Tek değişkenli zaman serileri tahmini için öznitelik tabanlı hibrit ARIMA-YSA modeli

Abstract

Highlights:Graphical/Tabular Abstract  A feature-based hybrid ARIMA-ANN method is proposed for time series forecasting.  Our new hybrid method calculates features of a given time series and selects most important ones.  The method achieved better forecasting accuracy than the examined individual and hybrid methods.The architecture of the proposed feature-based hybrid method is indicated in Figure 1. Figure 1 Proposed feature-based hybrid methodPurpose: In this study, it is aimed to improve the high-accuracy hybrid ARIMA-ANN model for univariate time series forecasting by converting it to a feature-based model. Theory and Methods:The proposed method firstly obtains a linear component from a given univariate time series by using moving average (MA) filter. Then, nonlinear residual series, which is the difference between given time series and linear component, is achieved. After that, a set of classical and advanced features is extracted from the observed data and their importance scores are computed by using gradient boosting trees. The most informative features are selected and used in the hybrid method as explanatory variables which lead to improve time series forecasting accuracy results. Results:The performance of the proposed feature-based hybrid method is evaluated on four different time series each which come from different fields and have different statistical characteristics. The proposed method improves the absolute accuracy of the hybrid method that we built on it, between 5.5% and 11.8% in the evaluated datasets. Conclusion:The univariate time series forecasting methods can use only values in the prior time steps to predict next value. They do not take account of underlying structural characteristics of past values which might provide a more meaningful summarized information that lead to yield better approximations in the forecasting. The developed method, extracts features which basically measure the characteristics of the given series. After that, the most explanatory features are selected and used in the hybrid method for improving forecasting performance results. Our experimental results indicate that our feature-based hybrid method gives superior accuracy as compared to well-known existing methods.High prediction accuracies at time series modeling and forecasting is of the utmost importance for a variety of application domains. Many methods have been proposed in the literature to improve time series forecasting accuracy. Those which focus on univariate time series forecasting methods use only the values in the prior time steps to predict the next value. In this study in addition to the historical values, it is aimed to increase the forecasting performance by using extra statistical and structural features which summarize characteristics of the time series. Feature importance scores are determined by gradient boosting trees (GBT).Features with the highest importance score are given as explanatory additional variable to the hybrid ARIMA-ANN model. The evaluation of the developed method is performe...

Solar Power Prediction with an Hour-based Ensemble Machine Learning Method

“…Time series models aim to predict future data from past observed data using statistical and machine learning techniques. In literature, there are various time series methods [8,9,10,11], some of them are directed towards for forecasting solar power [12,13,14,15,16,17,18]. Solar time series data includes high volatility and nonlinearity.…”