بهبود كارائي مدل شبكه عصبي مصنوعي در شبيه‌سازي رسوب معلق با استفاده از الگوريتم بهينه سازي ازدحام ذرات

Improvement of the efficiency of artificial neural network model in suspended sediment simulation using particle swarm optimization algorithm

برآورد ﺻﺤﯿﺢ ﻣﻘﺪار رﺳﻮب ﻣﻌﻠﻖ رودﺧﺎﻧﻪ ﻫﺎ، ﻧﻘﺶ ﻣﻬﻤﯽ در ﻣﻄﺎﻟﻌﺎت ﻓﺮﺳﺎﯾﺶ و رﺳﻮب، ﻫﯿﺪروﻟﻮژي و ﻣﺪﯾﺮﯾﺖ ﺣﻮزه ﻫﺎي آﺑﺨﯿﺰ دارد. ﺷﺒﯿﻪ ﺳﺎزي رﺳﻮب ﻣﻌﻠﻖ در ﺳﺎﻣﺎﻧﻪ ﻫﺎي ﻫﯿﺪروﻟﻮژﯾﮑﯽ، داراي ﭘﯿﭽﯿﺪﮔﯽ ﻫﺎي زﯾﺎد ﺑﻮده، در ﻋﯿﻦ ﺣﺎل درك و داﻧﺶ ﭘﮋوﻫﺸﮕﺮان از اﺟﺰاء و ﻓﺮاﯾﻨﺪﻫﺎي درون آنﻫﺎ ﻫﻤﻮاره ﺑﺎ ﻋﺪم ﻗﻄﻌﯿﺖ روﺑﺮو اﺳﺖ. اﯾﻦ اﻣﺮ، ﺳﺒﺐ ﮐﺎرﺑﺮد ﻓﺮاوان ﻣﺪل ﻫﺎي ﻫﻮﺷﻤﻨﺪ و از ﺟﻤﻠﻪ ﺷﺒﮑﻪ ﻫﺎي ﻋﺼﺒﯽ ﻣﺼﻨﻮﻋﯽ ﺷﺪه اﺳﺖ. ﺑﺎ اﯾﻦ ﺣﺎل، اﺳﺘﻔﺎده از اﯾﻦ ﻣﺪل ﻫﺎي ﻫﻮﺷﻤﻨﺪ ﻧﯿﺰ ﺑﺎ ﭼﺎﻟﺶ روﺑﺮو اﺳﺖ. ﺗﻌﯿﯿﻦ ﺳﺎﺧﺘﺎر ﻣﻨﺎﺳﺐ ﺷﺒﮑﻪ، ﻣﺴﺘﻠﺰم ﺑﻬﯿﻨﻪﮐﺮدن ﭘﺎراﻣﺘﺮﻫﺎي ﻣﻮرد اﺳﺘﻔﺎده در آن ﻧﻈﯿﺮ ﺗﻌﺪاد ﺑﻬﯿﻨﻪ ﻧﺮون ﻫﺎ و ﻻﯾﻪ ﻫﺎ، وزن و ﺑﺎﯾﺎس و ﻧﻮع ﺗﻮاﺑﻊ ﻓﻌﺎل ﺳﺎزي( ﺑﻮده ﮐﻪ واﺳﻨﺠﯽ ﻣﻨﺎﺳﺐ آن ﻫﺎ ﺑﻪ روش آزﻣﻮن و ﺧﻄﺎ، ﺿﻤﻦ ﮐﺎراﯾﯽ ﮐﻢ، ﻣﻨﺠﺮ ﺑﻪ ﺻﺮف زﻣﺎن زﯾﺎد ﻣﯽ ﺷﻮد. در ﭘﮋوﻫﺶ ﺣﺎﺿﺮ، ﺑﻪﻣﻨﻈﻮر ﺷﺒﯿﻪ ﺳﺎزي ﺑﺎر رﺳﻮب ﻣﻌﻠﻖ روزاﻧﻪ رودﺧﺎﻧﻪ ﻧﯿﺮﭼﺎي در ﻣﺤﻞ اﯾﺴﺘﮕﺎه آبﺳﻨﺠﯽ ﻧﯿﺮ در اﺳﺘﺎن اردﺑﯿﻞ(، از ﺷﺒﮑﻪ ﻋﺼﺒﯽ ﻣﺼﻨﻮﻋﯽ ﭘﺮﺳﭙﺘﺮون ﭼﻨﺪﻻﯾﻪ اﺳﺘﻔﺎده ﺷﺪ. ﺑﻪﻣﻨﻈﻮر آﻣﻮزش ﻣﺪل ﺷﺒﮑﻪ ﻋﺼﺒﯽ، ﻋﻼوهﺑﺮ روش ﻣﺮﺳﻮم ﭘﺲ اﻧﺘﺸﺎر ﺧﻄﺎ، از اﻟﮕﻮرﯾﺘﻢ ﺑﻬﯿﻨﻪ ﺳﺎزي ازدﺣﺎم ذرات )PSO(، ﺑﻪﻣﻨﻈﻮر ﺑﻬﯿﻨﻪ ﺳﺎزي ﻣﻘﺎدﯾﺮ وزن و ﺑﺎﯾﺎس ﻧﺮون ﻫﺎي ﻣﺪل ﻫﺎي ﺷﺒﮑﻪ ﻋﺼﺒﯽ اﺳﺘﻔﺎده ﺷﺪ. ﻫﻤﭽﻨﯿﻦ، ﺑﺮاي اﻓﺰاﯾﺶ ﻗﺪرت ﺗﻌﻤﯿﻢدﻫﯽ ﻣﺪل ﻫﺎ، از ﺧﻮﺷﻪﺑﻨﺪي ﻓﺎزي اﺳﺘﻔﺎده ﺷﺪ. ﻧﺘﺎﯾﺞ ﮔﺮﻓﺘﻪ ﺷﺪه از ﭘﮋوﻫﺶ ﺣﺎﺿﺮ ﻧﺸﺎن داد ﮐﻪ آﻣﻮزش ﻣﺪلﻫﺎي ﺷﺒﮑﻪ ﻋﺼﺒﯽ ﺑﺎ اﻟﮕﻮرﯾﺘﻢ PSO، ﺑﺎ ﮐﺎﻫﺶ ﺧﻄﺎي ﺑﺮآورد رﺳﻮب )ﮐﺎﻫﺶ ﺧﻄﺎي ﺑﺮآورد ﮐﻞ و رﯾﺸﻪ ﻣﯿﺎﻧﮕﯿﻦ ﻣﺮﺑﻌﺎت ﺧﻄﺎ ﺑﻪ ﺗﺮﺗﯿﺐ ﺗﺎ 0/3 درﺻﺪ و 10/4 ﺗﻦ در روز( ﮐﺎراﯾﯽ ﺑﯿﺸﺘﺮي ﻧﺴﺒﺖ ﺑﻪ ﻣﺪلﻫﺎي ﺷﺒﮑﻪ ﻋﺼﺒﯽ ﮐﻪ ﺻﺮﻓﺎ از روشﻫﺎي ﭘﺲ اﻧﺘﺸﺎر ﺧﻄﺎ اﺳﺘﻔﺎده ﻣﯽﮐﻨﻨﺪ، داﺷﺘﻪ اﺳﺖ. ﺑﺎ ﺗﻮﺟﻪ ﺑﻪ اﯾﻦﮐﻪ در ﺑﻬﯿﻨﻪ ﺳﺎزي ﭘﺎراﻣﺘﺮﻫﺎي ﺷﺒﮑﻪ ﻋﺼﺒﯽ، اﻟﮕﻮرﯾﺘﻢﻫﺎي ﺗﮑﺎﻣﻠﯽ )ﻧﻈﯿﺮ اﻟﮕﻮرﯾﺘﻢ PSO( ﻗﺎدر ﺑﻪ اراﺋﻪ راهﺣﻞﻫﺎي ﻣﻨﺎﺳﺒﯽ ﻫﺴﺘﻨﺪ، ﻟﺬا در شبيه ﺳﺎزي ﭘﺪﯾﺪهﻫﺎ و ﻣﺘﻐﯿﺮﻫﺎي ﭘﯿﭽﯿﺪه ﺣﻮزهﻫﺎي آﺑﺨﯿﺰ )ﻧﻈﯿﺮ رﺳﻮب ﻣﻌﻠﻖ( ﻣﯽ ﺗﻮان از اﯾﻦ ﺗﻮاﻧﻤﻨﺪي اﺳﺘﻔﺎده ﮐﺮد

