Electricity Price Forecasting: The Model Debate

Large-scale renewable energy deployment has brought significant volatility to electricity systems, transforming grid operation into a complex stochastic puzzle. Accurate electricity price forecasting (EPF) isn't just a technical necessity. it's important for operational decisions like optimal bidding strategies and balancing power preparation. But, most importantly, it's about reducing economic risk and enhancing market efficiency.

Probabilistic Forecasts: The Value Proposition

Probabilistic forecasts stand out because they quantify uncertainty. They account for renewable energy's intermittency, market coupling, and regulatory changes. This information enables market participants to make decisions that minimize losses and optimize expected revenues. So the question becomes: Which models should we trust for this task?

The Model Showdown

In the quest for the most reliable day-ahead probabilistic EPF, four models have been evaluated across European bidding zones. On one side, a deterministic NHITS backbone combined with Quantile-Regression Averaging (NHITS+QRA) and a conditional Normalizing-Flow forecaster (NF). On the other, two Time Series Foundation Models (TSFMs), Moirai and ChronosX, stand ready for comparison.

The data shows that TSFMs outshine task-specific deep learning models, at least Continuous Ranked Probability Score (CRPS), Energy Score, and predictive interval calibration across varying market conditions. Yet, task-specific models aren't to be dismissed lightly. Notably, NHITS combined with QRA can achieve performance levels close to TSFMs. In some scenarios, such as when enhanced with additional informative features or adapted via few-shot learning from other markets, they even surpass their TSFM counterparts.

Computational Expense vs. Performance Gains

The benchmark results speak for themselves. While TSFMs certainly offer expressive modeling capabilities, the conventional models hold their ground remarkably well. This leads us to a critical consideration: Is the computational expense of TSFMs justified by the marginal performance improvements in PEPF?

Western coverage has largely overlooked this nuanced debate, focusing instead on the glamour of new model introductions. But in the practical world of energy markets, the choice between model types isn't just academic. It's about real-world efficiency and cost-effectiveness.

What the English-language press missed: The decision isn't simply about which model is best in a vacuum. It's about context, application, and the specific needs of the market in question. So, should energy companies shift their resources to task-specific models or stick with TSFMs? The answer is neither straightforward nor universal. It depends on the specific scenario and what each company values more: computational simplicity or the potential for slightly improved accuracy.

This ongoing debate will continue to shape the future of energy forecasting. As renewable energy becomes a larger part of our grid, the need for precise EPF will only grow. Perhaps the real takeaway is that flexibility and adaptability in model choice could be the key to success.