ارزیابی روش های مختلف SOM-AI برای پیش‌بینی سطح آب‌زیرزمینی(مطالعه موردی: آبخوان دشت سلماس)

نویسندگان

چکیده

تغییرات سطح آب‌زیرزمینی، یکی از مهم‌ترین متغیرها در مدیریت آبخوان‌هاست که پیش‌بینی دقیق این متغیر می‌تواند در ارائه راهکارهای مدیریتی برای حفظ این مخازن آب شیرین استراتژیک به‌خصوص در مناطق خشک و نیمه‌خشک، مانند حوضة دریاچه ارومیه راهگشا باشد. با وجود توانایی بالای مدل‌های هوش مصنوعی (AI) در پیش‌بینی سطح آب‌زیرزمینی به‌دلیل ناهمگنی و ناهمسانی محیط‌های هیدروژئولوژیکی، گه‌گاه از کارایی پایینی برخوردارند؛ از این‌رو، استفاده از روش هوشمند نگاشت خود سازمان‌ده (SOM) برای خوشه‌بندی چاه‌های مشاهده‌ای و ترکیب آن با مدل‌های مختلف هوش مصنوعی می‌تواند باعث بهبود نتایج حاصل از مدل‌سازی شد. در این پژوهش، روش‌های مختلف SOM-AI، شامل ماشین ‌بردار پشتیبان (SOM-SVM) و مدل فازی ساگنو (SOM-SFL) برای پیش‌بینی تغییرات سطح آب‌زیرزمینی به‌کار گرفته شدند. بدین‌ترتیب، چاه‌های مشاهده‌ای (OW) در سه گروه G1، G2 وG3 دسته‌بندی و برای هر گروه از چاه‌های مشاهده‌ای مدل‌سازی سطح آب زیرزمینی اجرا شد. ارزیابی نتایج با استفاده از معیارهای RMSE، r2 و NSE نشان داد که حداقل در سه چاه مشاهده‌ای OW2، OW6 و OW9، مدل ترکیبی SOM-SFL عملکرد بهتری نسبت به بقیه داشت. در بقیة چاه‌های مشاهده‌ای مدل ترکیبی SOM-LSSVM برتری نسبی داشت.

کلیدواژه‌ها


عنوان مقاله [English]

Evaluation of SOM-AI models for prediction groundwater level(Case study: The Salmas plain aquifer)

نویسندگان [English]

  • Ataallhah Nadiri
  • Keiwan Naderi
  • Mehdi Kord
  • Asghar Asgari Moghaddam
چکیده [English]

In the recent decades, due to the complexity and nonlinearity of aquifers, artificial intelligence (AI) models have been extensively used in aquifer modeling. The purpose of this research is GWL predicting using hybrid of Self Organizing Map (SOM)-clustering method with Artificial Intelligence (AI) approaches, including support vector machine (SVM) and fuzzy logic (FL). The basic concept and theory of SVM was introduced by Vapnik (1995). The SVM implement the structural risk minimization (SRM) principle. The most important concept of SVM is minimizing an upper bound to the generalization error, instead of minimizing the training error. SVM has two outstanding of excellent generalization capability, and sparse representation. Zadeh (1965) introduced the fuzzy sets. A fuzzy system includes three parts: 1. Fuzzification, the degree of membership in a fuzzy set is defined through a membership function; 2. Fuzzy rule, A fuzzy rule-based model operates on an if-then principle; 3. Defuzzification. Basic FL models and clustering techniques have been combined to provide objective FL modelling techniques, but there are variations with respect to the type of their output membership function and the implication methods. As an efficient mathematical tool, SOM may be used to visualize a high-dimensional data set (Nourani et al, 2016). SOMs reduce data dimensions and show similar patterns. Each SOM network typically includes one input and Kohonen layer. This method may reduce the aquifer's heterogeneity due to unsupervised classification of the aquifers. Based on availability of data, 10 observation wells (OW1-OW10) in the study area were selected for predicting groundwater level. The monthly groundwater level at previous time (GWL_(t_0-1)), Monthly temperature at current time (T_(t_0 )), Monthly precipitation at current time (P_(t_0 )) and monthly discharge of the Zola river at current time (Q_(t_0 )) were used as inputs of models. All data were monthly, of 15 years (2001-2016). 80% of data were used for training and 20% for test step. In order to examine the effectiveness of the models in predicting GWL, the performance measure was quantified for all models using three indices: Root Mean Squared Error (RMSE), Coefficient of determination (r2) and Nash-Sutcliffe Efficiency (NSE).
The Salmas plain with an area about 550 km2, located at north of West Azarbaijan province, northwest of Iran. This area is a part of the Urmia Lake catchment and considered as a semi-arid and cold zone, with average annual temperature of 10.36 °C. Of the 27 observation wells (OW) in the aquifer of Salmas plain, 10 of them were selected to predict the monthly groundwater level. Due to hydrogeological and morphological heterogeneous nature of this aquifer, the SOM-clustering method was used to classify the observation wells. This method produces three groups, labelled as G1, G2 and G3. Sugeno Fuzzy Logic (SFL) models were implemented for each of the three groups of observation wells, using the Subtractive Clustering (SC) technique by systematically increasing the cluster radius from 0 to 1. The input and output clusters were created using the Gaussian and linear membership function, respectively. Least Squares-SVM (LSSVM) type was applied to predict GWL. Selecting suitable kernel function and optimized value of kernel (?) and regularization (C) parameters is important step in implementing SVM. The optimized value of ? and C were determined based on minimizing RMSE. The SFL and LSSVM models were built in three groups for predicting GWL at OW1-OW10 of the study area using training data. The three performance measures of RMSE, r2 and NSE for each of two models (SOM-SFL and SOM-LSSVM) at each of observation wells were calculated. The performance of SOM-SFL and SOM-LSSVM models during the testing was evaluated using the three performance measures for training the models and compared with the results obtained.
This research investigated the hybrids of SOM-clustering method with SFL and LSSVM approaches for predicting groundwater level. Performance measures of these models indicate that their results of two hybrid models (SOM-SFL and SOM-LSSVM) are acceptable. In group one (G1) of observation wells (OW1 and OW7), SOM-LSSVM performs better than SOM-SFL. In OW2, OW6 and OW9, obtained results show that SOM-SFL model has better performance. The proposed hybrid models in this research (SOM-SFL and SOM-LSSVM) could successfully be used in predicting GWL. Combine the results of two models and using of multiple models can improve the final results.

کلیدواژه‌ها [English]

  • Fuzzy logic
  • Groundwater level
  • Support Vector Machine
  • Self- organizing map