高级检索

基于ConvGRU的空气污染预测可视分析系统

A ConvGRU-Based Visual Analysis System for Air Pollution Prediction

  • 摘要: 预测细颗粒污染物浓度是制定防污减排措施的主要途径之一. 然而, 传统的用于预测的大规模数值模拟需要在超级计算机上计算数小时乃至数天, 成本高效率低, 甚至影响实效性. 针对上述问题, 提出一种基于卷积门控循环单元(convolutional gated recurrent unit, ConvGRU)的细颗粒物污染预测方法, 此方法通过设计一个全面损失函数(comprehensive-loss, C-Loss), 综合考虑预测结果与实况之间的绝对误差和相对误差; 通过与常用的均方损失函数对比, 证明C-Loss可以使预测模型更适合细颗粒物; 同时, 根据领域专家需求, 设计一个可交互的可视分析系统, 通过该系统, 领域专家可以高效地获取一系列时刻的预测结果, 从而交互式地深入探索大气污染的形成过程与气象因素之间的相关性, 为进一步制定防污减排方案提供了科学依据; 最后, 通过一系列应用示例全面地分析了污染物的形成原因, 并验证了预测模型的有效性.

     

    Abstract: Forecasting the concentration of fine particulate pollutants is one of the main ways to formulate measures of anti-pollution and emission reduction. However, traditional large-scale prediction simulations must be carried on supercomputer for hours or even days. The high cost and low efficiency even affect its timeliness. In order to resolve these issues, in this paper, we proposed a convolutional gated recurrent unit (ConvGRU for short) model based fine particle air pollution prediction method, the main idea is to design a loss function for fine particle prediction, named Comprehensive Loss Function (C-Loss Function for short), both the absolute error and relative error between the prediction results and the actual values have been evaluated by our C-Loss Function; by comparing with the commonly used Mean Square Loss Function, it is proved that C-Loss can make the prediction model more suitable for fine particles; furthermore, according to the requirements from domain scientists, an interactive visual analytics system has also been designed, in this system, domain scientists can efficiently obtain a series of prediction results, so as to interactively explore the correlation between the formation process of air pollution and meteorological factors, this can offer some scientific support for further making better anti-pollution measures; finally, the effectiveness of our proposed system has been demonstrated through analysis of a series of application examples.

     

/

返回文章
返回