Forecasting the water level of the Han River is essential to control traffic and avoid natural disasters. The stream flow of the Han River is affected by various and intricately connected factors. Thus, a simple forecasting machine frequently fails to capture its serial pattern. On the other hand, a complex predictive model loses the interpretability of the model output. This work proposes a neural network model with a novel transformer exploiting a causal relationship based on prior knowledge. The transformer consists of spatiotemporal attention weight that describes the spatial and temporal causation with multilayer networks with masking. Our model has two distinguished advantages against the existing spatiotemporal forecasting models. First, the model allows the heterogeneous predictors for each site such that a flexible regression is applicable to the causal network. Next, the model is adapted to partially identified causal structures. As a result, we have relaxed the constraints of the applicable causal network through our model. In real data analysis, we use the Han River dataset from 2016 to 2021, compare the proposed model with deep learning models, and confirm that our model provides an interpretable and consistent model with prior knowledge, such as a seasonality arising from the tidal force. Furthermore, in prediction performance, our model is better than or competitive with the state-of-the-art models.

翻译：预测汉江水位对于控制航运和避免自然灾害至关重要。汉江的水流受多种复杂关联因素的影响，因此简单预测模型常难以捕捉其序列模式，而复杂预测模型又会丧失模型输出的可解释性。本文提出了一种神经网络模型，该模型采用基于先验知识挖掘因果关系的创新Transformer架构。该Transformer由时空注意力权重构成，通过多层网络与掩码机制描述空间和时间上的因果关系。与现有时空预测模型相比，我们的模型具有两大显著优势：首先，模型允许各观测点采用异质性预测因子，从而实现对因果网络的灵活回归；其次，模型能够适应部分识别的因果结构。通过这两项改进，我们放宽了适用因果网络的约束条件。在真实数据分析中，我们使用2016年至2021年的汉江数据集，将所提模型与深度学习模型进行对比，证实该模型能提供可解释且与先验知识（如潮汐力导致的季节性）一致的预测结果。此外，在预测性能方面，我们的模型达到或优于当前最先进的模型。