Long-range forecasting of daily rainfall using machine learning techniques

Around the world, awareness of the value of rainfall forecasting and its social and economic advantages is rapidly growing. Nation like India, where agriculture is one of the main economic drivers, accurate rainfall forecasting is essential for managing water resources and reducing hydrological extremes. To predict daily rainfall for a year in the drought-prone Kangsabati river basin, long short-term memory (LSTM) and random forest (RF) techniques were utilised using the data of 52 years (1969 to 2020). Finding the right associated variable and a substantial lag in the time series that allows for future value prediction is essential when developing time series forecasting models. The partial auto correlation function and Pearson correlation technique were applied in this context. Furthermore, comparisons were made between the suggested models and the well-known statistical model, the seasonal auto-regressive integrative moving average (SARIMA). The study shows that for this region, the proposed LSTM model and SARIMA model had higher accuracy than the RF model. Additionally, this research suggests that machine learning algorithms may be used to analyse daily rainfall for a particular catchment or station and, to a certain extent, forecast extreme hydrological events.