Energy Price Forecasting with Uncertainty Estimation

Forecasting is the task of predicting future values, taking into account historical data. We forecast energy prices taking into consideration load, generation, historical prices and weather data. This thesis aims to solve this task using a Transformer model. While Transformers are typically used for NLP tasks, there have recently been some successful applications of Transformers for forecasting. In addition to predicting the prices, we try to estimate the uncertainty of this prediction by probabilistic forecasting. We experiment with different types of distributions and determine which distribution would be the best for probabilistic forecasting. We also study how probabilistic forecasting affects the model’s results compared to deterministic forecasting. We find that Transformers outperform other deep learning models. Additionally, probabilistic forecasting helps improve the accuracy in some cases and is helpful in understanding the uncertainty in the model’s prediction.