17:15 〜 19:15
[AHW22-P06] Frontier Exploration in Groundwater Level Forecasting: Precision Prediction Techniques Based on Transformer Models
キーワード：Transformer neural networks, machine learning, water resource management, regional groundwater level prediction
Groundwater is a vital resource for Taiwan, yet long-term groundwater over-extraction in the Zhuoshui River basin has led to severe land subsidence, posing significant challenges for sustainable water management. Developing reliable monitoring and predictive tools is crucial for mitigating these impacts. While Transformer neural networks have revolutionized natural language processing, their potential in environmental research remains largely unexplored. This study investigates the applicability of Transformer-based deep learning models for regional groundwater level forecasting in the distal and mid fans of the Zhuoshui River alluvial fan. Utilizing a 20-year dataset with hydrological observations recorded every 10 days, we compare the predictive performance of the Transformer model against traditional machine learning approaches. The results demonstrate that the Transformer-based deep learning model outperforms the comparative models in both 10-day and 20-day forecasts by effectively capturing complex spatio-temporal relationships between key hydro-meteorological variables, such as groundwater, river flow, and rainfall. The model's self-attention mechanism enhances the identification of groundwater level trends and extreme fluctuations, significantly improving predictive accuracy. Our findings validate the feasibility of applying Transformer-based deep learning models to groundwater level forecasting and further establish their superiority in capturing spatio-temporal dynamics. This study not only underscores the model’s practical utilization in sustainable water resource management under climate change but also opens new avenues for applying deep learning in environmental research, offering valuable insights for Taiwan and beyond.