Hourly Precipitation Prediction: Integrating Long Short-Term Memory (LSTM) Neural Networks with Granger Causality

Precipitation events are one of the most crucial processes governing the water cycle and therefore acts as a major input in majority of water resource studies. Furthermore, the modern era is witnessing an unprecedented increase in the frequency of extreme events highlighting the importance of understanding and predicting the precipitation events. Current precipitation prediction methods utilise complex physics-based models that require large number of input parameters as well as powerful computational facilities, making precipitation prediction a complex task. This scenario has not improved much over the years as advancements are often limited to either improving the model’s physics or input data quality. Hourly precipitation prediction is even more challenging due to increasing complexity and non-linearity with decreasing scale and therefore studies on understanding hourly precipitation is limited. Recent trends have shown a shift towards utilizing deep learning models in weather prediction owing to the ability of neural networks to capture complex patterns leading to high accuracy predictions. The current research introduces a Long Short Term Memory (LSTM) neural network adept at forecasting fine-scale hourly precipitation patterns up to two hours ahead, a critical development for real-time rain predictions. The Bi-LSTM's architecture, with its forward and backward processing capabilities, is particularly suited to capture the dynamic temporal relationships among the limited meteorological variables helping in effective precipitation prediction. Granger Causality analysis is done to capture relevant information for improving model performance. The model's performance is evaluated on its ability to accurately forecast weather conditions by learning from the historical inter-variable influences that were clearly detailed in the causality diagram. The findings from this study and the interlinks observed is expected to enhance our understanding of variable impact and improve the predictive power of precipitation models for future weather forecasting