The Research Of Prediction PM2.5 Mass Concentration Based On Neural Network

At present our country has been entered into the post-industrial era, but at same time extreme discord happened between environment and development. The environment has been huge destroyed in the development process, especially in the air, soil, water and so on several aspects. In recent years, air quality has been rapidly exacerbated, haze appears in countries at higher frequencies and larger ranges, and it cause serious harm to human health through respiratory system. With the improvement of human living standards, people start to pay more and more attention on the environment and the requirements of environmental quality of their own living regional also increase, air quality problems become the focus issue. This article use it as a starting point to mining air quality monitoring data, and find nonlinear relationship between PM2.5 and other air pollutants. Aim at problems such as, PM2.5 mass concentration has big gap between different areas, people are difficult to get forcast information timely and so on, to forcast PM2.5 mass concentration hourly, the paper main work is as follows:(1) According to the formation of PM2.5, the reason for the change of mass concentration was analyzed and the mathematical model was established. The formation of PM2.5 is complex, which contains a direct emission of a particle and two particles formed by the photochemical reaction. Mainly including organic carbon, elemental carbon, soil dust, ammonium sulfate or ammonium sulfite, ammonium nitrate, ammonium salt, semi volatile organic compounds, etc. Combined with the monitoring stations to monitor air pollutants, the final selection of CO, NO2, O3-1, O3-8, SO2, PM10 six kinds of pollutants, as the impact factor of PM2.5 mass concentration.(2) Getting environmental monitoring data of air pollutants, and preconditioning of the data. There are some occasional abnormal data in monitoring data, such as empty data which all values are zero. Before use them we need to remove abnormal data to avoid their affect in prediction results. To normalize after excluding, normalized can make the different magnitude data in the same range, avoid the prediction error caused by the gap of magnitude. The sample data are divided into training data set and test data set with appropriate proportion. Choosing Matlab as the main tool for predictive analysis, because of its high speed, efficiency and low cost in vast matrix operation(3) Research of neural networks, analysis its principle, executing process, parameter setting, calculating process and etc. Using back-propagation neural networks as predicting basis. Then through empirical formula and trial-and-error method to determine the best hidden layer neuron number, design the best network structure and select the appropriate transfer function, training function, learning function.Using the training function to train the network, after network training using sim() with a trained network to predict test data. Statistic the network forcast results, calculating its forecast acceptability, relative error, analyze network performance, advantages and disadvantages.(4) Aim at the disadvantage of back-propagation neural natwork, proposed a improved neural networks, using fuzzy system and genetic algorithm to optimize neural networks. Fuzzy system fuzzing the neural network's input and connection weight, clear the reasoning process of neural network, solve the black-box nature of neural network. Using neural network's input/output as fuzzy system's input/output, using neural network's hidden nodes to express membership functions and fuzzy rules. Genetic algorithm optimized the initial connection weights of neural network and improve the global searching ability, convergence speed of the network, solve the problem of neural network is easy to fall into local minimum. Accroding to the evolution target to select individual fitness function and genetic operation methods. Finally, assign the optimal initial connection weight to the neural network.(5) Write a complete optimized neural network's.m program under Matlab. Enter the pre-processing data into the network to train and test. If the network does not over-fitting, statistical comparison these three predict results. The result shows that genetic algorithm optimization neural network has the best performance in PM2.5 mass concentration, it improved the accuracy of the results and reduced the error rate.