Time series forecasting is crucial in various fields, with significant socio-economic implications, as accurate predictions can aid in better resource management, disaster preparedness, and economic planning. However, selecting an appropriate forecasting model remains a labor-intensive task demanding expertise. This research introduces a novel meta-learning approach to automate and enhance the model selection process. We curate extensive time series datasets specific to Ireland, spanning diverse temporal patterns and environmental attributes, including climate data, water level measurements, and landscape characteristics. The initial part of the research focuses on developing a systematic architecture using Extract, Transform, and Load (ETL) technology to integrate heterogeneous data from various sources while ensuring data quality and consistency. Then, this research concentrates on accurately predicting river water levels. Various Machine Learning (ML) models are employed, relying on previously observed river water level data. The research evaluates the predictive performance of these ML models across all hydrometric stations in Ireland and demonstrates the importance of careful model selection based on geographic and hydrological features. The results demonstrated that a universal ‘one-model-fits-all’ approach is not suitable for hydrological time series data. Subsequently, this research explores the core contribution of applying meta-learning to context-aware model selection for river water-level prediction. The study demonstrates that meta-learning enhances the accuracy and reliability of hydrologic time series forecasting, addressing the complexities of this task and providing valuable insights into applying ML in this domain. The efficacy of our meta-learning approach is evaluated across various real-world time series datasets, consistently demonstrating its superiority over traditional model selection techniques. Importantly, our approach streamlines and expedites time series forecasting, making it more accessible to researchers. In conclusion, this thesis significantly contributes to ML-based environmental time-series data prediction using a model-selection meta-learner approach and enhanced data integration techniques. The results show that our research aligns with the growing trend of automated machine learning and has the potential to revolutionise time series forecasting in diverse applications.