Multi-Horizon Electricity Price Forecasting with Deep Learning in the Australian National Electricity Market

TL;DR

This study extends electricity price forecasting in the Australian National Electricity Market from day-ahead to multi-day horizons (24h and 48h) using a diverse set of deep learning models, including standard DL architectures (LSTM, CNN-LSTM, Transformer) and state-of-the-art time-series models (TimeXer, TimeMixer, DLinear, TimesNet, iTransformer, Mamba). By evaluating all five regions with 30-minute data and conducting intraday interval-level analysis, it reveals that no single model consistently dominates across horizons or regions; standard DL models generally outperform SOTA models in accuracy, while SOTA models offer greater robustness to horizon extension. The results uncover pronounced diurnal error patterns: absolute errors spike during evening volatility, relative errors inflate during midday negative-price regimes, and directional accuracy degrades during frequent trend changes, underscoring the need for richer feature representations and horizon-robust modelling. The work suggests future EPF development should integrate multivariate inputs (weather, generation mix, interconnector flows) and explore asymmetric loss formulations to better capture heavy-tailed, regime-shifting price dynamics, thereby improving both short-term responsiveness and long-term forecasting stability.

Abstract

Accurate electricity price forecasting (EPF) is essential for operational planning, trading, and flexible asset scheduling in liberalised power systems, yet remains challenging due to volatility, heavy-tailed spikes, and frequent regime shifts. While deep learning (DL) has been increasingly adopted in EPF to capture complex and nonlinear price dynamics, several important gaps persist: (i) limited attention to multi-day horizons beyond day-ahead forecasting, (ii) insufficient exploration of state-of-the-art (SOTA) time series DL models, and (iii) a predominant reliance on aggregated horizon-level evaluation that obscures time-of-day forecasting variation. To address these gaps, we propose a novel EPF framework that extends the forecast horizon to multi-day-ahead by systematically building forecasting models that leverage benchmarked SOTA time series DL models. We conduct a comprehensive evaluation to analyse time-of-day forecasting performance by integrating model assessment at intraday interval levels across all five regions in the Australian National Electricity Market (NEM). The results show that no single model consistently dominates across regions, metrics, and horizons. Overall, standard DL models deliver superior performance in most regions, while SOTA time series DL models demonstrate greater robustness to forecast horizon extension. Intraday interval-level evaluation reveals pronounced diurnal error patterns, indicating that absolute errors peak during the evening ramp, relative errors inflate during midday negative-price regimes, and directional accuracy degrades during periods of frequent trend changes. These findings suggest that future research on DL-based EPF can benefit from enriched feature representations and modelling strategies that enhance longer-term forecasting robustness while maintaining sensitivity to intraday volatility and structural price dynamics.

Figures (17)

• Figure 1: An overview of the proposed EPF framework in this study, encompassing the stages of NEM data acquisition through data preparation, model selection, hyperparameter optimisation, training to forecasting, and performance evaluation.
• Figure 1: MAE, RMSE, sMAPE, and MDA of each interval of the day in the NSW region for all the evaluated models.
• Figure 2: Train, validation, and test splits of the electricity price data for the QLD region.
• Figure 2: MAE, RMSE, sMAPE, and MDA of each interval of the day in the VIC region for all the evaluated models.
• Figure 3: Average Performance of Models Across All Metrics.