Small-Scale Short-Term Spatio-Temporal Rainfall Prediction Based on EBK, GCN and Bi-LSTM Model

The field of meteorology and hydrology is significantly impacted by the spatial-temporal prediction of rainfall. Rainfall exhibits temporal correlation and periodicity in addition to significant spatial correlation. Currently, the primary objective of the majority of spatiotemporal rainfall prediction models is to extract the local temporal characteristics of rainfall events. We suggest an enhanced model for spatio-temporal rainfall prediction that is based on GCN and Bi-LSTM in order to overcome this obstacle. The model takes advantage of the EBK method in spatial interpolation to effectively generate the spatial distribution of small-scale short-term rainfall. The residual connection is implemented to derive spatial features on the basis of the GCN network, and the sequential operation of dilated convolution is converted to a parallel operation. Subsequently, the inflated sequence is fused into Bi-LSTM to extract short-term local features and macroscopic long-term characteristics of rainfall. To confirm the efficacy of the proposed model, the validation experiment employs historical rainfall data from the National Climatic Data Center for testing. The experimental results demonstrate that ESFEM-DDTM outperforms traditional prediction models in the prediction of spatio-temporal rainfall. The model's mean absolute error, mean absolute percentage error, and root mean square error are 2.6725, 123.153%, and 5.1447, respectively.