Precipitation Forecast Based on Multi-channel ConvLSTM and 3D-CNN

The short-term precipitation change has a profound impact on people's daily lives, thus it is vital to predict short-term precipitation accurately. Compared with conventional 2D sequence prediction based on radar echo images, a multi-channel ConvLSTM and 3D-CNN structure for multi-dimensional sequence prediction is proposed. Firstly, data preprocessing is carried out, and radar image dataset are denoised by ground-object elimination and 2D wavelet transform. To solve the problem of unbalanced distribution of original precipitation data, different weights are assigned to different precipitation rates and a new balanced loss function is established. Finally, the data fusion of radar echo images, grid pattern temperature and total precipitation extends 2D data to 3D data, and a short-term precipitation forecast algorithm based on multi-channel ConvLSTM and 3D-CNN structure is propsoed. Moreover, a meteorological data mapping method based on the power and logarithmic transformation is put forward, which balances the data distribution. Experimental results show that compared with some machine learning methods, the proposed model can obtain higher precipitation prediction accuracy.