基于门控循环单元神经网络的大气能见度临近预报技术

Research on Nowcasting Technology of Atmospheric Visibility Based on GRU Neural Network

针对大气能见度变化具有突变性以及复杂非线性问题,利用神经网络对复杂非线性过程拟合能力强且对映射关系变化反应速度快的特点,采用门控循环单元(GRU)神经网络为算法框架,将某省17个国家气象观测站近5年的地面能见度及相关要素数据预处理后形成本地化能见度数据集,通过该数据集对网络进行1~4 h预报时效的训练、测试与验证。实验结果显示,基于GRU神经网络的大气能见度短临预报算法其均衡平均数(F1-score)、准确率(accuracy)和风险评分(TS-score)指标明显优于长短期记忆神经网络(LSTM)、临近K指数(KNN)与支撑向量机(SVM)大气能见度短临预报算法。

Aiming at the problem of sudden change and complex nonlinearity of atmospheric visibility changes,this paper utilizes the characteristics of neural network that it has strong fitting ability to complex nonlinear processes and responds quickly to changes in the mapping relationship.It uses the gated recurrent unit(GRU)neural network as the algorithm framework,the ground visibility and related element data of 17 national meteorological observation stations in a province in the past five years are preprocessed to form a localized visibility dataset.The dataset trains,tests and validates 1 h,2 h,3 h and 4 h forecast aging time for the network.The experimental results show that the GRU neural network-based atmospheric visibility short-term forecasting algorithm has significantly better equilibrium average(F1-score),accuracy(accuracy)and risk score(TS-score)indicators than long short-term memory neural network(LSTM),nearby algorithm(KNN)and support vector machine(SVM)short-term forecasting algorithm for atmospheric visibility.