AI Tackles Electricity Price Prediction for Stable Power Markets

Scientists investigating electricity price forecasting recognise its crucial function in both power system management and informed market decisions. Runyao Yu, Derek W. Bunn (London Business School), and Julia Lin (Austrian Institute of Technology), alongside Stiasny et al., present a comprehensive review focusing on day-ahead, intraday, and balancing electricity markets. This research is significant because it addresses a gap in current understanding regarding the increasingly complex deep learning model architectures now employed in EPF. The authors introduce a novel taxonomy, dissecting models into core components to enable consistent evaluation and reveal emerging trends towards probabilistic, microstructure-centric, and market-aware designs, ultimately consolidating previous reviews and charting a path for future advancements.

This work addresses a critical need for structured analysis given the rapid proliferation of diverse model architectures and training objectives in the field.

The research introduces a taxonomy that decomposes these models into three core components, backbone, head, and loss function, allowing for consistent comparison of methodologies across different market timescales. By applying this framework, researchers have identified a clear shift towards probabilistic, microstructure-centric, and market-aware designs in electricity price forecasting.

The study highlights how deep learning is increasingly employed to capture the complex, nonlinear patterns inherent in electricity prices, which are influenced by factors like non-storability, network constraints, and weather-driven uncertainty. Initial applications focused on day-ahead forecasting using architectures such as multilayer perceptrons, long short-term memory networks, gated recurrent units, and convolutional neural networks.

Subsequent advancements have expanded the scope to encompass multi-timescale forecasting, multi-market settings, and probabilistic outputs, incorporating attention mechanisms, transformers, graph-based models, and mixture-of-experts. Analysis reveals a distinct evolution in day-ahead market models towards multi-country, multi-timestep, and multi-quantile approaches.

Intraday forecasting studies, while less numerous, are increasingly focused on orderbook-driven modelling and trajectory forecasting techniques. Notably, balancing market research prioritizes mechanism-aware pipelines and market-specific designs to account for the pronounced regime switching and unique settlement rules characteristic of this market. This comprehensive review consolidates existing studies and identifies key gaps, particularly the need for greater attention to intraday and balancing markets and the development of tailored modelling strategies for each.

Component decomposition of deep learning architectures for electricity price prediction

A unified taxonomy decomposes deep learning models for electricity price forecasting into backbone, head, and loss components to facilitate consistent evaluation across studies. This framework enabled a detailed analysis of recent trends in component design across day-ahead, intraday, and balancing markets.

The backbone, responsible for representation learning from input sequences X ∈ RL×F, where L represents the number of timesteps and F the number of features, was examined for its capacity to capture temporal, spatial, and structural patterns. Simplest backbones employed were Multi-layer Perceptrons, flattening the input sequence into a vector of size L · F, though these models forgo sequential processing.

Subsequently, the research investigated the head component, which defines the output structure for single- or multi-output forecasting. Variations in head design were observed, with models ranging from those predicting single market prices to those forecasting across multiple countries, timesteps, and uncertainty dimensions.

The loss function, encoding the forecasting objective, was categorised as either pointwise, such as Mean Absolute Error, or probabilistic, exemplified by quantile loss. A comprehensive review of 33 studies, spanning 2018 to 2023, was conducted, documenting the backbone, head, loss function, features used, and country of implementation for each.

This systematic analysis revealed a clear evolution towards multi-country, multi-timestep, and multi-quantile models in day-ahead markets. Intraday studies increasingly focused on orderbook-driven modelling and trajectory forecasting, while balancing market research favoured mechanism-aware pipelines and market-specific designs due to pronounced regime switching and heterogeneous settlement rules. The study identified a gap in attention given to intraday and balancing markets, alongside the need for microstructure-aware and trajectory-based forecasting, and market-specific design in deep learning models.

Evolution of deep learning architectures in electricity price forecasting 2018, 2022

Electricity price forecasting studies consistently employ Multi-layer Perceptrons as a foundational backbone, appearing in models across day-ahead, intraday, and balancing markets. Initial analyses reveal that in 2018, models utilising MLP backbones were prevalent in day-ahead markets, specifically in Belgium, Turkey, Lithuania, and Spain, often paired with a Simple Moving Prediction head and evaluated using Mean Absolute Error.

The research details a shift towards more complex architectures, with Long Short-Term Memory networks gaining traction in intraday forecasting by 2019, demonstrated in models applied to the Turkish market. Further investigation shows that by 2022, balancing market models began incorporating LSTM-Attention backbones, achieving Pinball loss scores in Belgium and LogLik scores in Germany.

The study highlights a growing trend of microstructure-aware designs, evidenced by the increasing use of Single-Stock Moving Prediction heads alongside various backbones. Detailed analysis of components across markets reveals that in 2023, intraday forecasting in Germany leveraged LSTM and CNN backbones, while balancing markets in the UK employed LSTM-Attention, both utilising Pinball loss functions.

The work identifies a notable increase in the application of Transformer architectures in day-ahead forecasting by 2024, specifically in Norway, Finland, and Poland, again with Pinball loss. Furthermore, the research documents the emergence of Graph Neural Networks in day-ahead markets, notably in Germany, Hungary, and Italy, paired with Mean Absolute Error evaluation. Recent advancements, extending to 2026, showcase the use of Kolmogorov, Arnold Networks and Cross-Attention Transformers for intraday forecasting in Germany and Austria, achieving Pinball loss scores, and RNNs for day-ahead forecasting in France, evaluated using Mean Squared Error.

Deep learning architectures for diverse timescales and market structures in electricity price prediction

A structured review of deep learning methods for electricity price forecasting across day-ahead, intraday, and balancing markets has been undertaken through a unified framework examining backbone, head, and loss design. This analysis reveals a discernible evolution in day-ahead forecasting towards probabilistic, multi-market, and foundation-style models, while intraday research is comparatively limited and increasingly focused on microstructure awareness and trajectory-based prediction.

Balancing markets demonstrate a preference for price-formation-aware designs, reflecting the impact of strong regime switching rules inherent in these systems. The review highlights a gap in attention given to intraday and balancing markets, suggesting that current research disproportionately concentrates on day-ahead predictions and individual national markets, limiting cross-country generalisation.

Further progress requires more advanced model structures that effectively integrate domain knowledge and market-specific mechanisms alongside the adaptability of data-driven learning approaches. Acknowledging the limitations of existing studies, the authors suggest that future work should prioritise greater attention to intraday and balancing markets.

Incorporating more sophisticated model structures, combining domain expertise with data-driven techniques, will be crucial for advancing the field and improving forecasting accuracy across all market timescales. These developments promise to enhance power system operation and facilitate more informed decision-making within electricity markets.