hinders pure real-time deployment.
- WaveNILM’s neural network inference can be optimised on modern hardware or embedded devices, while SSHMM inference is typically sequential and less parallelisable.
Summary
| Aspect | WaveNILM | SSHMM | |------------------|--------------------------------------------|---------------------------------------| | Model Type | Causal Neural Network | Semi-Supervised Hidden Markov Model | | Input | Waveform (high-resolution electrical data)| Typically steady-state power features | | Accuracy | Higher, especially on transient-heavy data | Good on steady loads but lower overall| | Real-time? | Yes, designed for low-latency online use | Often offline or not strictly causal | | Computational Effort | Efficient, lightweight neural inference | More computationally intensive | | Suitability | Embedded/edge deployment, waveform-rich data| Traditional NILM scenarios |
In essence, WaveNILM provides a more accurate and efficient solution than SSHMM for real-time NILM disaggregation, especially when waveform data is available and low-latency inference is required.
The implementation of WaveNILM can be found on GitHub, and it's worth noting that popular datasets like REDD and Pecan Street were not used in the study. However, WaveNILM was benchmarked against the SSHMM architecture for noisy and denoised input signals.
Technology, encompassing the causal neural network architecture of WaveNILM, is being leveraged to revolutionize the field of real-time Non-Intrusive Load Monitoring (NILM) using artificial-intelligence. In comparison to SSHMM, WaveNILM achieves higher accuracy and efficiency, thereby providing a more suitable solution for embedding or edge deployment and waveform-rich data disaggregation tasks within the realm of real-time NILM.
