A data-driven short-term prediction model called the Tiny-RainNet model is proposed to reduce the cumulative errors caused by multi-step prediction and the complexity of other models. We attempted to improve the accuracy of Doppler radar detection of short-term rainfall prediction using different radar echo maps and numerical model prediction. Rainfall prediction is a complicated temporal- spatial problem. Combined with the convolutional neural network in extracting image context information and the advantages of Bi-directional Long Short-Term Memory (BiLSTM) in processing timing information, 60×10×10 sequential radar echo maps were used as the input of Tiny-RainNet to predict the rainfall in the next 1 to 2 hours.. The proposed Tiny-RainNet, with a root mean square error (RMSE) of 9.67 mm/h, outperformed ConvLSTM, LSTM, FC-LSTM, and AlexNet, whose RMSE is 11.31, 11.50, 14.46, 15.88 mm/h respectively for rainfall prediction.