Artificial Neural Network Simulation And Prediction Of PM₁₀ Concentration In Air

Artificial Neural Network has three applications in the atmospheric science: Prediction, Function Simulation and Classify. We use BP Neural Network which is mature both in theory and application to forecast the concentration level of PM₁₀. This progress in research on Simulation and Prediction the concentration of PM₁₀ was reviewed. The analysis of the contribution rate of PM₁₀ impact factor, the establishing and the modifying of Neural Network models and the analysis of the experiment result were described in detail.The variety of the concentration of PM₁₀ was affected by the degree of atmosphere pollution and climate conditions. Therefore, apart from factors of the temperature, relative humidity, rainfall, wind speed and wind direction, the concentration of NO_x and SO₂ and other pollution influence the PM10 concentration greatly. In the experiment, we used Levenberg-Marquardt model to analyse the factors, such as relative humidity, wind speed ,wind direction, and the concentration of NO_x and SO₂. We obtained the five factors' relativity with the concentration of PM₁₀ are 0.503,0.213,0.159,,0.629,0.757. Obviously, the concentration of SO₂ has the greatest influence, and the lowest is wind direction.In the experiment, we ascertain the effect factors is relative humidity, wind speed and the concentration of NO_x and SO₂. In the simulation, we have trained the model using four neural network BP1-4 which are: Gradient Descent with Momentum Algorithm, Levenberg-Marquardt Optimization Algorithm, Bayesian Regularization method and Early Stopping method respectively. Meteorologic dates of the experiment were obtained form environment mornitoring center of Hangzhou .The result shows: changing number of neurons in hidden layers has evident effects to stability, and of modles. Regarding the functions, for testing 1, Mean Absolute Error (MAE), Mean Sguare Error (MSE) and their relative index(R) of BP3 and BP4 are: 0.0448, 0.0037, 0.3224 and 0.0568, 0.0064, 0.1457. For testing 2, Mean Absolute Error (MAE), Mean Sguare Error (MSE) and their relative index(R) of BP3 and BP4 are: 0.0473, 0.0041, 0.684 and 0.0591, 0.0068, 0.52.After comparing the effects of model's stability, function of the convergence, the spreading ability of model, we found a best model. In the base of best model, we build a forecasting system to forecasting the effect of PM₁₀ Concentration when the concentration of NO_x and SO₂ dropped gradually. The result is approach to theobserved, the error is about 5%. The result is satisfied to us, the model will also have a good application.