Centralized and Decentralized Approach to Monsoon Precipitation Forecasting in Pakistan

Rainfall, is one of the most important meteorological factors that affects many parts of our everyday lives including crop productivity, water quality, livestock availability, hydroelectric power generation to name a few. Rainfall prediction can significantly contribute to boosting the economy by enabling better planning, risk management, and resource allocation in various industrial sectors. In this study, forty years of monsoon precipitation data is gathered for 39 stations across five zones in Pakistan. We propose a multi-step Long Short-Term Memory (LSTM)-based prediction model capable of forecasting Monsoon yearly data. Three LSTM models stack, bidirectional and convolutional are applied on the dataset and the performance of these models are analysed using a centralized and a decentralized approach. It is observed that the RMSE score of the LSTM models across the centralized strategy was found better than the decentralized approach, whereby 100% of the models in the centralized had a lower RMSE as compared to the decentralized one. Moreover, in the centralized approach 78.7% of the models across the different zones exhibited R2 > 0.9 values indicating a general fit to the model.