基于图卷积记忆网络对珠海臭氧时空预测

The spatio-temporal prediction of ozone in Zhuhai based on graph convolutional memory network

臭氧(O_(3))已成为影响珠三角(乃至广东)空气质量达标的首要因素。数据驱动的统计模型(较数值模式)虽展现出改进的预报能力,但多数未能解析站点数据(非欧结构)之间的空间依赖性。本文基于珠海市6个环保国控站及其周边气象站监测数据,通过构建时空协同的图卷积记忆网络(GCN-LSTM)开展多站点未来3天逐小时O_(3)质量浓度预报。结果表明:GCN_LSTM在不同预报时效均准确还原了O_(3)的年、季节和昼夜变化特征,但对日变化的预报技巧随预报时效增加下降明显。通过与业务数值模式(GRACEs)和长短期记忆网络(LSTM)对比发现:GCN-LSTM表现最优,其72 h预报时效内RMSE和R均值分别为27.13μg/m^(3)和0.64,LSTM表现次之(RMSE=28.44μg/m^(3);R=0.61),而GRACEs与统计模型存在明显差距(RMSE=40.93μg/m^(3);R=0.33)。此外,相较于LSTM,GCN-LSTM全局考虑所有站点及其之间的相互联系,不仅将计算速度提高了71%,而且在不同站点的表现也更为优秀和稳定,同时捕捉秋季O_(3)污染事件的能力也有所提高。最后,敏感性实验揭示出考虑相关性较高的变量作为预报因子可以提高模型能力。

Ozone(O_(3))has become the primary factor affecting air quality over the Pearl River Delta and even the entire Guangdong Province.Although data-driven statistical models have shown improved forecast capabilities compared to numerical models,most of them operate grid-by-grid and cannot resolve the spatial dependence between site data of non-Euclidean structures.Based on in-situ measurements from national environmental stations and surrounding weather stations in Zhuhai,this study performs hourly O_(3)concentration forecasts for up to three days over multiple sites by constructing a graph convolution memory network(GCN-LSTM).The results show that GCN_LSTM forecasts at different lead times could accurately reproduce the annual,seasonal,and diurnal variations of O_(3),but the capability of capturing daily variations decreases significantly with the increase in lead time.Further comparisons with the operational numerical model(GRACEs)and Long Short-Term Memory(LSTM)reveal that GCN-LSTM performs the best,with mean RMSE=27.13μg/m^(3) and R=0.64,LSTM is the second(RMSE=28.44μg/m^(3);R=0.61),and GRACEs presents distinct results(RMSE=40.93μg/m^(3);R=0.33)in 72h forecasting.Compared with LSTM,GCN-LSTM considers all sites and their interconnections,it not only increases the calculation speed by 71%but also performs better and more stably over different sites.Moreover,it is also optimal for capturing O_(3)pollution events in cold seasons.Additional sensitivity experiments reveal that considering more correlated variables improves forecasting capabilities.