Tehnički vjesnik, Vol. 33 No. 4, 2026.
Izvorni znanstveni članak
https://doi.org/10.17559/TV-20250917002999
TAN: A Temporal Attention Network for Photovoltaic Power Forecasting
Chao Huang
; Beijing Fiberlink Communications Co., LTD., Beijing 100010, China
Hui Zhang
; Beijing Fiberlink Communications Co., LTD., Beijing 100010, China
*
Tian Luan
; Beijing Fiberlink Communications Co., LTD., Beijing 100010, China
* Dopisni autor.
Sažetak
The increasing penetration of distributed photovoltaic (PV) systems poses significant challenges to power grid stability, making accurate medium and long-term PV power forecasting essential. This paper proposes a Temporal Attention Network (TAN) for multi-step PV power forecasting. TAN adopts a pure temporal self-attention architecture with a learnable temporal embedding module, enabling effective modeling of periodic and non-stationary temporal dependencies without relying on recurrent or convolutional structures. The proposed model is evaluated on real-world PV data under 24-hour and 48-hour forecasting horizons. Performance is assessed using multiple metrics, including RMSE, MAE, MAPE, R2, and confidence intervals derived from rolling forecasting evaluation. Experimental results show that TAN achieves consistently lower forecasting errors and higher explanatory power than representative recurrent, convolutional, hybrid, and Transformer-based baselines in both forecasting settings. Moreover, TAN maintains competitive performance with moderate model complexity. These results indicate that TAN provides an effective and scalable attention-based solution for medium- and long-term PV power forecasting in renewable energy - integrated power systems.
Ključne riječi
attention mechanism; deep learning; photovoltaic power forecasting; temporal attention network
Hrčak ID:
348699
URI
Datum izdavanja:
30.6.2026.
Posjeta: 0 *