Rainfall Forecasting with Intermittent Approach using Exponential Smoothing Neural Network Ensemble

Rainfall forecasting is critical for agriculture, air resource management, urban planning, and emergency preparedness. Accurate forecasts allow decision-makers to adapt to changing weather conditions and reduce hazards. Traditional forecasting methods cannot represent meteorological data’s complex temporal dynamics and nonlinearity, especially intermittent rainfall patterns. This research explores the combination of Exponential Smoothing and Neural Networks in an irregular context by recognizing the unique characteristics of rainfall data. The data is divided into three areas. Each area will be forecasted using the model and evaluated by the metric value. The names of those three areas are Keputih, Kedung Cowek, and Gubeng. The Exponential Smoothing and Neural Network approaches are expected to contribute to more accurate and reliable rainfall forecasts, providing benefits for various applications and sectors that rely on precise weather predictions for decision-making and planning. We use the intermittent approach in input variables. Instead of using lag variables, we create a rolling mean as input variables. The evaluation procedure is using RMSSE value as a prior metric. That value is robust in the sparse data. Besides, other metrics such as RMSE, MSE, and MAPE are also provided. The results show that the ensemble model can form patterns better than individual exponential smoothing models or neural networks. The RMSSE values for all areas are 0.8867, 0.8626, and 0.7569.