Uzun kısa süreli bellek (LSTM) kullanarak İstanbul ve civarı için deprem eğilimi tahmini
Yükleniyor...
Dosyalar
Tarih
2024
Yazarlar
Dergi Başlığı
Dergi ISSN
Cilt Başlığı
Yayıncı
Bursa Teknik Üniversitesi, Lisansüstü Eğitim Enstitüsü
Erişim Hakkı
info:eu-repo/semantics/openAccess
Özet
Doğal afetler, genellikle insanların kontrolü dışında gerçekleşen ve gerçekleşmesi neticesinde başta can ve mal kayıpları olmakla birlikte ekonomik, çevresel, sağlıksal ve daha nice kayıplara yol açan olaylardır. Bu doğal afetlerden birisi olan depremler ile ülkemiz tarihinde sıkça karşılaşılmaktadır. Birçok ülke, kurum ve kişi tarafından depremlerin yıkıcı etkilerini en aza indirmek için çalışmalar yapılmaktadır. Bu çalışmalardan birisi de depremi gerçekleşmeden önceden yeri, zamanı ve büyüklüğüyle tahmin edebilen deprem tahmin çalışmalarıdır. Depremin önceden tahmin edilebilmesi can ve mal kaybını büyük oranda azaltma potansiyeline sahip olduğu için hayati öneme sahip ve depremlerin sürekli yaşanıyor olmalarıyla birlikte popülerliğini hiç kaybetmeyen bir konudur. Aynı zamanda depremlerin oluşumunun çok yüksek karmaşıklıktaki süreçleri içermesi ve analiz edilmesi zor olan çok sayıda faktöre bağlı olmasından dolayı oldukça zor bir konudur. Deprem tahmini üzerine mevcut çalışmalar, kullanılan metodolojilere göre matematiksel analiz, öncüllerin araştırılması, geleneksel makine öğrenmesi ve derin öğrenme olmak üzere dört farklı başlık altında incelenebilmektedir. Öncüller olarak isimlendirilen verilerin kullanıldığı çalışmalar her depremden önce meydana gelmediği için bu verileri kullanan algoritmaları genelleştirmek ve standart hale getirmek oldukça zordur. Matematiksel analiz yöntemleri kısıtlı verilerle çalıştıkları ve düşük başarı sonuçları elde ettikleri için pek tercih edilmemektedirler. Geleneksel makine öğrenmesi deprem tahmininde kullanılan önemli yöntemlerden birisidir. Derin öğrenme yöntemleri, karmaşık problemleri çözmedeki başarısı nedeniyle deprem tahmininde son zamanlardaki en popüler yöntemlerdir. Uluslararası mecralarda deprem tahmini çalışmaları gerçekleştirildiği gibi ülkemizde de çeşitli yöntemler kullanılarak deprem tahmini çalışmaları gerçekleştirilmiştir. Kahramanmaraş'ta meydana gelen depremlerle birlikte yaşamış olduğumuz can ve mal kayıplarının ardından gözler jeologların uzun zamandır üzerinde durduğu ve yapı stoğunun büyük bir çoğunluğunun eskiye dayandığı İstanbul'a çevrilmiştir. Ne yazık ki beklenen İstanbul depremi için İstanbul ve civarına yönelik yeterli sayıda deprem tahmini çalışması yapılmamıştır. Beklenen İstanbul depremi için bir deprem tahmin çalışması yapmak, deprem tahmini konusunda bir kaynak oluşturmak ve bu konuda çalışmalar yapmaya teşvik etmek amacıyla bu çalışmada İstanbul ve civarı için uzun kısa süreli bellek (LSTM) ve evrişimsel sinir ağları (CNN) kullanarak deprem tahmini çalışması yapılmıştır. Bu çalışmadan önce deprem ve deprem tahmin çalışmalarına dair genel bilgiler verilmiştir. Daha sonra deprem tahmini kısmı veri toplama, veri ön işleme, veri bölütleme, model eğitimi ve modellerin değerlendirilmesi aşamalarından oluşmaktadır. Veri toplama aşamasında İstanbul ve civarına ait son 25 yılın tarihsel deprem verileri AFAD'dan temin edilmiştir. Veri ön işleme aşamasında eksik veri tespiti ve veri normalizasyonu gerçekleştirilmiştir. Veri Bölütleme aşamasında veri xv seti eğitim ve test olmak üzere 2 kısma ayrılmıştır. Model eğitimi aşamasında LSTM ve CNN kullanılarak deprem tahmini modeli oluşturulmuştur. Deprem tahmin modeli girdi olarak deprem büyüklüğü, enlem, boylam ve derinlik verilerini ayrı ayrı ya da hep birlikte kullanarak deprem büyüklüğü, enlem, boylam ve derinlik tahminleri yapmaktadır. Model değerlendirme aşamasında deprem tahmin modeli ile birlikte parametreler, tarihsel veriler ve veri seti özellikleri ortalama mutlak hata, ortalama hata karesi ve ortalama mutlak yüzde hatası metrikleri kullanılarak test edilmiştir. Ek olarak eğitilen modeller, modellerin geçerliliğini ölçmek amacıyla bir diğer deprem beklenen yerler arasında olan Bingöl – Karlıova civarı için de test edilmiştir. Testler sonucunda İstanbul ve civarı için kabul edilebilir başarı oranına sahip bir deprem tahmin çalışması gerçekleştirilmiştir. Bununla birlikte bu problem özelinde LSTM ve CNN modellerin ne kadar başarılı oldukları, hangi zaman aralığına ait verilerin kullanımının daha başarılı sonuçlar üretebileceği, özelliklerin öğrenmeye olan katkısı ve hangi parametrelerin daha kullanışlı olduğuna yönelik bilgiler edinilmiştir.
