Talep tahmini için gri temelli bir yaklaşım

Abstract

Gri sistem teorisi ilk olarak 1989 yılında literatüre kazandırılmış bir çalışma alanıdır. Gri sistem teorisinin alt konularından biri olan gri tahmin ise, bu tez çalışmasının da temelini oluşturmaktadır. Talep tahmini çalışmaları literatürde sıklıkla karşılaşılan alanlardan olup, bu çalışmada talep tahmini için gri tahmin modellerinden yararlanılmıştır.Talep tahmininde ise enerji ve dolayısıyla elektrik talebi (tüketimi) tahmini çalışmaları günümüzde önem arz etmektedir. Kapasitenin doğru planlanması ve doğru fiyatlandırılma politikası, doğru tüketim tahminleriyle başarılı olacaktır. Tüm bu durumlardan yola çıkarak; bu tez çalışmasının temel amacı, talep tahmini için yeni bir model sunmaktır. Bu çalışmada gri modelin küçük boyutlu verilerle yapılan tahminlerde başarılı olmasının avantajına odaklanarak hata oranını daha da küçükleyecek melez modeller sunmak amaçlanmıştır. Tezde elektrik tüketimi tahmini üzerine yapılan uygulamada kullanılan ve önerilen modeller, literatürde ilk kez kullanılan hibrit modellerdir. Bu yönüyle çalışma literatüre önemli katkı sağlamaktadır. Literatür taraması ile gri tahmin modeller kullanılarak yapılan tahmin çalışmaları ve alanları, elektrik tüketim tahmininde kullanılan gri modeller ve Türkiye'de yapılan elektrik tüketimi tahmin çalışmalarında kullanılan modeller genel olarak incelenmiş ve bu bilgiler ışığında hibrit bir model olan bulanık GM (1,1) parametre optimizasyonu Güve- Işık Optimizasyonu Algoritması modeli önerilmiştir. Önerilen bu modele yuvarlanma mekanizması da eklenerek bir model önerisinde daha bulunulmuştur.Yapılan literatür taramasında gri modellerde parametre optimizasyonu çalışmaları da incelenmiş ve sıklıkla kullanılan metasezgisellerin olduğu çalışmalara yer verilmiştir. Çalışmada literatür taraması ve talep tahmini konusundan sonra literatürde en çok kullanılan Gri modellere ve bu tezin temel modeli olan Üçgen Bulanık GM(1,1) modeline detaylı biçimde yer verilmiştir. Sonrasında ise; metasezgisel algoritmalar konusu incelenmiş ve parametre optimizasyonunda kullanılan metasezgiseller detaylıca anlatılmıştır. Önerilen modellerin denklemleri ayrıntılı bir şekilde ifade edilmiş ve Türkiye'nin kısa dönem elektrik tüketimi tahmini konusunda bir çalışma yapılarak sonuçlar yorumlanmıştır. Uygulama, MATLAB 2018b programından yararlanılarak yapılmıştır. Önerilen modeller literatürde var olan gri model ile kıyaslanarak tahmin performansı başarımı ölçülmüştür. Kullanılan hata ölçütü literatürde sıklıkla kullanılan Ortalama Mutlak Yüzde Hatası (OMYH)'dır. Modellerde kullanılan ve literatürde yeni olan Güve- Işık Optimizasyonu Algoritması'nın performansı ise literatürde parametre optimizasyonu için sıklıkla kullanılan algoritmalarla kıyaslanarak ölçülmüştür. Söz konusu algoritma ile önerilen modeller Genetik Algoritma ile kıyaslandığında daha iyi bir tahmin performansı sergilemekte olup; PSO ile kıyaslandığında en az onun kadar güçlü olduğu sonucuna varılmıştır. Kıyaslama yapılırken programın çalışma süresi ve metasezgisellerin optimum noktayı yakaladıkları iterasyon sayıları dikkate alınmıştır. Uygulama sonuçlarına bakıldığında önerilen modellerin tahmin performansının geliştiği gözlemlenmiştir. GIOA, Genetik Algoritmaya göre süreler ve iterasyon sayıları açısından çok daha iyi sonuç verirken; PSO ile kıyaslandığında ise yakın süreler ve iterasyon sayıları vermekte olduğu gözlemlenmiştir. Son olarak, yapılan araştırma ve edinilen bilgiler ışığında tez çalışmasının sonuçları yorumlanmış ve gelecek çalışmalarla ilgili önerilerde bulunulmuştur.

