應用倒傳遞類神經網路於PM 10 預測之的研究 水土保持學系.docxVIP

應用倒傳遞類神經網路於PM10預測之研究 黃隆明(1) 古緯中(2) 摘 要 本研究以倒傳遞類神經網路進行PM10濃度之預測，為提升其預測力，除了以線西空氣品質監測站之氣象參數與環境監測參數作為輸入變數之外，並增加梧棲氣象站之氣象參數作為輔助之輸入變數；而在輸入倒傳遞類神經網路前，先進行正規化及數據篩選，而後經由訓練與測試，可獲得內部參數之設定值，當隱藏層單元數為17個、訓練次數為35,000次、學習速率為0.9與動量修正係數為0.1時，所得之預測值最佳。接著再進行變數之篩選，得到最佳輸入變數之組合為前一日PM10、濕度、風速、風向、日照時數、全天空輻射、降雨量與氣壓等八項；又於敏感度分析中發現，影響模式敏感度之主要因子為前一日PM10值。研究結果顯示，輸入變數經由篩選後，不但可提升預測值與實際值之相關性，亦可提升對極端值之預測能力，且不影響PM10季節性夏低冬高之分佈趨勢。 (關鍵詞：倒傳遞類神經網路、PM10、最佳輸入變數組合) Application of Back-Propagation Neural Network on the Prediction of PM10 Long-Ming Huang(1) Wei-Jhong Gu(2) Associate Professor(1), Graduate Student(2), Department of Soil and Water Conservation, National Chung-Hsing University, Taichung, Taiwan, R.O.C. ABSTRACT This study used back-propagation neural network to predict the concentration of PM10. For improving the prediction, meteorological parameters and environmental monitoring parameters from the Siansi air quality station were taken as input variables, and also used meteorological parameters from Wuqi weather station to be supporting input variables. Before setting parameters in the model, the data should be normalized and screened, and the results of setting parameters which could obtain the best prediction through training and testing were as follows: the number of hidden neurons and training times were 17 and 35,000 respectively, the learning rate was 0.9, and the momentum modified coefficient was 0.1. The next step was screening variables, and the results showed that the optimum input variables included PM10 of the previous day, humidity, wind velocity, wind direction, sunshine hours, surface radiation, rainfall, and atmospheric pressure. The sensitivity analysis showed that PM10 of the previous day was the main factor which affected the sensitivity of the model. The result of this study showed that after sieving variables, it could surely improve the correlation between predictions and actual observed values and the prediction of extreme values. Also, it didn’t affect the concentration o