Water shortage is a major problem faced by the world. Artificial rain enhancement can improve the precipitation conversion rate to increase precipitation and alleviate the problem of water shortage, but how to choose the operation time is the difficulty of artificial rain enhancement. With the development of detection technology, the main parameter that can be observed and obtained to characterize water vapor is the precipitable water vapor (PWV). Due to the influence of environment and observation, PWV series usually has the characteristics of nonlinearity and nonstationarity. These characteristics bring challenges to the accurate prediction of PWV. This paper constructs a combined model integrating data decomposition and multi model prediction, and carries out multi-step prediction experiment for the PWV data observed by microwave radiometer in Zhengzhou station. The model uses the variational mode decomposition (VMD) technology to decompose and denoise the PWV sequence, and uses the back propagation neural network (BPNN), long short term memory network (LSTM), bidirectional gated recurrent unit (BiGRU) and temporal convolutional network (TCN) models to predict the decomposed data respectively. Finally, the improved grey wolf optimization algorithm (IGWO) is used to determine the optimal weight of the model, and the final prediction value is obtained through the weighted combination. The results show that even in the 5-step prediction, compared with VMD-BiGRU, VMD-BP, VMD-LSTM and VMD-TCN, the root-mean-square error of the newly constructed combined model is at least reduced by 50.0%, 67.6%, 57.9% and 17.2%, respectively, which proves that the model has good stability and good generalization ability. It can provide support for precipitation forecasting and judging the time of artificial precipitation.