Natural disasters are events that generally occur beyond the control of human beings and as a result of their realisation, they cause economic, environmental, health and many other losses, especially loss of life and property. We frequently encounter earthquakes, one of these natural disasters, in the history of our country. Many countries, institutions and individuals are working to minimise the destructive effects of earthquakes. One of these studies is earthquake prediction studies that can predict the earthquake with its location, time and magnitude before it occurs. Predicting the earthquake in advance is of vital importance because it has the potential to greatly reduce the loss of life and property, and it is a subject that has never lost its popularity with the continuous occurrence of earthquakes. At the same time, it is a very challenging topic because the occurrence of earthquakes involves processes of very high complexity and depends on many factors that are difficult to analyse. Existing studies on earthquake prediction can be analysed under four different headings according to the methodologies used: mathematical analysis, antecedent search, traditional machine learning and deep learning. Since the studies using data called precursors do not occur before every earthquake, it is very difficult to generalise and standardise the algorithms using these data. Mathematical analysis methods are not preferred because they work with limited data and achieve low success results. Traditional machine learning is one of the important methods used in earthquake prediction. Deep learning methods are the most popular methods in earthquake prediction due to their success in solving complex problems. As earthquake prediction studies have been carried out internationally, earthquake prediction studies have also been carried out in our country using various methods. After the loss of life and property we have experienced with the earthquakes in Kahramanmaraş, the eyes have turned to Istanbul, which geologists have been focusing on for a long time and where most of the building stock is based on old buildings. Unfortunately, a sufficient number of earthquake prediction studies have not been carried out for Istanbul and its vicinity for the expected Istanbul earthquake. In order to make an earthquake prediction study for the expected Istanbul earthquake, to create a resource on earthquake prediction and to encourage studies on this subject, an earthquake prediction study using long short-term memory (LSTM) and convolutional neural networks (CNN) was carried out in this study for Istanbul and its vicinity. Before this study, general information about earthquake and earthquake prediction studies is given. Then, the earthquake prediction part consists of data collection, data preprocessing, data segmentation, model training and model evaluation. In the data collection phase, historical earthquake data of Istanbul and its neighbourhood for the last 25 years were obtained from AFAD. In the data preprocessing stage, missing data detection and data normalisation were performed. In xvii the data segmentation stage, the data set was divided into 2 parts as training and test. In the model training phase, an earthquake prediction model was created using LSTM and CNN. The earthquake prediction model uses earthquake magnitude, latitude, longitude and depth data separately or together to predict earthquake magnitude, latitude, longitude and depth. In the model evaluation phase, the parameters, historical data and dataset characteristics along with the earthquake prediction model were tested using the mean absolute error, mean squared error and mean absolute percentage error metrics. In addition, the trained models were also tested for the vicinity of Bingöl - Karlıova, which is among the other expected earthquake locations, in order to measure the validity of the models. As a result of the tests, an earthquake prediction study with an acceptable success rate for Istanbul and its neighbourhood was carried out. In addition, information has been obtained on how successful LSTM and CNN models are for this problem, which time interval data can produce more successful results, the contribution of features to learning and which parameters are more useful.
