Short-term forecasting of electricity consumption using Gaussian processes

Abstract

Forecasting of electricity consumption is considered as one of the most signi cant aspect of e ective management of power systems. On a long term basis, it allows decision makers of a power supplying company to decide when to build new power plants, transmission and distri- bution networks. On a short term basis, it can be used to allocate resources in a power grid to supply the demand continuously. Forecasting is basically divided into three categories : short-term, medium-term, and long- term. Short-term refers to an hour to a week forecast, while medium-term refers to a week to a year, and predictions that run more than a year refers to long-term. In this thesis, we forecast electricity consumption on a short-term basis for a particular region in Norway using a relatively novel approach: Gaussian process. We design the best feature vector suitable for forecasting electricity consumption using various factors such as previous consumptions, temperature, days of the week and hour of the day. Moreover, feature space is scaled and reduced using reduction and normalization methods, and di erent target variables are analysed to obtain better accuracy. Furthermore, GP is compared with two traditional forecasting techniques : Multiple Back- Propagation Neural Networks (MBPNN), and Multiple Linear Regression (MLR). Finally we show that GP is as better as MBPNN and far better than MLR using empirical results.