傅立葉神經運算子在短延時降雨預測之應用

Abstract

近年來，機器學習在語音識別、醫學、材料等多個科學領域皆有重大突破。 部分基於機器學習實現極短期定量降水預報研究指出，其模型之預報表現在特定 量化指標中優於傳統方法預報結果。儘管機器學習預報在研究數據上更具優勢， 但其產生之預報容易有空間細節流失、影像過於平滑等問題，使其難以活用於實務上。在預估流量、洪水預報等應用上，可描述局部降雨峰值變化之高解析度預報尤為重要。 綜上所述，本研究旨於探討機器學習應用於極短期降雨預報之能力，並提升機器學習預報描述空間細節之能力。通過分析基於生成對抗網路框架的機器學習降水預報方法 (Ravuri et al., 2021)，本文總結了致使預報影像平滑之原因及可能改善方法，並針對此缺失提出了基於傅立葉之神經網路模型。為評估其預報表現， 本文比較了傅立葉神經網路與諸多先進的極短期定量降水預報方法。儘管成果指出傅立葉神經網路模型無法提升的整體預報表現，但其有效的提升了機器學習模型描述空間細節之能力。

Machine learning (ML) has led to significant breakthroughs in various scientific fields, such as speech recognition, medical, materials, and many more. In recent years, a variety of attempts to apply ML to short-term rainfall forecasting (nowcasting) were also reported. These models have demonstrated the potential gains that might be achieved with ML-based nowcasting models; and in some literature, ML-based methods have been reported to outperform the state-of-the-art non-ML nowcasting methods. However, the predicted rainfall images from many of these models (and their variants) become overly smooth rather quickly; this is a common ’feature’ of many other ML models. This means that significant amount of spatial rainfall details is lost, which is undesirable for certain hy- drological applications, such urban flow and flood forecasting where small-scale rainfall variability in particular localised peaks may have tangible impacts. To address the above issues, this thesis focuses on exploring the capacity of ML-based nowcasting methods. More specifically, through investigating existing ML-based methods and through reproducing the state-of-the-art ML-based model a nowcasting model proposed by DeepMind in 2021 based upon a Generative Adversarial Network (GAN) framework (Ravuri et al., 2021), the key techniques employed by these methods are iden- tified and analyzed. Based upon this, a new ML-based model that incorporates a modified Fourier neural operator is proposed and compared with a number of cutting-edge nowcast- ing models. The comparison results suggest that, although the proposed model does not lead to the best overall performance, its ability to reproduce the observed rainfall features across various spatial scales demonstrates the potential of the proposed model.