Smart Meter Based Load Forecasting for Residential Customers Using Machine Learning Algorithms

The focus of this thesis is the use of machine learning algorithms to perform next step short term load forecasting on fifty five households in Stavanger, Norway. A dataset containing electricity consumption data for more than one year is used to train and evaluate a Feedforward Neural Network model and a Random Forest model. Weather data, atmospheric data and calendric variables are also used to aid the forecasting task. First, the implementation of the two models is introduced. Their architectures are given and the rationale behind the design principles are explained. Then, for every household, a separate neural network and random forest model are trained using the training dataset. The models are tested using the testing dataset, to evaluate the models’ accuracy. The models were trained and tested on three different but equivalent datasets. The difference between them was the time resolution of th edata. These resolutions are 1 hour, 15 minutes and 1 day. The implemented models achieved various levels of accuracy depending on the household and the data resolution. Generally, the implemented Neural Network achieved higher accuracy than its Random Forest counterpart. It was also discovered that the resolution has a big influence on the outcome of the next step short term load forecasting task