Price-aware artificial intelligence for market-aligned photovoltaic power nowcasting under quarter-hourly electricity settlement

The increasing penetration of distributed photovoltaic generation and the move toward quarter-hourly imbalance settlement, exemplified by the Italian Testo Integrato del Dispacciamento Elettrico reform, make ultra-short-term power forecasting a critical operational and economic task. Although machine-learning forecasters have improved statistical accuracy, they are commonly trained and evaluated with symmetric error indicators that do not reflect the asymmetric monetary effects of over- and under-forecasting. This paper addresses this gap by proposing a decision-oriented framework for fifteen-minute-ahead photovoltaic power nowcasting that embeds market asymmetry in both evaluation and learning. The framework introduces two methodological contributions: an Expected Monetary Impact indicator, which translates forecast deviations into monetary losses at quarter-hour settlement resolution, and a differentiable price-aware asymmetric loss function, which penalizes positive and negative forecast errors according to their market costs. The Italian reform is used as a realistic case study, while the formulation can be transferred to other markets with short settlement intervals and asymmetric imbalance prices. Long Short-Term Memory, Gated Recurrent Unit, and a Locally Recurrent Neural Network trained through the Binet formalism are evaluated as alternative forecasters against persistence under a rolling out-of-sample protocol using real measurements from a photovoltaic facility in Milan, with 2023–2024 used for training and 2025 for testing. The learning-based models outperform persistence: Long Short-Term Memory and Gated Recurrent Unit reduce the normalized monetary impact by 18.22% and 19.39%, respectively, while the locally recurrent model achieves a 54.72% reduction. These results show that market-aligned forecasting objectives provide greater operational value than conventional symmetric accuracy indicators.