The proper estimation of the amount of suspended sediment in rivers has an important role in erosion and sediment studies, hydrology and management of watersheds. The simulation of suspended sediment in hydrological systems that has a lot of complexity and at the same time our understanding of the components and processes within them is always uncertain led to the use of many intelligent models, including artificial neural networks (ANNs). However, the use of these smart models also faces challenges. Determining the proper structure of the network requires optimization of the parameters used (such as the optimal number of neurons and layers, weight and bias, and the type of activation functions), which their proper calibration, using test and error, leads to a lot of time spent in low efficiency. In this study, a multilayer perceptron (MLP) was used to simulate the daily sediment load of the Nirchai River at the site of the Nair hydrometric station in Ardebil province. In order to train the models, in addition to the error back propagation (BP) algorithm, Particle Swarm Optimization (PSO) algorithm was used to optimize the weight and bias of ANNs. The fuzzy clustering method was also used to increase the power of generalization of the models. The results showed that training of ANN models with PSO algorithm with decreasing estimation error (decreasing the PBIAS of estimation and root mean square error up to 0.3% and 10.4 tons per day respectively) is more effective than ANN models that use only error BP techniques. Due to insufficient recorded sediment data in most hydrometric stations of the country on the one hand and the need to train ANNs with sufficient data on the other hand, the use of evolutionary algorithms (e.g. PSO algorithm) can be a good solution for improving the efficiency of intelligent models.