The real-time availability of data from coastal meteo-marine stations is crucial for various stakeholders, including port authorities, government agencies, researchers, and the general public. While observation data is fundamental, short-term forecasts can significantly enhance planning and decision-making processes. This study explores the application of Machine Learning (ML) techniques to predict hourly values of air temperature, wind speed, atmospheric pressure, and humidity for the next 24 hours. We evaluate three ML models: Long Short-Term Memory Network (LSTM), Random Forest (RF), and Multivariate Linear Regression (LR). The models were trained using Python libraries and Optuna for hyperparameter tuning on datasets of varying lengths from stations in the Malta-Sicily channel. Additionally, we investigated transfer learning with the ERA5 dataset, which provides hourly values over an 83-year period, to address the challenge of limited data availability. The results show that models trained on longer datasets generally achieve better performance. Furthermore, the models demonstrated considerable generalizability, particularly across nearby stations, allowing models trained at one station to be effectively used for predictions at other proximate stations. To support further research and practical application, we have made our models and tools publicly available.