Doğrusal Ve Doğrusal Olmayan Metotlarla Bir Adım İleri Rüzgar Şiddeti Öngörüsü

Abstract

Rüzgar, bol bulunan ve temiz bir enerji kaynağı olması, kurulan rüzgar enerjisi sistemlerinin işletme maliyetlerinin düşük olması, alternatiflerine göre aynı miktarda enerji üretimini daha küçük bir alanda gerçekleştirebilmesi gibi avantajları sayesinde yenilenebilir enerji kaynakları arasında ön plana çıkmaktadır. Son yıllarda yaşanan teknolojik gelişmelerin beraberinde rüzgar enerjisi sistemleri Dünya'da gün geçtikçe daha fazla yer edinmektedir. Ancak rüzgar enerjisi üretiminin temel faktörlerinden olan rüzgar şiddetinin pek çok parametreden etkilenen, belirsiz ve kesikli yapıda olması rüzgarın enerji sistemlerinde kullanılması konusunda zorluk yaratmaktadır. Bu sorun doğrultusunda yürütülen çalışmalarda, rüzgar şiddetinin öngörüsünün yapılması konusuna ağırlık verilmiştir. Bu çalışmada doğrusal ve doğrusal olmayan toplam 8 adet modelden faydalanılarak bir adım ileri rüzgar şiddeti öngörüsü yapılması ve bu öngörünün doğruluğunun değerlendirilmesi amaçlanmıştır. Bu modeller doğrusal ve doğrusal olmayan modeller olarak iki grupta ele alınan; eksojen modelleri de kapsayan Box-Jenkins modelleri ve yapay sinir ağı modelleridir. Modellerin öngörü sonuçları arasında karşılaştırmanın yapılabilmesi amacıyla yaygın olarak kullanılan R2, MAE, MBE, RMSE gibi hata kriterlerinden faydalanılmıştır. Oluşturulan bütün modeller içinde öngörü performansı en iyi olan model ARX olarak belirlenmiştir. ARMAX modeliyle yapılan öngörü ARMA modelinden daha yüksek başarıma sahip olmuştur. Eldeki rüzgar verisi için eksojen giriş kullanarak öngörü yapılmasının daha uygun olduğu; hareketli ortalama (MA) terimlerinin ise bu veri için anlamlı bir katkı sağlamadığı görülmüştür. Öngörüde en başarısız olan model ise GRNN olarak belirlenmiştir. Doğrusal olmayan sistemlerdeki başarısı ve gürültü toleransının yüksek olması gibi sebeplerle yüksek bir öngörü performansı göstermesi beklenen yapay sinir ağları modelleri ise eksojen modellerin gerisinde kalmıştır.

Wind energy, due to the abundance of the resource and low operating costs of installed systems, is one of the most exploited renewable energy sources. Wind farms operate without producing carbon dioxide (𝐶𝑂2) and require relatively small terrain for the same amount of power compared to its alternatives. Besides, like other renewable energy sources, using wind energy has critical significance to increase energy security for the country. Wind energy has become one of the fastest growing renewable energy resources for electricity generation in recent years due to these advantages. Wind speed is the major factor that affects the wind power generation. However, it is affected by number of parameters such as pressure gradient, local weather conditions, frictional forces, Coriolis Effect. Electricity production from wind confronts many problems related to its uncertain and intermittent nature. In line with the problem, prediction is the key factor for scheduling maintenance of a wind farm, taking preventive actions for extreme wind speeds, integrating wind power in a conventional power grid, ensuring power supply quality while operation costs can easily increase with underachieved prediction. Thus, achieving accurate wind speed prediction is noteworthy research area tha wide variety of methods have been studied such as since it is known each method has different advantages and limitations. Accordingly, in the study, it was aimed to make one-step ahead prediction by using proposed linear and non-linear methods and to evaluate the accuracy of the prediction. Linear prediction models which are also called Box-Jenkins models includes AR model, ARMA model, ARIMA model, ARX model which is simply AR model with exogeneous input and lastly ARMAX model which is ARMA model with exogeneous input. Feed-forward Network, Radial Basis Function Network, and Generalized Regression Network that are the structures of artificial neural networks, were examined in non-linear models category. All of the one-step ahead wind speed prediction models are derived from raw wind speed data that was collected from Atatürk Havalimanı at 10 meters of the year 2005. Data has one minute sampling period and was obtained via Automated Weather Observing System (AWOS). Autoregressive (AR) model predicts the variable of interest (in this case, wind speed) using a linear combination of past values of the variable. Moving Average (MA) model, in a smilar way, predicts the wind speed value using past prediction errors in a model. With combining these two models, Autoregressive Moving Average (ARMA) model is obtained. These three models are exploited under the assumption of the time series is stationary. However, mostly time series may contain trends and seasonality which make the series non-stationary. In this case, differencing is used for making time series stationary. This operation is simply the replacement of data by the difference between its previous values and data itself. Letter I represents differencing in Autoregressive Integrated Moving Average (ARIMA) model. After differencing the time series, prediction is performed as ARMA model. In addition to conventional AR ve ARMA model, by exploiting exogenous input to make prediction, two models are obtained and investigated: ARX and ARMAX models. These models are used shifted wind speed data as exogeneus input. It should be expressed that exogeneous models have the same time series stationarity requirements. In the study, model orders were increased at a certain value and calculations were made for all of the aforementioned Box-Jenkins models to find best predictive model of each structure. Artificial neural network is frequently used nonlinear prediction tool because of having noise tolerance, its ease of use, and its capability of handling complex nonlinear characteristics. Its main difference from conventional Box-Jenkins methods is its adaptivity. Each time data are passed through the network, error between actual and predicted value is calculated and it is reduced by adjusting the weights until it comes to an acceptable level. This process is called learning and the artificial neural network models that were investigated in the study are in the category of supervised learning. Supervised learning is the task for mapping an input to an output by using example data set which consists of input-output pairs. For each ANN model, one-fourth of the wind speed data was seperated for testing of the network such that successive three data goes to train set and fourth data goes to test set. After the network is trained, each model is used for making prediction. Neural networks include input, output and hidden layers that are the groups of input, output and hidden neurons as their names referred. Feed-forward network architecture was constructed such as having one layer of each. Trials were conducted by changing the number of input neurons and hidden neurons, the FFN model with the best prediction performance was found. RBFN have input layer, RBF neuron layer and output layer in addition to one width parameter to adjust. The best model of this architecture was found by changing the number of embedding dimension and width parameter value. Similarly, best GRNN model was found by changing these two parameters. Once the model parameters are obtained for each proposed methods and predictions are performed, in order to assess the prediction performance, R2, MAE, MBE, RMSE are used as metrics in the study. At the conclusion, the comparison table with performance metrics of the best models of each model family and order of these models are given. Graphs that involves actual and predicted value of the best model and correlation graph were drawn for each model. Results show that among all predictive models, ARX has the best prediction performance with 𝑅12=0,999883. ARMAX models give better results than ARMA models. Prediction using exogenous input is more suitable for used wind dataset; and the sliding average (MA) model did not provide a significant contribution for the dataset. It is deduced that despite all the advantages of handling a nonlinear problem, exogenous models should be considered as an alternative to ANNs in prediction problems. Further studies might assess the wind speed prediction performance when prediction horizon is bigger than one as in this study. Additively, to consolidate the outputs of this study, a study might be conducted for wind speed data that is collected from terrains with different topographic factors. Last but not least, wind speed data that has different sampling periods can be used to research prediction performance of proposed models.