A Combined model based on data decomposition and multi-model weighted optimization for precipitable water vapor forecasting

Abstract

Water shortage is a major problem facing the world. Artificial precipitation enhancement is an effective way to improve precipitation conversion rate, but the selection of artificial precipitation enhancement operation timing is the main difficulty, and the precipitable water vapor(PWV) is a major index. The variation of PWV is nonlinear and unstable due to complex factors, especially in the Qilian mountains in the northeastern part of the Qinghai-Tibet Plateau, so it is difficult to predict it accurately. Therefore, based on the analysis of the observed data of microwave radiometer in Qilian Mountains, a new combined model is constructed which considers both data decomposition and prediction of several single models in this research. In the data preprocessing stage, the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) technique is used to decompose and de-noise the PWV sequence. In the prediction stage, four neural network with unique characteristics, back propagation neural network (BPNN), long short term memroy (LSTM), bidirectional gated recurrent unit (BiGRU) and temporal convolutional network (TCN), are selected to predict the decomposed data respectively. A variant of gray Wolf optimization algorithm (IGWO) is used to determine the optimal weight of the model, and finally the comprehensive predicted value is obtained by weighting calculation. Through the analysis of experimental results, in the longest 15-step prediction, compared with CEEMDAN-BP, CEEMDAN-LSTM, CEEMDAN-BiGRU, CEEMDAN-TCN, the prediction accuracy can be improved by 54.17%, 35.05%, 22.38%, 23.86%, respectively. Other step size prediction also achieves the highest prediction accuracy.