Demand Response Management Using Machine Learning Methods
The global shift from fossil and nuclear energy toward renewable sources brings a persistent challenge: keeping electricity supply and demand in balance, since most renewables generate power when conditions allow rather than when it’s needed. Demand Response (DR) addresses this by giving end-users incentives, such as time-based pricing, to shift their consumption — but this asks for a level of involvement that can be inconvenient, so machine learning has long been explored as a way to lower that barrier by monitoring household electricity use and turning it into recommendations or automated actions.
This thesis structures that process with the Machine Learning Demand Response Model (MLDR), which breaks it down into four steps: data monitoring, appliance identification, appliance usage segmentation, and appliance usage prediction, and contributes new approaches at each stage. A new monitoring system was built and used to collect DEDDIAG, an openly published dataset of 50 individual appliances across 15 homes, recorded at 1 Hz over periods of up to 3.5 years and enriched with usage annotations and household demographics. For appliance identification, the thesis presents both a low-sample-rate method using wavelet features with a k-Nearest-Neighbor classifier, and a high-sample-rate method that converts voltage-current cycles into Recurrence Plots classified via a Convolutional Neural Network with Spatial Pyramid Pooling. To segment individual appliance usage events — a step largely overlooked in prior research — the thesis introduces a new evaluation metric, the Jaccard-Time-Span-Event-Score (JTES), and a Support Vector Machine-based segmentation algorithm. Finally, a statistical model for predicting future appliance usage, based on time of day and time since last use, is evaluated on both DEDDIAG and the GREEND dataset.