Evaluation of New Methods of Time Series Forecasting: Exploring the Use of Transformers
Student
José Pedro de Olim Faria (B.S.)
Abstract
Accurate energy poverty prediction is necessary for the development of effective social strategies aiming to solve this problem. Current methods involve classical machine learning approaches that may fail to capture the non-linearity of socio-economic features related with energy poverty. Classical methods also often require a costly process of feature engineering, where, through many iterations, the most useful set of features to a given model are discovered. Deep Learning methods solve this problem by constructing their own features from raw data, shortening the data pre-processing pipeline.
In this study, we explore the use of recent deep learning models (such as the classic Transformer and Crossformer) for energy poverty prediction using the longitudinal Household, Income and Labour Dynamics in Australia (HILDA) dataset, uncovering possible accuracy gains. The work involved preprocessing complex variables, structuring hyperparameter searches for model optimization, positional encoding, and extensive model validation.
More information
- Date
- 2026
