Household electricity load forecasting for demand side management applications

As part of the energy transition, many residents have installed rooftop solar panels and started using more electrical appliances, such as electric vehicles (EVs) and heat pumps. This decentralized production and consumption puts pressure on the current electricity grid. To flatten load peaks and prevent overloading of grid components, demand side management (DSM) techniques can be employed to match electricity supply and demand. Some DSM techniques require house load predictions. However, there are challenges regarding the application of load predictions, the variation between household loads, and the volatility of a load within a single household. To overcome these challenges, this thesis proposes an online “rolling horizon” forecasting model that is continuously being trained on the smart meter data of a single household. It can therefore autonomously adapt to the load profile behaviour of that household and to structural changes in electricity consumption over time. The proposed forecasting model uses artificial neural networks (ANNs) to make day ahead predictions of the house load with a 15-minute granularity. Four different forecasting model architectures are compared (on neighbourhood- and house-level) that apply ANNs in a different way. Finally, this thesis examines the impact of the predictions on a DSM algorithm.