Emissions of nitric oxide and nitrogen dioxide, which are named as NOx, are a major environmental and health concern.To react to the climate crisis, the South Korean government has strengthened NOx emission regulations. An accurate NOx prediction model can help companies to meet their NOx emission quotas and achieve cost savings. This study focuses on developing a model which forecasts the amount of NOx emissions in Pohang, a heavy industrial city in South Korea with serious air pollution problems.In this study, the Long-short term memory (LSTM) modeling is applied to predict the amount of NOx emissions, with missing data imputation using stochastic regression. Two parameters (i.e., time windows and learning rates) necessary to run the LSTM model are tested and selected using the Adam optimizer, one of the popular optimization methods in LSTM. I found that the model that I applied achieved the acceptable prediction performance since its Mean Absolute Scaled Error (MASE), the most important evaluation criterion, is less than 1. This means that applying the model that I developed in predicting future NOx emissions will perform better than a naive prediction, a model that simply predicts them based on the last observed data point.

翻译：氮氧化物（NOx，包括一氧化氮和二氧化氮）的排放是重大的环境和健康问题。为应对气候危机，韩国政府已加强氮氧化物排放法规。精确的氮氧化物预测模型可帮助企业完成排放配额并实现成本节约。本研究聚焦于开发预测韩国重工业城市浦项（该市存在严重空气污染问题）的氮氧化物排放量的模型。研究中应用长短期记忆（LSTM）模型预测NOx排放量，并使用随机回归插补处理缺失数据。通过Adam优化器（LSTM中常用的优化方法之一）测试并选择了运行LSTM模型所需的两项参数（即时间窗口和学习率）。研究发现，所应用的模型达到了可接受的预测性能，其最重要的评估指标——平均绝对缩放误差（MASE）小于1。这意味着，应用该模型预测未来NOx排放量的效果将优于基于最近观测数据点进行简单预测的朴素模型。