The invention discloses a deep space-time prediction neural network training method for radar echo extrapolation. The method comprises the following steps: 1, constructing a model, namely constructingan initialization network, an encoding network and a prediction network; 2, setting training hyper-parameters; 3, inputting a training sequence sample; 4, by taking two front data of an input sequence as input, transmitting a concealed state and an initial memory required by the encoding network forwards by the initialization network, by taking later data input into the sequence and output of theinitialization network as input, transmitting the concealed state and the memory obtained through encoding input forwards by the encoding network, and by taking output of the encoding network as input, finally decoding an output prediction sequence by the prediction network; 5, according to the prediction sequence obtained through forwarding transmission in the step 4, calculating gradients of all network parameters according to loss functions, and finally updating all network parameters according to gradients and learning rates; 6, performing iteration training on a deep space-time prediction neural network.