Rainfall Prediction in Kathmandu City Using Machine Learning and Deep Learning Techniques

Abstract

Precipitation forecasting is a relatively important issue of preparedness to disasters and water resources management and adaptation to climate changes, especially in the geographical location such as Kathmandu Valley in Nepal where precipitation patterns have become erratic and acute with the effect of climate change. The work concerns the depth of rainfall forecasting through the use of machine learning (ML) and deep learning (DL) methods on the meteorological dataset that is 10 years long (2015-2025), with the characteristics that include temperature, humidity, atmospheric pressure, wind direction, and cloud cover. We compare the six predictive models (LightGBM, Support Vector Regression (SVR), K-Nearest Neighbors (KNN), Random Forest, Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) to find what patterns exist in the historical weather data and predict rainfall instances. We use analysis to show that the LSTM model reaches the best accuracy, which is RMSE of 3.73 and MAE of 2.48, even against the traditional ML models, such as Random Forest (RMSE: 3.89, MAE: 2.63) and SVR (RMSE: 4.28, MAE: 2.57). Its strength is seen in its capacity to reproduce temporal dependencies and nonlinear trends in the seasonal rainfall data, which has made LSTM highly effective. Also, our method minimizes the time of computation of the model, by exploiting optimized hyperparameters and preprocessing mannerisms specific to the zero-inflated nature of precipitation data. The results demonstrate the feasibility of deep learning models in enhancing the accuracy of rainfall predictions which can be used by policymakers and urban planners to help them prepare well ahead of time when it comes to a disaster-prone area.