Representation Learning in Power Time Series Forecasting

Abstract

Renewable energy resources have become a fundamental part of the electrical power supply in many countries. In Germany, renewable energy resources contribute up to $29\%$ to the energy mix. However, the challenges that arise with the integration of those variable energy resources are various. Some of these tasks are short-term and long-term power generation forecasts, load forecasts, integration of multiple numerical weather prediction (NWP) models, simultaneous power forecasts for many renewable farms and areas, scenario generation for renewable power generation, and the list goes on. All these tasks vary in difficulty depending on the representation of input features. As an example, consider formulas that express laws of physics and allow cause and effect of otherwise complex problems to be calculated. Similar to the expressiveness of such formulas, deep learning provides a framework to represent data in such a way that it is suited for the task at hand. Once the neural network has learned such a representation of the data in a supervised or semi-supervised manner, it makes it possible to utilize this representation in the various available tasks for renewable energy. In our chapter, we present different techniques to obtain appropriate representations for renewable power forecasting tasks, showing the similarities and differences of deep learning-based techniques to traditional algorithms such as (kernel) PCA. We support the theoretical foundations with evaluations of these techniques found on publicly available datasets for renewable energy, such as the GEFCOM 2014 data, Europe Wind Farm data, and German Solar Farm data. Finally, we give a recommendation that assists the reader in building and selecting representation learning algorithms for domains other than renewable energy.