River-dust induced airborne particulate matter forecasting using a hybrid model of improved complete ensemble empirical mode decomposition with adaptive noise and radial basis function neural network

Abstract

In the western plains of Taiwan, several rivers originate from the central Taiwanese mountains. Turbulent flow transports large quantities of river sediment that accumulates in the riverbed downstream. River-dust events are extreme dust events that are occasionally caused by drought along riverbeds. The Choushui and Kaoping Rivers demonstrate the highest potential for river-dust events in central and southern Taiwan, respectively. During monsoons, the accumulated river sediment in the estuary can easily become airborne. Consequently, vulnerable populations living along riverbanks are exposed to high PM10 concentrations, which may exert adverse health effects. To reduce the significant health risks associated with river-dust events, a hybrid pmodel composed of improved complete ensemble empirical mode decomposition with adaptive noise and radial basis function neural network (ICEEMDAN-RBFNN) is proposed. The hybrid model is compared with the RBFNN and Multilayer Perceptron (MLP) models for forecasting 3-h PM10 concentrations. In the hybrid prediction model, temperatures, relative humidity, wind speeds, wind direction indices, and previous hourly PM10 concentrations are accounted for. The results indicate that the hybrid model is more accurate than the RBFNN and MLP models in forecasting high and low PM10 concentrations in 3 h advance. The hybrid model achieves a reduction of up to 50% in both the MAPE (Mean Absolute Percentage Error) and RMSE (Root Mean Squared Error) compared to the RBFNN and MLP prediction models. Therefore, the hybrid ICEEMDAN-RBFNN model can forecast PM10 efficiently and can serve as an early-warning system for river-dust events. Finally, risk assessments are conducted to exemplify the application of the proposed hybrid model.