| dc.description.abstract | Although state-of-the-art models like linear regression and neural networks have been widely used for electricity consumption forecasting, the demand for improved prediction accuracy is still very high. A small improvement in prediction accuracy has a significant economic value for the energy industry. Gaussian processes (GPs) are becoming a more and more popular tool in machine learning, and in this thesis we will investigate how the GPs can be used for electricity consumption forecasting.
Its non-parametric nature make GPs a natural approach to addressing complex stochastic problems that are difficult to solve using the earlier parametric models. The drawback of the GP prediction model is that it requires computation which grows as , where n is the number of training points. To deal with such problems we used a novel approach known as the k-nearest neighbor (kNN) similarity search, which explores the training set and selects k elements of the dataset that are closest to a given query point. By using the output of the kNN search as a training set for GPs, the computational cost will be reduced from to , where .
The experimental sets in this thesis are based on a real life dataset obtained from Eidsiva Energy, and our goal is to forecast electricity consumption a day ahead (for the next 24 hours). In addition to the GPs, the neural networks (NNs) and local linear regression (LLR) models are also implemented and tested using test inputs over the same period of time as the GPs. Empirical analyses of the results show that the GPs outperform both the NNs and the LLR models.
Keywords: Gaussian Processes (GPs), k-Nearest Neighbors (kNN) similarity search, Neural Networks (NNs), Local Linear Regression (LLR), Short-Term Forecasting of Electricity Consumption (STFEC) | no_NO |
