基于改进双阶段注意力机制的降水智能预报

Intelligent precipitation forecast based on improved dual-stage attention mechanism

为提高现有时间序列算法降水预报的准确率,本文提出一种基于改进双阶段注意力机制的时间序列降水预报模型(DeepAMogLSTM).该算法分为两部分,在输入注意力机制中,使用三层注意力机制对输入序列进行多重关注,选择稳定的输入特征;在时间注意力中,通过选择与目标值最相关的隐层状态,捕获时间序列的长期相关性.算法同时引入形变长短时记忆网络(Mogrifier LSTM),增强模型特征表示能力.模型使用2016-2019年预处理的自动站点特征数据和欧洲中期天气预报中心(ECMWF)气象场模式资料进行集成预报,并利用同期实况观测资料进行模式预报订正.实验结果表明:该模型在时效为2 h的降水预报中,各项数值评价指标均有改善,其中均方根误差为1.877 mm,平均绝对误差为0.727 mm,拟合优度(R^(2))为0.783;同时与其他模型预报订正效果相比,该模型较好地拟合了实际降水空间分布.

In order to improve the existing time series algorithms for precipitation forecasting,this paper proposes a time series precipitation forecasting model(DeepAMogLSTM)based on improved dual-stage attention mechanism.The algorithm can be divided into two parts.In the input attention stage,a three-layer attention mechanism is designed to pay multiple attention to the input sequence;while in the time attention stage,the hidden state most relevant to the target value is selected to calculate the long-term correlation of the time sequence.In this manner,input features can be stably selected and input into the prediction structure.The algorithm also introduces Mogrifier LSTM(Long Short-Term Memory)to enhance the feature representation ability.The model uses preprocessed automatic station data from 2016 to 2019 and ECMWF weather field model data for integrated forecast,and corrects the model forecasts using observation data of the same period.The experimental results show that the evaluation indexes of the model are improved in the 2-hour precipitation nowcasting,in which the maximum square root error is 1.877 mm,the maximum average absolute error is 0.727 mm,and the goodness of fit(R^(2))is 0.783.At the same time,the modeled precipitation fits actual precipitation in spatial distribution,which is better than the correction effect of other models.