基于商空间粒度计算的大气质量组合预测分析-prediction and analysis of atmospheric quality combination based on quotient space particle size calculation.docxVIP

AbstractWith the rapid development of modern industry, environmental pollution problem is increasingly serious .More and more people are concerned about the impact of air quality on human health. Now the researches on air quality forecast at home and abroad are mostly based on the discussion of relationship between the concentration of the atmospheric pollutants and environmental parameters. Different methods are put forward, and models are built to forecast the future concentration of pollutants. Methods of statistical (including grey model, regression model etc.) and artificial neural model (including BP algorithms, SVM etc.) are used popularly. But from the view of granularity, they are the similar. These methods are merely researched in a single granular space, which just consider parts of characters of series and neglect the correlations among different series. Therefore they may be not able to discover the inherence of samples.In the artificial neural network models, though SVM can get the optimal solution, the results are merely discrete values with only two sorts. BP algorithm can get continuous values, but its disadvantages will confine its actual use, because it falls into local optimum easily and always forgets the former knowledge which has been learned.Accordingly, this dissertation puts forward to applying Granularity of the Quotient Space Theory and Combinatorial Prediction on air quality forecast. Samples are constructed and analyzed from granularity coarse to thin step by step, which based on the Quotient Space Theory. The samples are constructed in different granular space, which could mine the essence of information in or among the data. When learning the samples, the combinatorial prediction model is built by a constructive neural network and BP algorithm. This model not only takes advantage of the high accuracy of categorizations by the constructive neural networks, but also overcomes the shortcomings of the neural networks that the prediction r