Modern electricity markets are highly dynamic systems. Across a day, the electricity price can fluctuate rapidly and reach peaks several factors higher than the baseline or reach lows of negative prices. This volatility requires accurate forecasters in order to make decisions in such an uncertain environment.

In this project, I developed two time-series forecasters: A day-ahead electricity price forecaster and a load forecaster (energy demand) for a local distribution grid. The two forecasters allow a market participient to make well-grounded decisions.

I compared multiple statistical and machine learning methods, such as SARIMA, XGBoost, random forest, linear regression, MLP, and a transformer model (temporal fusion transformer) with different feature sets and did a careful performance analysis. I wanted to understand where the models are accurate and where they perform poorly, and therefore investigated both the temporal performance (model accuracy across the day and seasons, daily peak accuracy and daily peak timing accuracy), as well as model calibration

In a conclusion, I found that seq2seq models, such as the temporal fusion transformers, offer clear benefits against standard regression models and recursive (autoregressive) models, which suffer from high error accumulation. Another important finding is that all models would require periodic retraining due to distribution shift, as the increasing adoption of renewable energy sources and therefore strong weather dependency will amplify the volatility of market prices and energy demand/production.