基于阶段式时序注意力网络的PM2.5鲁棒预测

PM_2.5浓度的预测对于大气污染治理、改善环境质量等起到重要作用。受气象条件变化与大气污染物排放等多种因素的交叉影响，PM_2.5预测通常易受突变事件及噪声数据干扰。因此，基于对气象条件以及大气污染物与PM_2.5的相关性分析，提出阶段式时序注意力网络模型（staged temporal-attention network，STAN），该方法融合多段注意力学习模块与循环神经网络，建模气象因素与大气污染物对PM_2.5浓度的交叉影响。统计分析北京市、上海市、广州市预测结果的绝对误差值，可知:1）对比广泛使用的单一类模型支持向量机（support vector machine，SVM）、长短期时序记忆方法（long short-term memory，LSTM）和多层感知机（multilayer perceptron，MLP），STAN可达到10%以上的性能领先；对比最新的融合类模型U型网络（U-net），STAN领先了7%的优势。2）以北京市冬季预测结果为例进行统计分析，STAN的预测值与实测值之间的拟合系数可有95.2%的性能领先。此外，在鲁棒性分析中发现，STAN在含有10%噪声的数据上进行预测，误差上升幅度仅为9.3%。结果表明:注意力机制与时序学习模块相结合能够深度挖掘PM_2.5变化规律并抑制噪声数据，且STAN模型可以进行PM_2.5浓度的鲁棒预测。

PM2.5 ROBUST PREDICTION BASED ON STAGED TEMPORAL ATTENTION NETWORK

The forecast of PM_2.5 concentration plays an important role in air pollution control and improvement of environmental quality. Affected by multiple factors such as changes in meteorological conditions and air pollutants emissions, the PM_2.5 forecast was usually susceptible to sudden changes and noise data. Therefore, in-depth exploration of the PM_2.5 concentration change law and modeling robust prediction models have become key steps in this task. Based on the analysis of the correlation between meteorological conditions and atmospheric pollutants on PM_2.5, a staged temporal-attention network (STAN) was proposed. This method combined a multi-stage attention module and a recurrent neural network (RNN) to model the cross-influence of meteorological factors and atmospheric pollutants on PM_2.5 concentration. Statistical analysis of the absolute error values of the prediction results of Beijing, Shanghai, and Guangzhou showed the following results:1) compared with the widely used support vector machine (SVM), long short-term memory (LSTM), and multilayer perceptron (MLP), the performance of the proposed method increased more than 10%. 2) compared with the latest fusion model U-net, the proposed STAN still achieved a decrease in the error of 7%. Taking Beijing's winter forecast results as an example for statistical analysis, the fitting coefficient between the predicted value of STAN and the measured value could reach 95.2%. In the robustness analysis, it was found that the error increased by only 9.3% when the proposed method ran on data with 10% noise. The above results proved that the combination of the attention mechanism and the temporal learning module could deeply mine the change law of PM_2.5 and suppress noise data. It also showed that the STAN could achieve the robust prediction of PM_2.5 concentration.