کارایی مدل هیبریدی nna-ann در مدل‌سازی تنش خاک در بدنه ‌سد‌های ‌خاکی‌ در زمان ساخت و مقایسه با مدل عددی (مطالعه موردی سدکبودوال)

تخمین دقیق مقادیر تنش خاک در بدنه سد خاکی در زمان ساخت از اقدامات ضروری برای مدیریت پایداری آن است. در این پژوهش، تاثیرگذارترین ویژگی‌ها در مدل‌سازی تنش خاک به صورت مطالعه موردی (سدکبودوال) با استفاده از الگوریتم هیبرید شبکه عصبی- شبکه عصبی مصنوعی(nna-ann) تعیین شد و مقایسه‌ای بین نتایج مدل هیبریدی با مدل عددی صورت پذیرفت. پنج ویژگی شامل تراز مخزن، تراز خاکریزی، زمان ساخت، سرعت آبگیری و سرعت خاکریزی برای ورودی مدل هیبریدی هوشمند انتخاب گردید. با استفاده از الگوریتم هیبریدی و روش انتخاب ویژگی، ترکیب سه ویژگی، شامل تراز خاکریزی، زمان ساخت سد و تراز آب گیری (با rmse برابر با 5024/0) موثرترین ویژگی‌ها در مدل‌سازی تنش کل در سلول‌ منتخب بودند. نتایج نشان داد که مدل هیبریدی در سد کبودوال با مقادیرr^2، rmse، mae وns به ترتیب برابر با 9943/0، 5653/2، 9973/1 و 9999/0 دارای عملکرد بهتری در مدل‌سازی تنش کل خاک نسبت به مدل عددی با مقادیرr^2، rmse، mae وns به ترتیب برابر با 9625/0، 2567/26، 1667/25 و 9772/0 است. این پیش‌بینی برای سایر سلول‌ها در مقاطع مختلف سد مذکور، نیز قابل استناد است. نتایج حاصل از این تحقیق برای ساخت گاه جدید با مشخصات ژئوتکنیکی جدید یعنی سد مسجد سلیمان نیز معتبر بود ولی برای هر ساخت گاه از ترکیب مناسب به خود باید استفاده کرد.

efficiency of nna-ann hybrid model in modeling soil stress in the body of earth dams during construction and comparison with numerical model (case study of kaboud-val dam)

the stress created in the soil significantly affects its engineering behavior. changing its value during the construction of earthen dams causes volume change and shear strength, causing rupture, soil compaction and settlement in earthen dams. so measuring soil stresses of dams is essential that it is done by instrumentation installed. artificial intelligence models such as artificial neural networks for modeling many engineering applications. also by extending the meta-heuristic algorithms, combined with neural networks have become very popular due to more accurate results.methodologyin this study, the cross-section 19 (cross-section of the middle part of the body and dam foundation) for the modeling of soil stress were used during the construction of the dam kaboud-val. also kaboud-val dam instrumentation data (derived from golestan regional water co.) was used at the time of construction during the period of 4 years. type and number of input data is the most important thing in modeling artificial intelligence. by examining data tpc19.1 cells in section 19 in kaboud-val dam, embankment alignment (f), the water level of the reservoir (r), the construction of the dam (t), speed filling and dewatering speed was selected for input. the soil stress (p) in the body of the dam during construction, intelligent model was selected for output. this process is most effective in a subset of features from the set of input features according to the least error, selected and additional features will be removed. in this research, a meta-algorithm (artificial neural network (nna) algorithm) is combined with an artificial neural network (ann) that has the ability to predict complex and nonlinear relationships and extracts effective features for modeling soil stress with appropriate accuracy. in this study, the most effective features in soil stress modeling were determined in a case study (kaboud-val dam) using the nna-ann hybrid algorithm and a comparison was made between the results of the hybrid model and the numerical model. five features include reservoir level, fill level, dam construction time, impounding velocity and fill velocity was selected for the input.results and discussionusing hybrid algorithm and feature selection method, a combination of three features, including reservoir level, fill level, dam construction time (with rmse equal to 0.5024) were the most effective features in modeling soil stress in the selected cell. the results showed that the hybrid model in kaboud-val dam (with values of r^2, rmse, mae and ns equal to 0.9943, 2.5653, 1.9973 and 0.9999, respectively) has better performance in modeling soil stress than the numerical model. (with values of r^2, rmse, mae and ns are equal to 0.9625, 26.2567, 16.6725 and 0.9772, respectively). the results showed that the reduction in the input features to reduce the time and cost reduction is more economical and more effective. because with the increase in the number of features in the hybrid model, the increase in modeling accuracy did not occur. sensitivity analysis showed that the dam construction time and fill level, of the highest sensitivity factor, the most important feature of the model is the total stress in cells. modeling with the mentioned features, in another dam with a new construction site and new geotechnical specifications (masjed soleiman dam) showed that the use of artificial intelligence model according to statistical indicators has more accurate answers than the numerical model.conclusionthe results showed that the use of artificial intelligence methods in the design and initial estimates of soil stress parameters in earthen dams instead of using numerical methods has high reliability and accuracy. the combination of input data in the hybrid model under study is suitable for kaboud-val dam and masjed soleiman dam and the appropriate combination should be used for each construction site. by completing the number of data in different sections of the dam and the number of construction sites in areas with similar climate and geotechnical conditions, a design software can be obtained to predict the amount of soil stress during construction in the body and foundations of earth dams.