AGST-LSTM: The ConvLSTM Model Combines Attention and Gate Structure for Spatiotemporal Sequence Prediction Learning

Spatiotemporal sequence prediction learning generates one or more frames of images by learning from multiple frames of historical input. Most current spatiotemporal sequence prediction learning methods do not adequately consider the importance of long-term features for spatial reconstruction. Based on the convolutional LSTM (ConvLSTM) network unit, this paper adds a memory storage unit that updates information through the original memory unit in the ConvLSTM unit and uses the same zigzag memory flow as the PredRNN network, which can focus on long-term and short-term spatiotemporal features at the same time. Then, an attention module is proposed to extract important information from the long-term hidden state and aggregate it with the short-term hidden state, expand the temporal feeling field of the hidden state, and propose the attention gate spatiotemporal LSTM (AGST-LSTM) model, which further enhances the model’s capacity to catch the spatiotemporal correlation. This paper validates the model through two different prediction tasks. The AGST-LSTM model has competitive performance compared to the comparative model to some degree, as exhibited in experiments.