Natural disasters are events that generally occur beyond the control of human beings and as a result of their realisation, they cause economic, environmental, health and many other losses, especially loss of life and property. We frequently encounter earthquakes, one of these natural disasters, in the history of our country. Many countries, institutions and individuals are working to minimise the destructive effects of earthquakes. One of these studies is earthquake prediction studies that can predict the earthquake with its location, time and magnitude before it occurs. Predicting the earthquake in advance is of vital importance because it has the potential to greatly reduce the loss of life and property, and it is a subject that has never lost its popularity with the continuous occurrence of earthquakes. At the same time, it is a very challenging topic because the occurrence of earthquakes involves processes of very high complexity and depends on many factors that are difficult to analyse. Existing studies on earthquake prediction can be analysed under four different headings according to the methodologies used: mathematical analysis, antecedent search, traditional machine learning and deep learning. Since the studies using data called precursors do not occur before every earthquake, it is very difficult to generalise and standardise the algorithms using these data. Mathematical analysis methods are not preferred because they work with limited data and achieve low success results. Traditional machine learning is one of the important methods used in earthquake prediction. Deep learning methods are the most popular methods in earthquake prediction due to their success in solving complex problems. As earthquake prediction studies have been carried out internationally, earthquake prediction studies have also been carried out in our country using various methods. After the loss of life and property we have experienced with the earthquakes in Kahramanmaraş, the eyes have turned to Istanbul, which geologists have been focusing on for a long time and where most of the building stock is based on old buildings. Unfortunately, a sufficient number of earthquake prediction studies have not been carried out for Istanbul and its vicinity for the expected Istanbul earthquake. In order to make an earthquake prediction study for the expected Istanbul earthquake, to create a resource on earthquake prediction and to encourage studies on this subject, an earthquake prediction study using long short-term memory (LSTM) and convolutional neural networks (CNN) was carried out in this study for Istanbul and its vicinity. Before this study, general information about earthquake and earthquake prediction studies is given. Then, the earthquake prediction part consists of data collection, data preprocessing, data segmentation, model training and model evaluation. In the data collection phase, historical earthquake data of Istanbul and its neighbourhood for the last 25 years were obtained from AFAD. In the data preprocessing stage, missing data detection and data normalisation were performed. In xvii the data segmentation stage, the data set was divided into 2 parts as training and test. In the model training phase, an earthquake prediction model was created using LSTM and CNN. The earthquake prediction model uses earthquake magnitude, latitude, longitude and depth data separately or together to predict earthquake magnitude, latitude, longitude and depth. In the model evaluation phase, the parameters, historical data and dataset characteristics along with the earthquake prediction model were tested using the mean absolute error, mean squared error and mean absolute percentage error metrics. In addition, the trained models were also tested for the vicinity of Bingöl - Karlıova, which is among the other expected earthquake locations, in order to measure the validity of the models. As a result of the tests, an earthquake prediction study with an acceptable success rate for Istanbul and its neighbourhood was carried out. In addition, information has been obtained on how successful LSTM and CNN models are for this problem, which time interval data can produce more successful results, the contribution of features to learning and which parameters are more useful.
Açıklama
Anahtar Kelimeler
Deprem tahmini, Derin öğrenme, LSTM, CNN, İstanbul, Earthquake prediction, Deep learning, Istanbul
