Vessel performance monitoring and fuel consumption optimization have become critical in the maritime industry due to evolving environmental regulations and the need for operational efficiency. Accurate prediction of the propeller shaft power, provided by the main engine, is fundamental to these optimization routines. In recent years, Physics-Informed Neural Network (PINNs) and knowledge guided models, utilizing sea-trial baselines, have demonstrated high efficacy in modeling vessel parameters [1, 2]. However, their performance is inherently tied to the quality and distribution of the training data. Specifically, the imbalance between ballast and laden loading conditions, often present in real world operational datasets, can significantly bias model predictions and limit its robustness. This work investigates the impact of draft distribution within a dataset by analyzing how varying the percentage of ballast versus laden data points affects the model's accuracy in predicting the main engine power and its ability to capture the underlying physics, namely the cubic relationship of the propeller law. To further evaluate the operational impact of these predictions, a novel utility function is introduced that translates power prediction errors into potential fuel consumption deviations and associated costs. Preliminary results indicate that although physics-guided models partially mitigate the effects of data imbalance, the specific ratio of loading conditions remains a primary driver of generalization error. This study provides a practical framework for assessing dataset utility and optimizing data collection strategies for maritime decision making and planning.