@Article{cmes.2020.012818, AUTHOR = {Weiwen Kong, Baowei Wang,2,3}, TITLE = {Combining Trend-Based Loss with Neural Network for Air Quality Forecasting in Internet of Things}, JOURNAL = {Computer Modeling in Engineering \& Sciences}, VOLUME = {125}, YEAR = {2020}, NUMBER = {2}, PAGES = {849--863}, URL = {http://www.techscience.com/CMES/v125n2/40325}, ISSN = {1526-1506}, ABSTRACT = {Internet of Things (IoT) is a network that connects things in a special union. It embeds a physical entity through an intelligent perception system to obtain information about the component at any time. It connects various objects. IoT has the ability of information transmission, information perception,andinformationprocessing.Theairqualityforecastinghasalways been an urgent problem, which affects people’s quality of life seriously. So far, many air quality prediction algorithms have been proposed, which can be mainly classifed into two categories. One is regression-based prediction, the other is deep learning-based prediction. Regression-based prediction is aimed to make use of the classical regression algorithm and the various supervised meteorologicalcharacteristics toregressthemeteorologicalvalue.Deeplearn- ing methods usually use convolutional neural networks (CNN) or recurrent neural networks (RNN) to predict the meteorological value. As an excellent feature extractor, CNN has achieved good performance in many scenes. In the same way, as an effcient network for orderly data processing, RNN has also achieved good results. However, few or none of the above methods can meet the current accuracy requirements on prediction. Moreover, there is no way to pay attention to the trend monitoring of air quality data. For the sake of accurate results, this paper proposes a novel predicted-trend-based loss function (PTB), which is used to replace the loss function in RNN. At the same time, the trend of change and the predicted value are constrained to obtain more accurate prediction results of PM2.5. In addition, this paper extends the model scenario to the prediction of the whole existing training data features. All the data on the next day of the model is mixed labels, which effectively realizes the prediction of all features. The experiments show that the loss function proposed in this paper is effective.}, DOI = {10.32604/cmes.2020.012818} }