استفاده از ترکیب الگوریتم ژنتیک و شبکه عصبی مصنوعی جهت برآورد عمق آب شستگی اطراف پایه‌های پل

پل ها یکی از مهمترین سازه های رودخانه ای هستند که به عنوان کلید راه های ارتباطی اهمیت زیادی دارند. سالانه بسیاری از پل ها در مواقعسیلابی تخریب شده و مشکلات عدیده ای را به وجود می آورند. یکی از مهمترین و موثرترین عوامل تخریب و شکست پل ها، آب شستگیاطراف پایه های پل و تکیه گاه ها است. روند آب شستگی اطراف پایه های پل بسیار پیچیده است و رابطه جامع و کاملی برای تخمین آن وجودندارد. امروزه با پیشرفت علم و تکنولوژی، استفاده از سیستم‌های هوشمند کامپیوتری برای مدل‌سازی پدیده‌های پیچیده و غیرخطی از اهمیت روز افزونی برخوردار شده‌اند. در این تحقیق با استفاده از داده‌های واقعی، کارایی سیستم‌های هوش مصنوعی که شامل ترکیبی از شبکه عصبی پرسپترون چندلایه و الگوریتم ژنتیک بوده مورد بررسی قرار گرفته شده است. از میان مدل‌های با تعداد نرون‌های مخفی متفاوت، شبکه عصبی مصنوعی با سه نرون مخفی دارای کمترین خطا می‌باشد. مقایسه مقادیر نسبت اختلاف میان مدل عصبی-ژنتیک پیشنهادی و معادلات متداول موجود نشان می‌دهد که دقت مدل عصبی-ژنتیک از کارایی بالاتری در مقایسه با سایر معادلات برخوردار می‌باشد. جذر میانگین مربعات خطا در مدل پیشنهادی 0.51 محاسبه گردید در حالیکه این مقدار برای معادلات تجربی موجود بالای 0.89 محاسبه شد.

using the combination of genetic algorithm and artificial neural network to estimate scour depth around bridge foundations

introduction scouring is a natural phenomenon that occurs as a result of the erosive action of water flow in alluvial waterways. this phenomenon is considered a serious threat to the stability of structures located in the flow path, such as the foundations of bridges. among the various reasons for the destruction of hydraulic structures, hydraulic factors play a major role, and among the hydraulic factors, scour plays the most important role.methodology in this research, the combination of genetic algorithm and multi-layer perceptron neural network is used. determining the number of neurons in the input and output layers is simple, as it is induced to the model by the input and output variables. but it is not possible to determine the number of hidden layers and their neurons easily. in most of the past studies, the trial and error method has been used to determine the number of neurons and hidden layers. in this research, since two activation functions (sigmoid function) and tanh (hyperbolic tangent) are used, the number of executions of different models is large. based on this and according to the studies done in the past, the number of hidden layers was considered equal to one layer. in this research, genetic algorithm was used to optimize the error function.results and discussion in this research, 1.5>dr>0.5 is considered as accuracy in each model. the relationship of shen et al. (1969) has an accuracy of 32%. this is while the overestimation and underestimation occurred in this model is 58.7% and 9.5%, respectively, which can be understood that in most cases, the scour depth is calculated more than the actual value. the relationship of melville (1997) and melville and chiu (1988) has an equivalent accuracy of 20.5% and 24.5%, respectively. this is while the overestimation and underestimation occurred in these models are 79.5% and 75.6%, respectively, which means that these two equations can be classified as medium accuracy equations. the relationship of melville and sutherland (1988) with an accuracy equal to 45.3% has been the most accurate experimental relationship in this section. also, in this model, overestimation and underestimation equal to 54.7% occurred, which shows the relatively good symmetry of this model. the relationships of hec-18 and ansari and ghadar (1994) with an accuracy of 11 and 12.6% respectively have the lowest prediction accuracy among the selected empirical relationships. in these models, with an overestimation of 89 and 82.9 percent, respectively, it can be seen that a lot of overestimation has occurred in these two models, and these models do not have proper symmetry, and often the estimates of these two models are higher than the actual values. have been. the intelligent model used in this research with an accuracy of 53.7% has the highest level of accuracy compared to experimental relationships. the overestimation and underestimation occurred in this model are 43.7% and 2.6%, respectively.conclusionscomparing the values of the difference ratio between the proposed neuro-genetic model and the existing common equations shows that the accuracy of the neuro-genetic model has a higher efficiency compared to other equations. the root mean square error in the proposed model was calculated as 0.51, while this value was calculated above 0.89 for the existing experimental equations.