A Spatiotemporal Sequence Prediction Model Based on Conv-LSTM and Causal Inference

Spatiotemporal forecasting, through unsupervised learning, infers future states from historical dynamics, providing significant decision-making potential for applications like weather prediction and urban traffic management. Methods based on Convolutional Long Short-Term Memory (Conv-LSTM) units suffer from performance degradation in long-term forecasting due to the “saturated forgetting” characteristics of their gating mechanisms. Transformer-based methods often capture spurious correlations, which negatively impact predictive accuracy. In this paper, we introduce a causal attention unit that integrates the causal relationships between historical memory and the current state, enabling the prediction of future feature vectors based on causal inference, thereby improving forecasting performance. Additionally, we propose an end-to-end prediction model based on CNN-RNN-CNN architecture, where the encoder/decoder serves as the interface, and the internal structure comprises re-designed stacked WM-LSTM and causal inference units in a dual-branch configuration for enhanced prediction. Beyond the original LSTM units, the newly designed RNN includes a wave-like memory path that connects hidden states across different layers in a manner akin to wave propagation, thereby improving long-term forecasting performance. Ablation experiments are conducted to validate the effectiveness of this component. The results demonstrate that our approach achieves superior performance on prominent public datasets, including Moving MNIST, KTH, and RADAR.