人工智能在水文预报中的应用研究

Hydrological Forecasting Using Artificial Intelligence Techniques

摘要: 全面论述了数据驱动水文模型中人工智能的关键技术及其适应范围，分析了机器学习在水文预报中遇到的技术瓶颈。采用Gamma Test对数据驱动模型进行输入优选，降低了模型的白噪声误差影响；提出了长短期记忆神经网络与批量学习、正则化、筛选神经元技术相结合的深度学习网络，以解决变化环境下降雨–洪水过程统计特征的非线性、随机性和时变性问题。长江上游向家坝~三峡水库区间流域的应用结果表明：在不考虑未来降雨预报的前提下，仅以前期和现时已知的降雨–洪水资料为模型输入，长短期记忆动态神经网络结合三种深度学习辅助算法，防止模型的过参数化和过拟合，有效提高了三峡水库入库洪水的预报精度，1~3 d预报精度均达到了甲等水平。

Abstract: The key techniques and bottlenecks of artificial intelligences in data-driven hydrological model were reviewed thoroughly. Gamma test method was used to optimize the input combination of data-driven model to reduce the white noisy error. The machine learning techniques, such as batch-size learning, regularization and drop out neuron were incorporated into a long-short-term memory (LSTM) deep learning neural network to simulate nonlinear, stochastic and non-static processes in hydrological forecast under changing environment. The application results in study area between Xiangjiaba and Three Gorges Reservoir inter-basin indicate that the forecasting accuracy from one to three days lead-time reaches A-grade (reliability ≥ 85%) and is improved effectively by integrating LSTM neural network and three deep learning auxiliary algorithms in the interests of conquering model over parameterization and overfitting bottlenecks.