In time series analysis, data aggregation is an essential preprocessing step that consolidates data points over specified time intervals, simplifying the data structure and reducing noise. This process is vital for enhancing the manageability of large, complex datasets, commonly encountered in various domains such as energy consumption forecasting. However, the choice of the most appropriate temporal aggregation (TA) frequency can significantly impact the performance and the reliability of predictive models. Different TA frequencies may perform better in different scenarios, and there is a lack of evidence on what indicators may determine the superiority of each one. In this study, we design and execute an empirical experiment framework to first explore the performance of various machine learning (ML) models using different TA frequencies on 04 multivariate time series datasets of ship fuel consumption. We further investigate the reliability of those models by studying the marginal contribution of the time series features in each aggregation period. Using the Random Forest Regressor, XGBoost Regressor, LightGBM, and Extra Trees predictive models along with the Shapley Additive Explanations to analyze feature contributions, we found that time series aggregation consistently produces accurate results regardless of the TA period. However, the models may exhibit significant biases, highlighting the need to examine their inner workings to select the appropriate TA period.