基于半监督生成对抗网络的最优卷积化数据驱动水库群调度方法

Deep learning provides theoretical and method support for big data-driven dynamic scheduling of complex systems. Aiming at a series of key scientific issues about the urgency of delicacy operation of multi-basin reservoir group, the accuracy of reservoir group operation with multi-objective including ecological protection, flood control, irrigation, water supply and electricity generation and the profoundness of data feature with different origin and structure, this project will conduct an intensive study on the unsupervised generative adversarial networks with law of adversarial loss to adapt to scheduling field. By revealing closed-loop mechanism of cross-domain deep learning, the native supervised generative adversarial networks model will be improved to formulate the mechanism of semi-supervised generative adversarial networks basing clustering by density peaks. Optimal convolution data-driven method for multi-objective optimization based on decomposition will be proposed for theoretical analysis of dynamic performance of optimal matrix chain. After that, global margin ranking strategy will be used in label data processing in order to build generative model and discriminative model of reservoir group operation scenarios driven by big data. Therefore the big data-distributed delicacy operation of reservoir group based on generative adversarial networks will be formed and accomplished with minimum manual intervention. The research achievements of this project will play a vital role in optimal operation of reservoir group for ecological protection, irrigation, flood control and electricity generation, and offer an efficient way to develop the theory of modeling and intelligentialize of big data scheduling based on deep learning.

深度学习理论的涌现为大数据驱动的复杂系统优化调度提供了新的理论和方法支持。本项目针对多流域水库群精细化调度任务的紧迫性、生态防洪灌溉与供水发电并存的精准性、多源异构数据特征呈现的深层次性等一系列关键科学问题，深入研究适应调度领域的无监督生成对抗网络及其对抗损失规律，揭示跨领域深度学习的闭环机制，改进有监督生成对抗网络模型，建立基于密度峰值聚类的半监督范式下的生成对抗网络机制，提出基于分解多目标优化的最优卷积化数据驱动方法，开展最优矩阵链动态性能理论分析，将全局余量排序策略用于标签数据处理，创建大数据驱动的水库群调度方案生成模型和判别模型，从而在最小人工干预条件下，提出并实现基于生成对抗网络的大数据分布式水库群精细化调度方法。本项目的研究成果对水库群生态、灌溉、防洪、发电优化将发挥重要的作用，并为形成基于深度学习的大数据调度模型化和智能化理论提供有效的途径。

水库、水资源生态调度生产是利国利民的基础建设，解决深度学习结合图像的多源异构数据特征呈现的深层次性、多模态生态防洪灌溉与供水发电并存的精准性、大数据驱动的复杂系统优化调度所需的紧迫性是实现水库、水资源生态调度生产的关键措施。且对国家各级民生调度有着重要的意义。本课题针对上述三者特性，提出了结合生成对抗网络的最优矩阵链图像处理研究、多模态多目标优化的水库群调度方法和大数据框架下水库调度方法研究。针对多源异构图像的深层特征，使用生成对抗网络完成图像的深度特征提取，实现对不同复杂图像的归类并提取水流速度、水域流速等多方位上下游水资源信息，实现图像深层次特征数据化、可视化。针对多模态生态调度的精准性。通过最小生镜等方法，解决多模态多目标的生态防洪灌溉于供水发电问题。实现在复杂条件下，精准优化水资源利用的问题。针对大规模水资源复杂调度出现的计算机卡顿、溢出等单台性能计算机失效等情况，搭建在大数据驱动框架下的水资源合理开发与运用的平台，提升调度算法鲁棒性、安全性。最后，将理论成果应用于多级水库，设计一套大数据驱动下结合深度学习的多模态多目标水库、水资源的调度算法。 本课题的研究成果为应用与实际的国家多级水库调度方法提供理论基础，具有重要的理论意义和工程应用价值。