Application research of combined model based on VMD and MOHHO in precipitable water vapor Prediction

Abstract

The change characteristics of precipitable water vapor (PWV) are closely related to the timing of artificial rainfall enhancement operations, and accurate prediction of PWV changes is of great practical significance. The instability and nonlinearity of PWV are the difficulties in its accurate prediction. However, most of the current PWV prediction models are individual models and simple hybrid models, without considering the limitations of a individual prediction model, efficient data preprocessing strategies, and weight optimization. At the same time, the research on the application of PWV prediction in artificial rainfall enhancement operations also needs to be strengthened., therefore, this paper proposes a new combination forecasting system, which is effectively applied to solve the problem of PWV forecasting, and analyzes the job timing with the forecasted value. The system consists of variational mode decomposition (VMD), multi-objective Harris hawks optimization (MOHHO) and six different types of individual prediction algorithms. Among them, VMD can accurately separate the noise sequence and the main feature sequence of the original PWV sequence. MOHHO overcomes the shortcoming that the single-objective optimization algorithm can only achieve one criterion, and it can optimize the accuracy and stability at the same time. Autoregressive integrated moving average (Arima) and Exponential smoothing (ES), as representatives of statistical models, are responsible for predicting the linear trend of PWV, while Back Propagation neural network (BP), Radial Basis Function neural network(RBF), Long Short Term Memory(LSTM) and Temporal Convolutional Network(TCN), as representatives of neural networks, are responsible for capturing the nonlinear characteristics of PWV. Through experiments and analysis, compared with the individual model such as Arima, BP and LSTM, the prediction accuracy of the proposed combined model is improved significantly in all prediction steps. The PWV threshold and the PWV hourly slope threshold can be used as indicators for the timing selection of artificial rainfall enhancement operations. Using the proposed combined model to predict PWV in precipitation events can accurately obtain the timing of the reflection index, which can provide technical reference for artificial rainfall enhancement operations.