Abstract:
Forecasting precipitation is highly challenging for scientific modellers due to the complexity and uncertainty of atmospheric data and weather prediction models. To investigate the hydrological alternations such as rising sea levels, increasing floods and evaporation, and changes in snowpack caused by climate change, it is essential to accurately predict precipitation, a function of several interrelated climatic variables. This study presents a unique approach to predicting precipitation with minimum uncertainty by performing a comparative assessment of long-short-term memory (LSTM) approaches. The LSTM prediction models were run using quarterly, semi-annual, annual, and biannual precipitation data and other data such as temperature, vapour pressure, cloud cover, rainy days, and potential evaporation. Bivariate models using potential evaporation and temperature produced equivalent results to the multivariate model as the mean absolute error (MAE) was found to be 23.89% and 26.35%, respectively, compared to the univariate model (MAE 76.29%).