基于GRU神经网络的雷州半岛近海岸能见度短临预报研究

近海岸大气能见度变化具有复杂的非线性和局地性特征，且近海岸气象观测站少，一直是精细化预报业务的难点。利用GRU(Gated Recurrent Unit)神经网络，采用广东省湛江市国家基本气象站及其周边上下游观测资料，构建了雷州半岛近海岸能见度1 h时效短临预报的多站GRU模型、单站GRU模型和逐步回归预报模型，并进行了检验评估。结果表明，相比传统的逐步回归方法，GRU神经网络能更好地识别上下游能见度的时空变化特征，多站GRU模型平均绝对误差(MAE)、均方根误差(RMSE)、决定系数(R2)评分均明显好于多元逐步回归模型。模型结构对能见度短临预报效果至关重要，将上下游的气象特征引入到能见度短临预报模型可显著提升预报效果。多站GRU模型在个例检验中较单站GRU模型的MAE、RMSE分别下降了36%和29%,R2提高了30%，表明多站GRU神经网络对能见度预报具有明显优势，为近海岸能见度的精细化短临预报提供了新思路。

RESEARCH ON SHORT-IMPENDING FORECAST OF NEAR-COAST VISIBILITY FOR LEIZHOU PENINSULA BASED ON GRU NEURAL NETWORK

Atmospheric visibility changes in areas near the coast have complex, non-linear and local characteristics. With few near-coastal meteorological observation stations, it has been difficult for fine forecasting operations. In this paper, a multi-station GRU model, a single-station GRU model and a stepwise regression forecast model for 1 h valid, short-impending forecast of near-coastal visibility for the Leizhou Peninsula were constructed, tested and evaluated using a GRU neural network with the national basic meteorological station of Zhanjiang and its surrounding upstream and downstream observations. The results show that compared with the traditional stepwise regression method, the GRU neural network can better identify the spatiotemporal meteorological characteristics of upstream and downstream visibility changes, and the mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R2) scores of the multi-station GRU model are significantly better than those of the multiple stepwise regression model. Model structure is crucial to the effectiveness of short-range visibility forecasting, and the introduction of upstream and downstream meteorological features into the visibility short-impending forecasting model can significantly improve the forecasting effectiveness. The multi-station GRU model decreased MAE and RMSE by 36% and 29%, respectively, and improved R2 by 30%, compared with the single-station GRU model in a typical case, indicating that the multi-station GRU neural network model has obvious advantages for visibility forecasting and provides a new idea for refining short-impending forecast of near-coastal visibility.