كليدواژه :
پيش بيني ماهانه بارش , سامانه اطلاعات مكاني , كلان داده , توزيع مكاني زماني , شبكه ي عصبي كم عمق , شبكه باور عميق
چكيده فارسي :
در جهان امروز، توسعه‌ي سريع، همه‌جانبه و پايدار، هدف اصلي تمامي كشورها از جمله ايران ميباشد. اصلي‌ترين محدوديت پيش روي توسعه‌ي پايدار، وضعيت اقليمي و شرايط محيط زيستي كشورها است. از جمله‌ي مهم‌ترين محدوديت هاي اقليمي در كشور ايران، بارش ناكافي همراه با پراكندگي نامناسب مكاني- زماني است. هدف اصلي اين تحقيق، پيش‌بيني ماهانه‌ي ميزان بارش در منطقه‌ي شمال غربي ايران با استفاده از روش هاي يادگيري ماشين شبكهي عصبي كم عمق و شبكهي باور عميق و همچنين تهيه ي نقشه‌ي پيش‌بيني توزيع مكاني- زماني آن در منطقه‌ي مطالعاتي ميباشد. در اين تحقيق از داده‌هاي بارش به همراه ساير داده‌هاي محيطي و هواشناسي، ازابتداي سال 1951 تا پايان سال 2014 استفاده شده است. ارتفاع و مختصات ايستگاه‌هاي هواشناسي نيز به عنوان پارامتر توپوگرافي از مدل رقومي ارتفاعي بدست آمد. همچنين به منظور انجام پژوهشي جامع و افزايش دقت نتايج، از داده‌هاي بزرگ مقياس اقليمي در كنار ساير داده‌ها بهره برده شده است. وجود پارامترهاي متنوع محيطي، توپوگرافي و اقليمي اثر گذار بر پديدهي بارش، سبب شده است كه در اين تحقيق با كلان داده‌هاي مكاني- زماني مواجه باشيم. در اين پژوهش ابتدا داده هاي مورد نياز جمع آوري شد و بعد ازآماده سازي در پايگاه دادهي غير رابط هاي كاساندرا ذخيره سازي گرديد. در گام بعدي دو شبكهي عصبي كم عمق و شبكهي باور عميق پياده‌سازي شد و آموزش، تست و پيش‌بيني با هر دو مدل صورت پذيرفت و نقشه‌ي پيش‌بيني توزيع مكاني- زماني بارشِ هر دو شبكه براي دوازده ماه سال 2014 تهيه گرديد. مقايسهي مقدار بارش ماهانه ي به وقوع پيوسته با ميزان بارش ماهانه ي پيش بيني شده توسط شبكهي عصبي كم عمق و شبكه ي باور عميق نشان داد كه شبكه ي باور عميق از توانايي بيشتري در مواجه با كلان داده هاي مكاني- زماني و حل پيچيدگي هاي مساله ي پيش بيني بارش برخوردار است. همچنين در اين تحقيق از معيارهاي Accuracy، Precision، Recall و F1 score جهت ارزيابي عملكرد روش‌هاي پيشنهادي استفاده شد. نتايج نشان داد كه مقادير Accuracy براي شبكهي عصبي كم عمق و شبكهي باور عميق به ترتيب برابر 0/67 و 0/71، مقادير Precision به ترتيب 0/69 و 0/69، مقادير Recall به ترتيب 0/7 و 0/8 و مقادير F1 score به ترتيب 0/69 و 0/74 ميباشد.
چكيده لاتين :
In today’s world, rapid and sustainable development is on the top of all countries agenda, including Iran. The major obstacles of sustainable development are climate and environmental conditions. One of the most important climatic constraint in Iran is insufficient rainfall. In addition to world water restrictions, Iran has approximately one third of the average global precipitation. Also the spatial distribution of rainfall due to natural conditions is very heterogeneous and temporal distribution of rainfall as well as the spatial distribution shows a similar trend. For these reasons, the water crisis has become a national predicament. In addition Floods and Droughts are the two faces of the same coin. Over the last few years, the overwhelming majority of disasters have been caused by floods in Iran. So Iran is amongst the few countries that is facing floods and drought simultaneously. Hence, long-term meteorological forecasting is of prime importance and plays a significant role in water resource management and sustainable development. This study presents an approach to forecast the monthly rainfall of north-western part of Iran and produce the spatio-temporal prediction maps in the study area. In this research, precipitation data along with environmental and meteorological information such as minimum monthly temperature, maximum monthly temperature, average monthly temperature, maximum wind speed and mean monthly wind speed from 1950 to 2014 were considered as affecting input parameters. Additionally, topographic parameters, elevation, latitude and longitude were computed from Digital Elevation Mode (DEM). In order to increase the prediction accuracy, large-scale climate data such as North Atlantic Oscillation (NAO), Antarctic Oscillation (AO), Extreme Eastern Tropical Pacific SST (Nino 1+2), Eastern Tropical Pacific SST (Nino 3), Central Tropical Pacific SST (Nino 4) and East Central Tropical Pacific SST (Nino 3.4), is used along with other environmental and topographical data. Considering the diversity of sources along with amount of the input data, we were facing challenges involving big data storage and processing. Hence for data storage, Cassandra as a NoSQL database was used. The two main reasons to choose a No SQL database are robust and reliable architecture and providing a mechanism for storage and retrieval of data. Then, shallow artificial neural network and deep belief network as two branches of machine learning were trained and tested. Forecasted precipitation maps for twelve’s months of the 2014 were produced afterward using both shallow neural network and deep belief network. In order to evaluate and compare the performances of the networks, four criteria, including accuracy, precision, recall and f1 score, were used. The comparison between monthly precipitation forecast and measured precipitation shows that deep belief network is capable of handling very large spatial-temporal data sets and is also able to solve the complexities of forecasting precipitation. The results indicate that the accuracy of shallow and deep neural networks were 0.67 and 0.71, the precision were 0.69 and 0.69, the recall were 0.7 and 0.8 and the f1 score were 0.69 and 0.74, respectively.
