Exergy analysis provides a thermodynamically rigorous measure of energy quality by explicitly accounting for irreversibilities, entropy generation, and the true useful work potential of energy conversion systems. Although widely used for performance assessment of renewable energy technologies, exergy evaluation is typically conducted through system-specific thermodynamic modeling or detailed numerical simulations. Such approaches, while accurate, are computationally intensive and often tailored to individual configurations, limiting scalability and rapid optimization across diverse renewable platforms. In contrast, machine learning methods have been increasingly adopted for renewable energy forecasting and performance prediction; however, most existing studies focus primarily on energy efficiency and rely on purely data-driven architectures that lack thermodynamic consistency and physical interpretability. This work presents a physics-guided machine learning framework for predicting exergy efficiency across solar thermal, wind, and biomass-based systems. A feedforward neural network architecture comprising three hidden layers with ReLU activation functions is trained on a curated dataset assembled from published case studies and publicly available performance data. Input variables include key operational and thermodynamic parameters relevant to each system type, enabling cross-technology applicability within a unified modeling structure. To ensure physical consistency, constraints derived from the first and second laws of thermodynamics are incorporated directly into the loss function as penalty terms. These regularization components enforce non-negative entropy generation and maintain consistency between exergy destruction, input exergy, and exergy efficiency definitions during training. Model performance is evaluated using five-fold cross-validation and standard regression metrics, including root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). The proposed framework achieves an average R² of 0.93 and an MAE of 0.037 in exergy efficiency prediction. Compared with a conventional neural network baseline, the physics-guided model demonstrates improved predictive stability and a lower incidence of physically inconsistent outputs, particularly under off-design operating conditions. Unlike prior system-specific exergy prediction approaches, this study proposes a unified physics-guided framework applicable across multiple renewable technologies. The findings indicate that embedding thermodynamic constraints within the learning process enhances robustness and interpretability while preserving computational efficiency, providing a scalable alternative to purely simulation-based exergy assessment for renewable energy optimization and sustainability analysis.