方衛華1 2，張 慧1 2，夏童童1 2

(1.水利部南京水利水文自動化研究所，江蘇 南京 210012;2.水利部水文水資源監控工程技術研究中心，江蘇 南京 210012)

　　摘 要：針對水工程現場運行環境惡劣，監測儀器損壞率高、實時性分析不及時，不具備大型大型計算分析設備，從而導致統計分析和建模樣本容量不夠的問題，基于有限稀疏樣本采用輕量化生成式對抗網絡(GANs)進行數據現場生成試驗研究。在優化生成器和判別網絡結構及優選激活函數的基礎上，搭建輕量化GANs模型，采用筆記本電腦在工程現地實現數據快速生成。KL散度及Wasserstein距離分析表明，與時間序列預測方法相比，生成數據與原始數據概率分布之間的距離最高減少33.3%。研究表明：采用基于輕量化GANs的數據生成方法可用于現場數據便捷生成，解決樣本數量不夠問題，為今后應用大樣本建模和實測數據總體特征分析提供有效的解決方法，對進一步推廣人工智能在水利工程上的應用具有重要意義。

　　關鍵詞：輕量化GANs;現場數據生成;充水試驗;時間序列;概率距離;引水隧洞

　　Generation analysis of water filling data of diversion tunnel based on light GANs

　　FANG Weihua1，2, ZHANG Hui1，2, XIA Tongtong1 2

　　(1. Nanjing Automation Institute of Water Conservancy and Hydrology, Ministry of Water Resources, Nanjing 210012, China;

　　2. Research Center on Hydrology and Water Resources Monitoring, Ministry of Water Resources, Nanjing 210012, China)

　　Abstract: To solve the problem of insufficient sample size during statistical analysis and modeling, this paper conducts experimental research using light Generative Adversarial Networks(GANs ) based on finite sparse samples, against the problems in hydraulic engineering site such as poor operation environment, high damage rate of monitoring instruments, no timely real-time analysis and no large computing and analysis instruments. A light GANs model is built to realize fast filling of engineering field data by using laptop, based on the optimization of generator and discriminant network structure and the activation function. Compared with the time series prediction method, the KL divergence and Wasserstein distance analysis can reduce the distance between generated data and original data up to 33.3% in probability distribution. Research indicates that the proposed method is suitable for convenient and efficient filling of field water hammer data and able to solve the problem of insufficient sample size. It provides an effective solution for the future application of large sample modeling analysis and the overall characteristics analysis of measured data, which is of great significance to further popularize the application of artificial intelligence in water conservancy projects.

　　Key words: light GANs; field data generation; water filling test; time series; probability distance; diversion tunnel