基于BP神经网络的农田大气氨浓度预测

农业源氨排放是大气氨最主要的来源，其中氮肥施用是最主要的农业氨排放源之一。预测大气氨浓度的变化，确定影响大气氨排放的因素，可为科学管理农田，减轻环境污染提供参考。本文利用BP神经网络分析农田大气氨浓度及与各气象因素的关系，以便清晰地了解农田大气氨浓度的变化规律，为研究农田大气氨提供一种新的思路与方法。首先选取2015年5—10月的农田大气氨浓度数据及气象监测数据，建立以气象因素（气压、气温、相对湿度、降水量、风速和日照时数）为输入变量，农田大气氨浓度作为输出变量的预测模型。其次采用主成分分析法筛选出对农田大气氨浓度影响较大的气象因素，分别为气温、相对湿度、降水量和风速，然后把筛选出的4个主要因素和原来的6个因素分别作为BP神经网络预测模型的输入变量，利用神经网络模型对农田大气氨浓度进行预测。结果显示，农田大气氨浓度的实际值为0.148 5 mg·m^-3，4个因素的预测值为0.159 4 mg·m^-3，6个因素的预测值为0.173 2 mg·m^-3，预测误差分别为7.35%、16.65%，并且4个因素的预测相对误差为1.4%~27.0%，而6个因素的预测相对误差为1.1%~45.0%。预测的农田大气氨浓度在前5 d内变化较大，但随着时间的推移，农田大气氨浓度逐渐变小趋于平缓，且预测值与实际值的变化趋势基本相符。利用4个因素作为输入变量建立预测模型，预测得到的农田大气氨浓度值比6个因素作为输入变量得到的农田大气氨浓度值与实际值更吻合，相对误差值较小。可见，通过主成分分析法去除冗余因素后建立的神经网络模型更加有效，预测结果比筛选之前的预测效果更好，所建立的模型对甄选关键因素具有较好的适用性，并且神经网络预测模型对农田大气氨浓度的预测精度较高。本文构建的农田大气氨浓度预测模型可为农田大气氨浓度分析及相关研究提供方法和思路上的指导。

Prediction of atmospheric ammonia concentration in farmlands using BP neural network

The emission of ammonia from agricultural system is the main source of atmospheric ammonia and nitrogen fertilizer application is the main sources of ammonia emission in agriculture. The prediction of the changes in atmospheric ammonia concentration and the determination of the factors driving atmospheric ammonia emission will be benefit to the basis for scientific and rational farmland management and for control of environmental pollution. In this paper, BP neural network was used to analyze the concentration of ammonia in farmlands and its relationship with various meteorological factors. The aim was to better understand the changes in ammonia concentration in farmlands and provide new idea and method of study of ammonia in farmlands. First, farmland ammonia measured with Laser Analyzer and meteorological monitoring data from May to October 2015 were used to establish a model for the prediction of farmland ammonia with meteorological factors (air pressure, air temperature, relative humidity, precipitation, wind speed and sunshine hours) as input variables. Secondly, principal component analysis was used to screen meteorological factors with the strongest effect on the ammonia concentration in farmlands, including air temperature, relative humidity, precipitation and wind speed. Then four main factors and six original factors were used as input variables to predict ammonia concentration in farmlands in the region. The results showed that the actual ammonia concentration in farmlands was 0.148 5 mg·m^-3, the predicted value based on four factors was 0.159 4 mg·m^-3 (with predicted error of 7.35%) and the predicted value based on six factors was 0.173 2 mg·m^-3 (with predicted error of 16.65%). The range of the relative error of four prediction factors was 1.4%-27.0% and that of six factors was 1.1%-45.0%. The predicted concentration of ammonia in farmlands varied greatly in the first five days, decreased gradually and apparently flattened out with time, which was basically consistent with the measured value. Eventually, four factors were used as input variables in building the prediction model. The predicted values of farmland ammonia concentration used four factors were more consistent with measured values than that used six factors. It was noted that the established neural network model after removing redundant factors by principal component analysis method was more effective, and the prediction results were better than those before screening. The model established had better applicability for selecting key factors, and the prediction accuracy was higher. The model constructed in this paper for predicting ammonia concentration in farmlands provided more accurate method and newer idea than before on the analysis of farmland ammonia concentration and the related research.