基于数值-循环神经网络相结合的回南天客观预报方法

A objective prediction method of Huinantian based on numerical weather prediction and recurrent neural network combination

摘要:

利用2015-2021年河源国家站资料和东源县回南天观测资料，基于循环神经网络构建以未来24和48小时室内地板温度为预测目标的回归预测模型，并以2022年的观测数据作为测试集对模型性能进行评估。结合使用人工智能与数值预报产品的方法，提升回南天预报水平。检验结果表明：相比传统线性模型，GRU效果更佳。敏感性实验结果表明更准的数值预测结果会提升GRU模型预测的鲁棒性。该方法结合了人工智能和数值预报，其准确率取决于数值预报的露点温度和气温的精确度。

Abstract:

Based on the data from Heyuan meteorological station and the observation data of "Huinantian" in Dongyuan from 2015 to 2021, a recurrent neural networks (RNN) regression prediction model is constructed to predict the indoor surface temperatures for the next 24 and 48 hours , and evaluated on the data from 2022. By combining artificial intelligence with numerical forecasting products, the level of "Huinantian" prediction is improved. The verification results show that the Gated Recurrent Unit (GRU) performs better than traditional linear model. Sensitivity experiments demonstrate that more accurate numerical forecast results enhance the robustness of the GRU model. This method combines artificial intelligence and numerical forecasting, whose accuracy depends on the precision of numerical forecasts for dew point temperature and air temperature.