• Energy Research

Abstract Forecasting the future price of crude oil, which has an important role in the global economy, is considered as a hot matter for both investment companies and governments. However, forecasting the price of crude oil with high precision is indeed a challenging task because of the nonlinear dynamics of the crude oil time series, including chaotic behavior and inherent fractality. In this study, a new forecasting model based on support vector regression (SVR) with a wrapper-based feature selection approach using multi-objective optimization technique is developed to deal with this challenge. In our model, features based on technical indicators such as simple moving average (SMA), exponential moving average (EMA), and Kaufman’s adaptive moving average (KAMA) are utilized. SMA, EMA, and KAMA indicators are obtained from Brent crude oil closing prices under different parameters. The features based on SMA and EMA indicators are formed by changing the period values between 3 and 10. The features based on the KAMA indicator are obtained by changing the efficiency ratio (ER) period value, which is considered as fractality efficiency, between 3 and 10. The features are selected by the wrapper-based approach consisting of multi-objective particle swarm optimization (MOPSO) and radial basis function based SVR (RBFSVR) techniques considering both the mean absolute percentage error (MAPE) and Theil’s U values. The obtained empirical results show that the proposed forecasting model can capture the nonlinear properties of crude oil time series, and that better forecasting performance can be obtained in terms of precision and volatility than the other current forecasting models.

In the last decade, the significant increase in the use of renewable energy sources based on wind energy has increased the importance of wind speed forecasting studies to ensure that these resources can respond to the needs in an uninterrupted and predictable manner. In order to be able to utility from wind energy technologically, it is very important to knowing the facilities of utilization, determining the regions, which have high potential of wind energy, being predictable the wind characteristics and speeds. The reliable and high accuracy wind speed forecasting is of vital to the efficient exploitation and utilization of wind power. The non-stationary and stochastic structure of the wind speed raise to the forefront the decomposition methods in wind speed forecasting. In this study, the effect of empirical mode decomposition, ensemble empirical mode decomposition and empirical wavelet transform on the performance of wind speed forecasting model obtained with long-short term memory from deep learning methods is investigated. The data collected from five wind farms in Marmara region, Turkey are decomposed to subband signal by these three decomposition methods, and the combined wind speed forecasting model is obtained with the long-short-term memory model structure. The performance of the combined models obtained by each decomposition method has been evaluated according to the statistical error criteria, and the decomposition method that is the highest effective to performance of wind speed forecasting model is suggested for the studies of obtaining the hybrid wind speed forecasting model. Son on yılda, rüzgâr enerjisine dayalı yenilenebilir enerji kaynaklarının kullanımındaki kayda değer artış, bu kaynakların ihtiyaçlara kesintisiz ve tahmin edilebilir bir şekilde cevap verebilmesini sağlamak için rüzgâr hızı tahmin çalışmalarının önemini arttırmaktadır. Rüzgâr enerjisinden teknolojik olarak faydalanmak için; yararlanma imkânlarının bilinmesi, yüksek rüzgâr enerjisi potansiyeline sahip bölgelerin belirlenmesi, rüzgâr karakteristiklerinin ve hızlarının tahmin edilebilir olması oldukça önemlidir. Güvenilir ve yüksek hassasiyetli rüzgâr hızı tahmini, rüzgâr gücünün verimli kullanımı ve kullanılması açısından hayati önem arz etmektedir. Rüzgâr hızının durağan olmaması ve stokastik yapısı, rüzgâr hızı tahmininde ayrıştırma yöntemlerini ön plana çıkarmaktadır. Bu çalışmada, ayrıştırma yöntemlerinden ampirik kip ayrışımı, topluluk ampirik kip ayrışımı ve ampirik dalgacık dönüşümünün derin öğrenme yöntemlerinden uzun-kısa süreli bellek ile elde edilen rüzgar hızı tahmin modeli başarımına etkisi incelenmektedir. Türkiye'nin Marmara bölgesindeki üç rüzgâr istasyonundan toplanan veriler her bir ayrıştırma yöntemi ile alt bant sinyallerine ayrıştırılarak uzun-kısa süreli bellek model yapısı ile kombine rüzgâr hızı tahmin modeli oluşturulmaktadır. Her bir ayrıştırma yöntemi ile birlikte elde edilen kombine modellerin başarımları istatistiksel hata ölçütlerine göre değerlendirilmekte ve rüzgâr hızı tahmin modeli başarımına etkisi en yüksek ayrıştırma yöntemi, melez rüzgâr hızı tahmin modeli elde edilmesi çalışmalarında önerilmektedir.

Abstract The price forecasting of the digital currencies in the financial market is of great importance, especially after the recent global economic crises. Due to the nonlinear dynamics, which is including inherent fractality and chaoticity of the digital currencies, it is understood from the research conducted by many researchers that a single model is not sufficient in forecasting the digital currencies with very high accuracy. Since the single models used in the forecasting of digital currencies have weaknesses as well as their own strengths, they might not grant the best forecasting achievement in all situations for all the time. A new hybrid-forecasting framework has been proposed in digital currency time-series to minimize this negative situation and increase forecasting achievement. In this study, a novel hybrid forecasting model based on long short-term memory (LSTM) neural network and empirical wavelet transform (EWT) decomposition along with cuckoo search (CS) algorithm is developed for digital currency time series. The model is obtained by combining the LSTM neural network and EWT decomposition technique, and optimizing the intrinsic mode function (IMF) estimated outputs with CS. The price of the four most traded digital currencies such as BTC, XRP, DASH and LTC, is estimated by the proposed model and the performance of the model has been tested. The experimental results show that the hybrid model proposed for digital currency forecasting can capture nonlinear properties of digital currency time series.

In this study, Bitcoin prediction is performed with Linear Regression (LR) and Support Vector Machine (SVM) from machine learning methods by using time series consisting of daily Bitcoin closing prices between 2012–2018. The prediction model with include the least error is obtained by testing with different parameter combinations such as SVM with including linear and polynomial kernel functions. Filters with different weight coefficients are used for different window lengths. For different window lengths, Bitcoin price prediction is made using filters with different weight coefficients. 10-fold cross-validation method in training phase is used in order to construct a model with high performance independent of the data set. The performance of the obtained model is measured by means of statistical indicators such as Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Pearson Correlation. It is seen that the price prediction performance of the proposed SVM model for Bitcoin data set is higher than that of the LR model.