基于3D卷积神经网络的区域降雨量预报

3D Convolutional Neural Network for Regional Precipitation Nowcasting

摘要: 准确的区域降水量预报，在气象服务领域一直是非常重要的问题。短时降雨量预报的目标是在未来短期(0~6小时)内，对当地区域的降雨强度进行精确和及时的预测。气象站通过预测的短期降雨量数据，与观测的天气预报气象数据进行整合，能够发布城市紧急降雨警报，提供有效的防汛防洪信息。本文根据自动站检测的周边历年降水量数据，以及气象站观测的区域上空不同高度的多普勒雷达回波外推图，提出一种基于深度学习方法的降雨预测模型。所提出的模型基于3D卷积神经网络(3D Convolution Neural Network)，将所建立的网络模型应用于降雨预测的回归问题，并利用合适的指标对模型精度进行评价，对高精度下特定区域的短时期降雨量进行预测。通过实验，在不同网络结构下进行分析对比实验预测值与观测值的均方根误差达到了6以下。该方法能够对区域上空未来短期的降雨量进行准确的预测。该训练模型在气象站整年的数据中预测稳定。

Abstract: Accurate regional precipitation forecast has been a very important issue in the field of meteoro-logical services. The goal of short-term rainfall forecasting is to make accurate and timely predic-tions about the intensity of rainfall in local areas in the short-term future (e.g., 0 - 6 hours). Weather stations can issue emergency urban rainfall alerts and provide effective flood prevention information by integrating the predicted short-term rainfall data with the observed weather fore-cast meteorological data. In this paper, according to the surrounding historical rainfall data of au-tomatic station detection and weather observation area of different heights above the Doppler radar echo extrapolation figure, we proposed a rainfall prediction model based on the deep learning method. Proposed model is based on 3D Convolution Neural Network, the established network model was applied to the regression problem of the rainfall forecast, use the appropriate index to evaluate the accuracy of model under the high precision of short-term rainfall forecast in a particular area. Through experiments, this model can accurately predict the short-term rainfall over the region. With the experiments under different network structure, the root mean square error of predicted value and observed value is below 6. The training model predicts stability in weather station data throughout the year.