Akademik Veri Yönetim Sistemi
8th International Conference on Mathematical Advances and Applications (ICOMAA-2025), İstanbul, Türkiye, 7 - 09 Mayıs 2025, cilt.1, sa.1, ss.1-9, (Tam Metin Bildiri)
In this study, a hybrid ensemble machine learning model integrating statistical learning and time series modeling is proposed for hourly forecasting of power generation from photovoltaic (PV) systems. The model architecture combines the high-accuracy regression capabilities of the Extreme Gradient Boosting (XGBoost) algorithm with the temporal dependency learning capacity of Long Short-Term Memory (LSTM) networks. The final prediction is obtained by dynamically aggregating the outputs of both models using a weighted averaging technique. The modeling process utilizes hourly meteorological variables collected over one year, including solar irradiance, ambient temperature, relative humidity, and cloud cover. In addition, domain-specific feature engineering based on solar physics is applied to enhance the explanatory power of the input variables. Hyperparameter optimization for both XGBoost and LSTM models is conducted using a parametric grid search algorithm and cross-validation techniques. The performance of the model is evaluated using statistical metrics such as Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and the Coefficient of Determination (R²). The hybrid model demonstrated a high prediction accuracy with an RMSE of 0.0057 and an R² of 0.9994. For comparative analysis, conventional methods, including Support Vector Machines (SVM), k-Nearest Neighbors (k-NN), Random Forest (RF), and ARIMA, were also implemented, and the proposed hybrid approach outperformed all these traditional techniques. The results obtained indicate that statistical learning-based hybrid models possess significant potential for smart decision support systems in distributed energy applications and contribute to the more efficient and predictable management of renewable energy resources.