تخمين نيترات آب زيرزميني دشت همدان- بهار با استفاده از شبكه عصبي مصنوعي و اثر تفكيك داده‌ها بر دقت پيش‌بيني

Estimation of Nitrate in Hamedan-Bahar Plain Groundwater Using Artificial Neural Network and the Effect of Data Resolution on prediction Accuracy

كسب اطلاعات لازم در خصوص نيترات در منابع آب زيرزميني مستلزم اندازه‌گيري‌هاي دقيق دوره‌اي است كه با وجود اندازه‌گيري آن در برخي مناطق، به‌علت حساسيت‌هاي اجتماعي و سلامتي جامعه گزارش نمي‌شود. بنابراين مدل‌سازي آن به علت اطلاع از وضعيت كيفي آب هر منطقه ضروري به‌نظر مي‌رسد. هدف اين مطالعه استفاده از روش شبكه عصبي مصنوعي در برآورد نيترات و مقايسه آن با مقادير اندازه‌گيري شده و بررسي تاثيرپذيري برآورد نيترات از تعداد و ماهيت اطلاعات ورودي به مدل شبكه عصبي بود. داده‌ها از آمار و اطلاعات كمي و كيفي 53 حلقه چاه آب زيرزميني دشت همدان- بهار در دو گروه اطلاعات پرهزينه و كم هزينه، طي سالهاي 1382 تا 1387 اخذ شد. در گروه اطلاعات پرهزينه از 13 متغير مستقل شيميايي به عنوان ورودي شبكه عصبي مصنوعي و در گروه كم هزينه از 7 و 8 متغير به تفكيك براي مدل‌سازي نيترات استفاده شد. مقايسه نتايج آزمون هر سه آرايش، حاكي از توانايي بالاي مدل شبكه عصبي در پيش‌بيني غلظت نيترات است. مقايسه ميانگين خطاهاي حاصل از هر سه مدل شبكه عصبي با آزمون t و آماره Z نشان داد كه تفاوت معني‌داري بين نتايج مدل‌ها وجود ندارد. بنابراين استفاده از اطلاعات گروه دوم در ورودي شبكه عصبي قابل توجيه است. مشخصه‌هاي ورودي مدل پيشنهادي شامل خصوصيات ژيومرفولوژي عمق استاتيك، عمق چاه، مختصات جغرافيايي و اطلاعات كيفي دما، pH ، هدايت الكتريكي نمونه‌هاي آب اندازه‌گيري شده است كه موفق به پيش‌بيني غلظت نيترات با اطميناني بيش از 80 درصد شد كه مويد كارايي مناسب مدل در آبخوان دشت همدان-بهار است.

