Hierarchical Forecasting Methods in Day-Ahead Electricity Consumption Forecasting
Name
Carel Kuusk
Abstract
In various applications, several time series can be organized into one hierarchy, such that lower-level time series can be aggregated into higher-level time series. Forecasting such hierarchical time series requires reconciliation of the final forecasts to ensure that the aggregation constraints present in the original time series are satisfied with the forecasted time series as well. The aim of this thesis is to develop and analyze hierarchical forecasting methods in the context of hourly electricity consumption time series. As a result, hierarchical models based on LightGBM and ridge regression are developed, and their performance is analyzed. Two complex linear reconciliation methods – OLS and Minimal Trace (MinT) reconciliation – are compared against the bottom-up method, and the severe limitations of the OLS and MinT approaches are discovered. Limitations arise due to the electricity consumption forecasting error covariance structure. However, the analyzed reconciliation methods can be used to find forecasts for intermediary levels in the hierarchy.
Graduation Thesis language
English
Graduation Thesis type
Master - Data Science
Supervisor(s)
Meelis Kull
Defence year
2024