Interpretable Transformer Model for Capturing Regime Switching Effects of Real-Time Electricity Prices

Forecasting the Occurrence of Electricity Price Spikes: A Statistical-Economic Investigation Study

“…Authors have used interpretable models in electricity price forecasting. For example, the authors of [5] present an interpretable Transformer-based probabilistic forecasting model to predict the imbalance electricity prices in the Belgian market. Here, interpretability plays a key role because the Transformer-based attention mechanism allows it to learn the non-linear temporal (past and future) relationships between the input features and the target variable at specific time steps.…”

“…Machine learning models require a careful fine-tuning of several hyperparameters [5]; this is typically achieved by following some heuristic optimization [30]. Here, different algorithms may be used, however, in electricity price forecasting, common approaches involve Bayesian optimization algorithms [11,30].…”

“…The occurrence of electricity price spikes in the short term can be associated with one or a combination of several factors. For example, forced outages of power generators and transmission lines [4,5]; transmission line congestion [6]; forecast errors in the production of intermittent generation (i.e., wind and solar) [7]; sudden increments of the electrical demand [8]; and strategic bidding of market participants [9]. Overall, these extreme prices are shaped by the complex interactions between the technical and economic forces driving the operation of the power system and the electricity market.…”

“…Conversely, explainability refers to post hoc explanations that are generated on top of black box models [27]. In electricity price forecasting, model interpretability has recently started to gain inertia; some of these models are based on interpretable neural networks, e.g., see the studies from [5,28]. Other works, like [24], introduce a form of model explainability as a post-processing stage.…”

“…We evaluate the performance of all models against both, data-driven and economic measures. Furthermore, inspired by the recent works in interpretable machine learning in electricity price forecasting [5,28], we explore the costs and benefits of interpretability in price spike occurrence forecasting. Here, we use the decision rules generated from an interpretable model [31] to predict the occurrence of spiky prices.…”