Our Client:

A technology company specialized in energy analytics

Situation:

Our client wanted to explore the use of recent advanced techniques such as the Temporal Fusion Transformer to predict future electricity prices across Europe. European energy market is characterized by a highly dynamic nature with interconnected factors and potential dependencies across different regions, therefore using transformer architecture is sensible given the task at hand. This is of vital importance for understanding and responding to fluctuations in the market. The scope of our study included both hourly and daily price predictions, reflecting the complexity and continual change inherent to the energy sector.

Task:

DSL was commissioned to build a time series prediction model employing the cutting-edge transformer architecture. Our system was designed to process past electricity prices from various European countries, providing future price forecasts for each respective nation within the dataset (24 nations from Europe). We enhanced the model's performance by implementing feature augmentation using logarithmic differences and cyclical representations of time (including day of the week, day of the month, and hour in the day). The model encodes 20 past days of data and predict the electricity price for the following 3 days, with either daily or hourly forecast. This was accomplished by employing established libraries like pandas and PyTorch Forecasting, widely recognized for their effectiveness in data science and time series analysis for prediction.

Result:

The training data spanned from 2015 to 2023, and we reserved the final 20 days of the dataset for evaluation. Although the prediction outcomes were satisfactory, they were not without fault, a result that is understandable given the intricacy of deriving precise forecasts solely based on data. The model was capable of capturing daily seasonality in hourly predictions and discerning trends in daily predictions, but these were still subject to some degree of error and uncertainty.

We compared the results when predicting raw prices and when predicting changes of prices (through log-diff representations). We verified that for both hourly and daily predictions, performances improve when predicting the change in price, both when measuring the change and the actual predicted price based on the change.

While the complexity of the energy market presented substantial challenges, our approach marked a significant step in leveraging modern technology to shed light on future utility prices, positioning DSL at the forefront of predictive analysis in this ever-evolving field. It also helped setting up baselines to compare to in following works that will incorporate data of different sources and nature.