2026-10-19T00:02:51Zhttps://ebyu-rdlopen.rawdatalibrary.net/oai/request

oai:ebyu-rdlopen.rawdatalibrary.net:ebyu/101662026-10-18T20:36:44Zopenairedrivercom_ebyu_10000col_ebyu_10001

Data-driven forecasting of photovoltaic performance with explainable artificial intelligence: Experimental validation using environmental and thermal predictorsKarakoyun, YakupKatipoğlu, Okan MertDogan, AhmetCanbaz, AliMachine LearningRegression AnalysisSHAPSolar Energy ForecastingSustainable EnergyAtmospheric temperatureConvolutionConvolutional neural networksForecastingPhotovoltaicsRandom forestsSolar energySupport vector regressionSustainable developmentWindEnergy forecastingExperimental validationsMachine-learningPanel temperaturePerformanceRandomized searchThermalLearning systemsPV systems meet energy demand, but environmental and operational factors affect their performance. This study used a three-scenario ablation framework to predict PV power from baseline meteorological and operational variables to under-panel temperature and finally to the full thermal configuration, including water temperature (a newly studied variable). Scenario 1 included time, air temperature, panel angle, relative humidity, and average wind speed; Scenario 2 added under-panel temperature; and Scenario 3 added water temperature. Direct thermal predictors were evaluated separately from meteorological inputs using this design. Four sophisticated ML algorithms, Random Forest (RF), CatBoost, Support Vector Machine (SVM), and one-dimensional Convolutional Neural Network (1D CNN), were refined by Randomized Search (RS) and evaluated using metrics. According to the first group of power prediction results under the full Scenario 3 configuration, the CatBoost-RS model demonstrated superior performance with RMSE = 10.61, MAE = 8.41, AIC = 222.22, KGE = 0.957, and R2 = 0.99. For the second group, the SVM-RS model outperformed others, achieving RMSE = 8.02, MAE = 5.50, AIC = 200.36, KGE = 0.994, and R2 = 0.992. Comparative analysis showed that CatBoost-RS was the most robust predictor, while RF exhibited proficiency in managing nonlinearity in this experimental setting. One-way ANOVA, paired t -test, Wilcoxon signed-rank test, bootstrap confidence intervals, and effect size (Cohen’s d) analyses across both groups show that optimised ensemble and kernel-based models outperform the convolutional approach in predictive performance. Experimental validation and high predictive accuracy improve PV performance prediction in the proposed method. © 2026 Elsevier Ltd.Elsevier Ltd2026Articlehttps://rdlopen.rawdatalibrary.net/r/ebyu/10166https://doi.org/10.1016/j.enconman.2026.1217340196-8904enEnergy Conversion and Managementinfo:eu-repo/semantics/openAccessinfo:eu-repo/semantics/articlehttps://ebyu-rdlopen.rawdatalibrary.net/tr/items/ebyu/10166