SARNN: A Spatiotemporal Prediction Model for Reducing Error Transmissions

Spatiotemporal prediction has become an important research topic in weather forecasting and traffic planning. Due to the cyclic structure for prediction images frame by frame, the error generation and accumulation has often led to blurred images. In this paper, we propose a new end-to-end spatiotemporal attention recurrent neural network (SARNN) to overcome this problem. A new cyclic core mechanism based on long-short term memory (LSTM) is used for extracting the directions of spatial correlation and temporal evolution feature separately. Specifically, an attention mechanism added in temporal direction allows for adaptively choosing highlight input time step of hidden state, instead of decoder just relying on the output of previous time step; a scale change convolution block has been added in the spatial direction to enhance the capability of extraction multi-level semantic features. The validation experiment on Moving-Mnist and KTH dataset demonstrates that SARNN can output more accurate and clearer prediction frames.