Akademik Veri Yönetim Sistemi
8th International Congress on Human-Computer Interaction, Optimization and Robotic Applications, ICHORA 2026, Ankara, Türkiye, 21 - 23 Mayıs 2026, (Tam Metin Bildiri)
The production performance of photovoltaic systems depends not only on historical data but also directly on complex environmental dynamics such as temperature, solar radiation, wind, and cloud cover. In this study, a unique dataset was created by integrating production records from the Yulara Solar system in Australia with CAMS-based particulate matter measurements. Four different deep learning architectures were established: LSTM, GRU, Hybrid_CNN_LSTM and Hybrid_CNN_Transformer. All models were evaluated under the same input size and similar training strategies. WMAPE, MAE, RMSE, $\mathbf{R}^{\mathbf{2}}$, and MAPE metrics were used in the performance comparisons of the models. The results were evaluated according to the WMAPE metric. The most successful model in this study was the Hybrid_CNN_Transformer model. The highest $\mathbf{R}^{\mathbf{2}}$ value was observed in the GRU model. This finding shows that model success can vary depending on the performance metric used. XAI methods were used in the evaluation of the results. It has been observed that intraday cyclical time variables, radiation-based inputs, and air pollution data have an impact on model decisions. The combined use of hybrid deep learning architectures and explainable artificial intelligence approaches demonstrates a supportive effect in terms of both accuracy and interpretability in solar energy forecasting.