Work

Multi-Horizon Inverter-Level AC Power Forecasting Using Inverted Attention Transformers (iTransformer): A Real PV Case Study in Konya, Türkiye

Year: 2026

Type: article

Abstract: Accurate multi-horizon forecasting of inverter AC output power is essential for the reliable operation and efficient management of grid-connected photovoltaic (PV) power plants. However, inverter-level forecasting remains challenging due to the highly nonlinear and rapidly varying nature of solar irradiance, particularly in high-resolution operational data. In this study, an iTransformer-based forecasting framework is adopted for multi-horizon prediction of inverter AC output power using one year of real-world 15-minute resolution data collected from a utility-scale PV power plant located in Konya, Türkiye. A comprehensive experimental design is established under four different feature configurations, ranging from univariate input to multivariate settings incorporating physically meaningful exogenous variables, including irradiance ramp features. The iTransformer model is systematically compared with Informer and Long Short-Term Memory (LSTM) networks under identical training protocols and forecasting horizons. Experimental results demonstrate that iTransformer consistently outperforms both baseline models across all evaluated settings. In particular, the applied approach achieves relative MSE reductions ranging from 27.9% to 49.8% compared to Informer, with similar improvements observed over LSTM, especially under longer forecasting horizons. The analysis further reveals that incorporating irradiance ramp information significantly enhances prediction accuracy, confirming the dominant role of short-term irradiance fluctuations in inverter power variability. These findings validate the effectiveness of feature-wise inverted attention for multivariate-to-univariate energy forecasting and highlight its practical value for grid scheduling, reserve planning, and smart PV monitoring systems.

Source: IEEE Access

DOI: 10.1109/access.2026.3702901

Authors: Bülent Haznedar, Hilal Haznedar, Ali Eren Kök

Institutions: Gaziantep University, Hasan Kalyoncu University

Language: English

Topic: Solar Radiation and Photovoltaics

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

SDGs: Affordable and clean energy, Industry, innovation and infrastructure, Climate action

Open Access status: gold

Estimated APC paid: $2,075