แบบจำลองเชิงบูรณาการผลกระทบของไนโตรเจนสำหรับพื้นที่ลุ่มน้ำย่อยคลองอู๋ตะเภา

Abstract

The aim of this research was to develop the integrated artificial intelligence (Al) model to predict nitrogen effects at U-tapao canal. The research was split into two parts: 1) the water quality data analysis and 2) the development of Al model to predict the nitrate nitrogen (NO, -N) concentration in the U-Tapao canal. A water quality index (WQL) which consists of five parameters, namely dissolved oxygen, biochemical oxygen demand, total coliform bacteria, fecal coliform bacteria, and ammonia nitrogen was used to analyze the water quality. Water samples were collected from 21 sampling sites during 2011-2015. K-means clustering analysis was used to cluster the groups of surface water quality in each sampling site by the percentage of WQl classification criteria. The results show that the surface water quality tends to be decadence, especially in the downstream of U-tapao canal. The problem areas are industrial areas and urban areas. The relationships between the daily values of six water quality parameters (EC, pH, turbidity, DO, ammonia-nitrogen (NH,-N), and NO, -N) and the flow rates were analyzed using correlation analysis. The correlation shows a very strong positive correlation (r=0.87) between EC and NO, -N, while the flow rate and NO, -N produced the largest negative relationship (r=-0.59) and moderate positive correlation. Two types of four layer ANNs of the feed-forward back propagation (FFBP) and cascade-forward back propagation (CFBP) types were constructed for the prediction of the level of NO, -N in the U-Tapao canal. Ninety six of EC and flow rate data, which were collected daily from December 2014 to March 2015, were used as inputs of the ANNs. It is found that the four layer FFBP with 2 neurons in the input layer, 20 neurons in the first hidden layor, 30 neurons in the second hidden layer (2-30-30-1), and a single neuron in the outpur layer with a lan-sigmoid transfer function is the optimal model. The FFBP model produces more accurate results than the CFBP model. Linsar regression analysis was usod to predict NO, -N, the regresaion aulysis posults were compared with the results of the ANNa and do potocomos of do ANDs was botor tan that of the lincar regression analysis, To model nitrogen loading under limited relevant water data, in both quality and quantity, GWLF was selected. The output data, which are comparable to the data obtained from existing complex models, obtained from running this model were then used to train the neural network which was used as the reference model for the developed Al-based nitrogen loading model. This model employs a fuzzy logic unit with an adjustable output membership functions. This adjustment is dependent on the differences of monthly predicted nitrogen loads and the values from both the reference model and measurement. The advantage of this approach is that, without recalibration, the system model will adjust the rule base of the fuzzy unit to reflect changes in parameters in the canal. The system model was then used to study the nitrogen loading under two scenarios, 1) the investment of municipal central waste water treatment systems and 2) the regulation of waste water discharges from factories. It is found that the investment has small impact to nitrogen loading due to main nitrogen loads are from point sources in the area.