Reliable fault diagnosis in photovoltaic systems is compromised when measurement data are corrupted by noise. This study assesses the robustness of Support Vector Machine-based hybrid classifiers SVM+KNN, SVM+LR, SVM+MLP, SVM+DT, and SVM+RF subjected to controlled Gaussian noise injection. A dataset of 13,767 records collected at the Ngaoundéré weather station was used, partitioned 80%/20% for training and testing. The hybridization strategy relies on a parallel probabilistic fusion scheme in which prediction probabilities from each base classifier are averaged. Model performance was evaluated along three complementary axes: F1-score, empirical error rate, and temporal stability of the classification rate. Results show that the SVM+RF hybrid achieves the best overall accuracy (91.7%) and AUC (0.991), with the greatest resilience to noise, while SVM+KNN exhibits the weakest robustness. Importantly, probabilistic fusion does not consistently outperform the strongest individual model; it mainly moderates instability when the base classifiers offer genuine complementarity. One-way ANOVA confirms that the performance differences between individual and hybrid configurations are statistically significant. Temporal analysis further reveals that fused models maintain a more regular classification rate across samples, a key advantage in unstable PV operating environments. This work contributes to the development of robust and reliable hybrid systems for real-world diagnostics.