Nitrate is an important pollutant in groundwater that the costs and consequences of health problems and damages due to this pollutant are not measured. Nitrate is a good indicator of water pollution to organic materials. It is important in drinking water hygiene. Thus, utilizing indirect and low-cost methods for accurate estimation of nitrate is considered. Collecting data regarding nitrate in groundwater requires a series of periodic accurate measurements that consumes a lot of time and costs. Materials and Methods E-mail: zareabyaneh@gmail.com The purpose of this study is the application of artificial neural network in estimating nitrate, and comparing the results with measured values. Another purpose is to study sensitivity related estimates of nitrate entered to the neural network model. Data were obtained from the quantitative and qualitative information of 53 drinking water wells in Hamedan-Bahar plain between 2003 and 2008, divided into two groups of costly and low-cost methods. Costly information group used 13 independent chemical variables as artificial neural network input. Low-cost group used two neural models with 7 and 8 variables for nitrate modeling. In general, independent variables were divided into three groups, shown in figure (1). Fig. 1: structure using artificial neural network Corresponding author: Tel: 09188183441 38 Zare Abyaneh, H., et al. Neuro Solution software was used for modeling artificial neural network. This software apply multilayer perceptrons network (MLP) together with Feed-Forward Back Propagation algorithm (FFBP). This model contains an input layer, a hidden layer and an output layer. For each model, the number of input neurons according to Figure (1) is 13, 7 and 8 neurons. Also, 1 to 24 neurons were entered to the middle layer to repeat, and test method was used. For Network training Levenberg Marquet algorithm and sigmoid activation function were used. For neural network implementation, all existing data were randomly divided into two categories of education (70 percent) and calibrated (30 percent). Discussion of Results Results of implementation of each pattern to the separate training models and test data are given in Figure (2). Testing 1 RMSE=4.93, MAE=3.95 MPE=17, r=0.82 ♦ „ y=0.7359x +6.0615 10 20 30 40 50 - ^ 40 - "3d e u =3 20 -u 1. Cu 10 0 Traning 1 RMSE=5.11, MAE=3.98 y = 0.6648x + 7.7239 10 20 30 40 Observation (mgl"1) 50 50 40 et> S3» ? 2® ■ a. 10 0 ■ 50 _ 40 "©JD S30 50 50 40 30 - Training 2 RMSE=5.35, MAE=4.16 MPE=18.1, r=0.8 ♦ H^S* ♦ * y=0.6411x+ 8.1144 jo | 20 Cu 10 0 1 2® Observation (mgl1) Testing 2 RMSE=5.79, MAE=4.4 ^ MPE=18.4, r=0.78 y=0.6487x +8.0463 20 30 40 Observation (mgl1) 10 50 10 20 30 40 Observation (mgl1) 50 50 _ 40 w> S3o e u ■3 20 - a> 1. Cu 20 30 40 Observation (mgl1) 10 - 0 10 20 30 40 Observation (mgl1) Testing 3 RMSE=5.73, MAE=4.48 MPE=17.9, r=0.8 ^ # ♦ ♦ ♦♦ v = 0.6166x +8.4658 Training 3 RMSE=5.40, MAEf=4.23 MPE=18, r=0.79 ^ ♦ ^ % ♦ ♦ 10 y = 0.6066x + 9.039 50 50 40 Fig. 2: Scatter plots of observed versus predicted nitrate for the three models Figure 2 shows that the lowest error rate is in group 1 with RMSE=4.93 mgl-1, and the highest error rate is in group 2 with RMSE=5.79 mgl-1. Added the amount of rain in the artificial neural network third model, decreased 2.6 percent error. O 100 200 300 400 500 Significant differences were not reported between the results of the three groups. To identify the most suitable model, differences of results were investigated by t-test and Z Index in the significant level of 0.05. The results of t values calculated in t table values for the slope of fit line, and Z Pearson for a significant correlation coefficient (Z>1.96) is given in table (1). Group number Index 2 and 3 1 and 3 1 and 2 0.56 ns 0.61 ns 1.18 ns Z 0.79 ns 0.39 ns 0.29 ns t ns : Non-significant in the level of 0.05 The comparison of t and Z values shows that there is no significant difference between the number of neurons in the first layer (13, 8, 7 neurons) and the nature of it (high-cost and low-cost) (p<0.05). Therefore, the second model can be suggested as the appropriate input in simulation of nitrate. Model input parameters including well depth, static depth, geographical and qualitative temperature information, pH and electrical conductivity of water samples were measured by the successful prediction of nitrate concentration with the confidence of more than 80 percent. To identify the diagnosis effects of the precipitation agent against non-periodic factors in the nitrate concentration of groundwater, observed nitrate concentration values and corresponding rainfalls were drawn, and the estimated coefficient of determination (Figure 3) was determined. Due to low coefficient of determination (R2=0.139), the effective nitrate from rainfall is not high. Low coefficient of determination could have occurred because of the sensitivity of nitrate to some agents other than rainfall. Similar results by Naseri et al. (2006) have been reported among the effective factors in the study of nitrate concentration changes in the groundwater of Golestan province. According to the investigation, nitrate leaching from irrigated agricultural lands, sewage and wastewater, and rainfalls, leaching process is completed. Thus, according to the rainfall decrease due to the recent droughts and municipal wastewater discharges into the plain area, the nitrate concentration observed in Hamedan-Bahar plain groundwater through the non-rain factors seems to be logical. 35 30 25 £ 20 e 4. 15 .1 ,0 Z 5 O Rainfall (mm) Fig. 3: correlation between annual rainfall and nitrate concentration Therefore, the input model 2 that eliminates the effect of rainfall factor is found to be suitable. Increasing trend in the nitrate concentrations during the investigation may be due to other factors such as urban sewage and wastewater of agricultural activities. Conclusions The results in this aquifer showed that the predicted values for all neurological models were less than 4.5 percent. The amount of error can be reduced with more accuracy in the measurement of input parameters and running software in different environments. Changes of mean amount of error in the nitrate simulation process derived from the fluctuations of neural models.