Grey system theory is a field of study that was first introduced in 1989. Grey forecasting, which is one of the subtopics of grey system theory, forms the basis of this thesis. Grey system theory is a new method that focuses on problems with small samples and lack of information. Obtains useful information by using partial information available in uncertain systems. In this way, grey system theory analyzes indeterminate systems. Thus, the operational behaviors of the system and the evolution principles of these behaviors can be accurately defined and followed effectively. Demand forecasting studies are frequently encountered in the literature, and in this study, grey forecasting models are used for demand forecasting. In demand forecasting, energy and therefore electricity demand (consumption) forecasting studies are really important today. Correct planning of capacity and correct pricing policy will be successful with forecasting the consumption correctly. The models used and suggested in the application on electricity consumption estimation in the thesis are the hybrid models used for the first time in the literature. In this respect, the study makes an important contribution to the literature. With the literature review, forecasting studies and fields using grey forecasting models, grey models used in electricity consumption estimation and models used in electricity consumption forecasting studies conducted in Turkey were examined in general, and in the light of this information, a Triangular Fuzzy GM (1,1) parameter optimization model based on Moth- Flame Optimization Algorithm, which is a hybrid model, is proposed. By adding the rolling mechanism to this proposed model, another model has been proposed. In the literature review, parameter optimization studies in grey models were also examined and studies with frequently used metaheuristics were included. In the study, after the literature review and demand forecasting are explained, the most used Grey models in the literature (GM (1,1) and rolling mechanism GM (1,1)) and the Triangular Fuzzy GM (1,1) model, which is the basic model of this thesis, are given in detail. In the traditional GM (1,1) model, the entire data set is used for forecasting. Sometimes, however, data may exhibit different trends and characteristics. It is recommended to use the GM (1,1) model with a rolling mechanism instead of the traditional GM (1,1) model in such chaotic and different data. Afterwards; the subject of metaheuristic algorithms is examined and metaheuristics used in parameter optimization are explained in detail. Metaheuristics are algorithmic constructs that are applied to various optimization problems, usually with only a few modifications to fit the given problem. Metaheuristics have been classified in various ways in the literature. In this study, the most frequently used classifications are included. GA and PSO algorithm are the most widely used metaheuristics in the literature. GA is one of the most popular evolutionary algorithms. GA is also a population-based algorithm. PSO is the most widely used algorithm among swarm intelligence based algorithms with its flexibility and simplicity. While imitating herd behaviors, it also has ties to evolutionary programming and genetic algorithms. In the study, both algorithms are explained in detail. The working principles of the algorithms are explained step by step and pseudo-codes are written. Finally, the working principles are visualized with flow charts. The Moth-Flame Optimization Algorithm used in the thesis is given in more detail. Algorithm principles are explained and the working mechanism is explained step by step. As in PSO and GA, pseudo -codes are explained in this algorithm and a flowchart is given. Moths fly by taking advantage of the moonlight. Moths, which are insects similar to the butterfly family, have a special way of finding their way at night called "transverse orientation." The Moth-Flame Optimization (MFO) Algorithm was inspired by these special flight methods of moths. The equations of the proposed models have been expressed in detail. Three basic models are explained in detail and PSO, GA and MFO based parameter optimized models are also explained separately. Each algorithm is integrated into the proposed models. The decision variables in the proposed models are the horizontal correction parameter λ, which is a triangular fuzzy number, and α, β and γ which constitutes the improvement parameter a. After the detailed explanation of the proposed models, an application has been made on Turkey's short-term electricity consumption forecasting, and the results have been interpreted. Turkey's electricity demand has been estimated using hourly electricity consumption data. The models presented in the study were used in demand forecasting. The forecasting power of the proposed models was measured using the mean absolute percent error (MAPE), which is an error measure. The TFGM (1,1) model, which exists in the literature and is explained in detail in this study, and the proposed models were compared using the MAPE error criterion. In this study, since Turkey's hourly electricity consumption is considered, it is taken into account that EMRA divides one-day electricity demand into three periods. The hours between 22:00 and 06:00 are called night hours, between 06:00 and 17:00 are called daytime hours, and between 17:00 and 22:00 are called peak hours. Night hours are considered as intermediate load, peak hours as peak load, and daytime hours as base load. The application was made by considering the data size in two different dimensions as 80 % education-20 % test and 70 % education-30 % test. The application has made using the MATLAB 2018b program. The error criterion used is the Mean Absolute Percent Error (MAPE), which is frequently used in the literature. The performance of the Moth-Flame Optimization Algorithm, which is used in the models and is new in the literature, has been measured by comparing it with the algorithms frequently used for parameter optimization in the literature. The models proposed with the said algorithm show better prediction performance when compared to the Genetic Algorithm; It has been concluded that it is at least as strong as PSO when compared to it. While comparing, the running time of the program and the number of iterations in which the metaheuristics reached the optimum point were taken into account. When the application results are examined, it has been observed that the prediction performance of the proposed models has improved. When the results of the application are examined, it has been determined that the estimation error obtained with the MFOA-Rolling mechanism TFGM (1,1) among the proposed models has the lowest estimation error in both data sizes and in all three load blocks. When the run times were examined, it was concluded that MFOA had shorter times than the times obtained with GA and obtained close times with the PSO Algorithm. Based on all these analyzes, it is concluded that the performance of the MFOA used in the proposed and newly introduced models is as successful as the algorithms frequently used in the literature. While MFOA gives much better results in terms of durations and iteration numbers compared to the Genetic Algorithm; when compared to PSO, it has been observed that it gives close times and iteration numbers. Finally, the study was interpreted in the light of the results obtained and suggestions were made for future studies.