A Study of Electrical Load Forecasting by Synergetic Time Series Clustering in a Temporal Convolutional Network

Abstract

In this thesis time series forecasting is reviewed and performed on electrical load time series. The main dataset that is used consists of 4074 load time series, each collected from a secondary substation. The time series in this set cover hourly observations spanning more than 2 years, and these time series all have different patterns, some being more similar to each other. We explore how we can use this similarity and dissimilarity in order to group the time series, and find that a clustering-like behaviour would be desired. We also explore different possibilities with regards to forecasting the time series, and find that temporal convolutional networks (TCNs) present good promises for doing such tasks. Two methods, in addition to a simple baseline, are then presented and used, a regular TCN, and the TCN-based model DeepGLO, which combines TCNs with clustering-like behaviour by use of matrix factorization. Ultimately we find that the regular TCN outperforms DeepGLO, and speculate that TCNs themselves might exhibit behaviour similar to clustering.