ML4GE explores machine learning techniques to analyse smart-meter data. The idea is to use deep learning based clustering approaches to find similar energy consumption patterns and to put these similar patterns in the same group or cluster and then use these patterns to forecast total demand for each cluster. Recently smart meters have been widely deployed in many countries. These devices replace conventional electrical meters and are able to provide measurements for time intervals of typically less than one hour and can send these to the utility. This technology provides utilities a large amount of data and the opportunity to use this data to improve the way they run the grid.
ML4GE is a web-based tool that offers state-of-the-art methods for load profiling and total demand forecasting on time-series data in a user-friendly manner:
Using cutting-edge methods for clustering and forecasting on time series data: Improved Deep Embedded Clustering (IDEC) and a probabilistic machine learning model called State Space Model (SSM).
Real-world smart-meter dataset: London Smart Meter Dataset used to demonstrate the effectiveness of using deep embedded features for clustering and probabilistic machine learning models for forecasting.
Individual and comparative analysis of the methods.
Explanation based on intuitive visualisation and charts: Cluster visualisation, t-SNE visualisation, forecasting graphs.
The tool is targeted to show to any company or organisation the potential of unsupervised learning methods like clustering as well as forecasting on time series data. The demonstrator presents a series of visualisations and controls that allow understanding the functioning of these methods and the importance of defining properly the hyper-parameters (number of clusters) for clustering time series data and finding interesting usage patterns depending on business considerations.
Dr Oisin Boydell, University College Dublin,
Dr Seyed Naser Razavi , University College Dublin,