Studying the relationship between window size and other hyperparameters in deep learning-based energy consumption forecasters

RWTH Publications (RWTH Aachen)

Abstract

The forecasting of time series data has brought numerous benefits since its inception in statistics. Specifically, the energy sector has witnessed the deployment of Machine Learning (ML) algorithms. As electrical grids are dynamic systems that need to be balanced, forecasting their fluctuations in demand is important. However, the notorious difficulty with time series analysis lies in window size selection. While training ML models, engineers must select an uninterrupted window from which the model predicts the next value. This thesis investigated the relationship between the window size and other hyperparameters in deep learning-based forecasters. Subsequently, Bayesian Optimisation (BO) was implemented using novel surrogate models such as the combined surrogate model, which balances the strengths and weaknesses of a Random Forest (RF) and a Gaussian Process (GP). In some cases, this new combined model outperformed existing surrogate models. Additionally, conformal methods were integrated into the SMAC library, but their performance in optimising forecasters did not beat that of pre-existing methods. Using the optimisation data collected, an analysis of the importance of the window size and its influence on performance in combination with other hyperparameters was conducted.

Authors 0

1. Author list not loaded yet.