EMD-KNN Model for Annual Average Rainfall Forecasting

The prediction of rainfall is the premise of responding to water resources management and flood defense. The hydrological system is complex, and rainfall time series as a member of it has characteristics of being nonlinear and nonstationary, so accurate rainfall prediction still a much difficult job at present. The paper proposes a conjunction model named EMD-KNN for forecasting annual average rainfall. The model is improved by combining two methods, empirical mode decomposition (EMD) and the K-nearest neighbor bootstrap regressive model (K-NN). It is applied to case studies of forecasting annual average rainfall for Nanjing city and the Dahuofang reservoir basin. Where Nanjing city is water-rich area in East China, Dahuofang reservoir basin is water-deficient area in Northeast China. Three performance evaluation measures results revealed that the EMD-KNN model reduces the prediction mean absolute error (MAE), mean relative error (MRA), and root-mean square error (RMSE) with respect to the single K-NN model by almost 50% each, so the suggested model can effectively improve the forecast accuracy of single K-NN in forecasting the annual average rainfall.