Precipitation analysis and forecasting using singular spectrum analysis with artificial neural networks

Abstract

The temporal variability monthly precipitation time series for Korea is decomposed using singular spectrum analysis (SSA) to detect hidden periodicity information in the data, and to compare the forecasting performance of combining linear recurrent formulas (LRFs) and artificial neural networks (ANNs). The SSA technique is used on monthly precipitation data to decompose and reconstruct the components, including a special inerratic feature for reconstruction and successful forecasting using LRF and ANN analysis. These components obtained using SSA indicate the behavior of the monthly precipitation data as a trend, or as periodic and/or quasi-periodic oscillations. The LRF and ANN methods were applied to several leading components to forecast the monthly precipitation. Results show that reconstruction and forecasting using the SSA-ANN model is more accurate than using the SSA-LRF model, especially for peak value forecasting. This validates the use of the SSA-ANN combined model for effective reconstruction and forecasting of monthly precipitation.