The increasing adoption of highly variable renewable energy has introduced unprecedented volatility into the National Electricity Market (NEM), rendering traditional linear price forecasting models insufficient. The Australian Energy Market Operator (AEMO) spot price forecasts often struggle during periods of volatile demand, renewable variability, and strategic rebidding. This study evaluates whether transformer architectures can improve intraday NEM price forecasting. Using 34 months of market data and weather conditions, several transformer variants, including encoder–decoder, decoder-only, and encoder-only, were compared against the AEMO’s operational forecast, a two-layer LSTM baseline, the Temporal Fusion Transformer, PatchTST, and TimesFM. The decoder-only transformer achieved the best accuracy across the 2–16 h horizons in NSW, with nMAPE values of 33.6–39.2%, outperforming both AEMO and all baseline models. Retraining in Victoria and Queensland produced similarly strong results, demonstrating robust regional generalisation. A feature importance analysis showed that future-facing predispatch and forecast covariates dominate model importance, explaining why a decoder-only transformer variant performed so competitively. While magnitude estimation for extreme price spikes remains challenging, the transformer models demonstrated superior capability in delivering statistically significant improvements in forecast accuracy. An API providing real-time forecasts using the small encoder–decoder transformer model is available.

The increasing adoption of highly variable renewable energy has introduced unprecedented volatility into the National Electricity Market (NEM), rendering traditional linear price forecasting models insufficient. The Australian Energy Market Operator (AEMO) spot price forecasts often struggle during periods of volatile demand, renewable variability, and strategic rebidding. This study evaluates whether transformer architectures can improve intraday NEM price forecasting. Using 34 months of market data and weather conditions, several transformer variants, including encoder–decoder, decoder-only, and encoder-only, were compared against the AEMO’s operational forecast, a two-layer LSTM baseline, the Temporal Fusion Transformer, PatchTST, and TimesFM. The decoder-only transformer achieved the best accuracy across the 2–16 hour horizons in NSW, with nMAPE values of 33.6–39.2%, outperforming both AEMO and all baseline models. Retraining in Victoria and Queensland produced similarly strong results, demonstrating robust regional generalisation. A feature importance analysis showed that future-facing predispatch and forecast covariates dominate model importance, explaining why a decoder-only transformer variant performed so competitively. While magnitude estimation for extreme price spikes remains challenging, the transformer models demonstrated superior capability in delivering statistically significant improvements in forecast accuracy. An API providing real-time forecasts using the small encoder-decoder transformer model is available at https://nem.redaxe.com

Price spikes are rare but economically significant events observed across electricity, financial, commodity, and cryptocurrency markets. Their abrupt magnitude, heavy-tailed distributions, and severe class imbalance make them difficult to forecast using conventional time-series methods. This systematic literature review, conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) framework, synthesises recent deep learning approaches to forward-looking price-spike prediction and classification. Searches of Scopus, Web of Science, and IEEE Xplore identified studies published between 2020 and 2026. Following screening and full-text eligibility assessment of approximately 300 studies, only 20 met the inclusion criteria and were included in the final synthesis, comprising 19 peer-reviewed papers and one doctoral thesis. The review develops a structured taxonomy spanning spike definitions, task formulations, model architectures, input design, and evaluation practices. A central finding is that predictive performance is driven more by problem formulation, label construction, and evaluation design than by model architecture. While architectures have diversified to include recurrent networks, transformers, graph neural networks, and hybrid frameworks, improvements are often attributable to differences in how the prediction problem is defined rather than the models themselves. Key limitations stem from inconsistent spike definitions and insufficient treatment of class imbalance, leading to a misalignment between modelling objectives and evaluation practices, further exacerbated by the absence of standardised benchmarks. These issues hinder comparability and can lead to overstated model performance by masking poor detection of rare but economically critical spike events. The review therefore identifies clear directions for future research, including standardised spike labelling, adoption of rare-event-appropriate evaluation frameworks, and problem formulations that explicitly target extreme-event prediction.