This study aims to evaluate whether ComBat data harmonization can improve the reproducibility and scalability of autoencoder-based normative modeling of resting-state fMRI connectivity across multiple sites and diagnostic groups. Resting-state fMRI data from three public datasets — UCLA (schizophrenia, bipolar disorder, ADHD), COBRE (schizophrenia), and CPDS (chronic pain) were preprocessed using identical pipelines, parcellated with the Yeo 17-network atlas and functional connectivity was estimated using several methods. ComBat harmonization was applied considering healthy control data only, and multilayer autoencoder models were trained to learn normative connectivity patterns. Model accuracy and stability were assessed through reconstruction error (MAE) and a Total Score metric summarizing reproducibility across preprocessing configurations. Harmonization reduces site-related variability, slightly improves reconstruction accuracy, and increases the model stability compared to single-site training. Across diagnostic groups, harmonization led to more stable and interpretable connectivity deviation patterns, aligning with known large-scale network dysfunctions reported in psychiatric and pain research. Overall, ComBat effectively reduced site-related variability and improved the cross-cohort consistency of normative modeling results, while preserving meaningful biological differences. It thus represents a practical and scalable strategy for transdiagnostic fMRI connectivity analysis. However, residual site and cohort variability persist, underscoring the need for validation on larger and more diverse datasets to confirm its generalizability and fully address multi-site heterogeneity.