Akademik Veri Yönetim Sistemi
Journal of Energy Storage, cilt.138, 2025 (SCI-Expanded, Scopus)
Solar ponds are systems consisting of layers of increasing density with depth, where solar energy is stored thermally in a very dense saltwater solution. Integrated systems with solar collectors have been proposed to increase the thermal performance of the solar pond. This paper presents a high accuracy ensemble learning approach to predict the temperature distribution in the Heat Storage Zone (HSZ) of an integrated solar pond-collector system. Using a vertically layered thermal monitoring structure equipped with thirteen thermocouples, the model predicts five key target temperatures (HSZ; T9-T13) based on eight input variables. These inputs are Non-Convective Zone (NCZ) temperatures (T1-T4), Upper Convective Zone (UCZ) temperature (T5), heat exchanger inlet/outlet temperatures (T6-T7) and ambient temperature (T8). Support Vector Regression (SVR), Nearest Neighbor (KNN), Decision Tree (DT), Random Forest (RF), Gradient Boost Machine (GBM), XGBoost (XGB), Artificial Neural Network (ANN), Ridge, A Super Learner ensemble model consisting of ten different regressors, Gaussian Process Regression (GPR) and LASSO, was trained by 10-fold cross-validation and evaluated on seven statistical metrics (RMSE, MAE, R2, MAPE, SMAPE, EVS and Maximum Error). The results revealed that the Super Learner model outperformed all individual models in every layer. Especially for the deepest layer, HSZ-T13, the model achieved high performance with RMSE of 0.4584, MAE of 0.2936, R2 of 0.9985 and MAPE of only 0.76 %. These findings reveal that the proposed ensemble model is able to model nonlinear heat dissipation patterns and vertical thermal gradients with high accuracy. Additional time analyses show that SL offers low extraction delay relative to its training cost and is suitable for real-time monitoring/energy management scenarios. The proposed method improves operational efficiency by providing reliable real-time temperature prediction and supports energy optimization in renewable thermal energy storage applications. The proposed approach has provided lower error and higher stability across layers compared to single models. Specifically, in the deepest layer (HSZ-T13), RMSE = 0.4584 °C, MAE = 0.2936 °C, R2 = 0.9985, MAPE = 0.76 % were obtained (10-fold CV). Fold-based comparisons and distribution similarity (KDE) analyses demonstrate the consistent superiority of the Super Learner.