Statistical electricity price forecasting: A structural approach

TL;DR

This paper introduces a structured approach to electricity price forecasting that incorporates domain knowledge, leading to improved accuracy over unstructured models, especially in dynamic European market conditions.

Contribution

It demonstrates how encoding domain knowledge through a structured model improves forecasting accuracy with limited data compared to purely statistical methods.

Findings

NRMSE reduced by 0.1 during daytime hours

Significant improvements in the first day of the forecast horizon

Structured models adapt better to changing market conditions

Abstract

The availability of historical data related to electricity day-ahead prices and to the underlying price formation process is limited. In addition, the electricity market in Europe is facing a rapid transformation, which limits the representativeness of older observations for predictive purposes. On the other hand, machine learning methods that gained traction also in the domain of electricity price forecasting typically require large amounts of data. This study analyses the effectiveness of encoding well-established domain knowledge to mitigate the need for large training datasets. The domain knowledge is incorporated by imposing a structure on the price forecasting problem; the resulting accuracy gains are quantified in an experiment. Compared to an "unstructured" purely statistical model, it is shown that introducing intermediate quantity forecasts of load, renewable infeed, and cross-border exchange, paired with the estimation of supply curves, can result in a NRMSE reduction by 0.1 during daytime hours. The statistically most significant improvements are achieved in the first day of the forecasting horizon when a purely statistical model is combined with structured models. Finally, results are evaluated and interpreted with regard to the dynamic market conditions observed in Europe during the experiment period (from the 1st October 2022 to the 30th April 2023), highlighting the adaptive nature of models that are trained on shorter timescales.