Enhanced rainfall forecasting for Mumbai using data-driven ConvLSTM2D models at fine spatial and temporal scales

Forecasting rainfall in tropical areas is challenging due to complex atmospheric behavior, elevated humidity levels, and the common presence of convective rain events. In the Indian context, the difficulty is further exacerbated because of the monsoon intra-seasonal oscillations, which introduce significant variability in rainfall patterns over short periods. Earlier investigations into rainfall prediction leveraged numerical weather prediction methods, along with statistical and deep learning approaches. This study introduces a nuanced approach by deploying a deep learning spatial model aimed at enhancing rainfall prediction accuracy on a finer scale. In this study, we hypothesize that integrating physical understanding improves the precipitation prediction skill of deep learning models with high precision for finer spatial scales, such as cities. To test this hypothesis, we introduce a physics-informed ConvLSTM2D (Convolutional Long Short-Term Memory 2D) model to predict precipitation 6 h and 12 h ahead for Mumbai, India. We utilize ERA-5 reanalysis data with hourly time steps spanning from 2011 to 2022 to select predictor variables, including temperature, potential vorticity, and humidity, across various geopotential levels. The ConvLSTM2D model was trained on the target variable precipitation for 4 different grids representing different spatial grid locations of Mumbai. The Nash–Sutcliffe Efficiency (NSE), utilized to evaluate the precision of 6 and 12 h ahead precipitation forecasts, yielded ranges of 0.61–0.68 for 6-h predictions and 0.58–0.66 for 12-h predictions during the training phase. In the testing phase, the NSE values range from 0.42 to 0.51 for 6-h forecasts and from 0.47 to 0.58 for 12-h forecasts, respectively. These values highlight the model’s high accuracy and its capacity to capture variations. Thus, the use of the ConvLSTM2D model for rainfall prediction, utilizing physics-informed data from specific grids with limited spatial information, reflects current advancements in meteorological research that emphasize both efficiency and localized precision.