A fuzzy series‐parallel preprocessing (FSPP) based hybrid model for wind forecasting

Abstract

Wind power is one of the most important renewable energy sources that is widely used in many developed and developing countries. However, it is generally stated in the literature that providing accurate forecasts for large-scale planning purposes is not a simple task, especially by single models. It is the main reason for this fact that why researchers in recent years have sought to propose hybrid models for increasing the accuracy of predictions. In general, choosing the appropriate type and the number of components, as well as the proper type of hybridization methodology, are the most effective factors in the performance of the developed hybrid models. Although in the literature, numerous attempts have been made in order to answer these questions, there is no general consensus on this matter. For this reason, the main idea of this paper is to concurrently combine different hybrid methodologies as well as different single models in order to benefit from the advantages of these models and methodologies, simultaneously. In this way, three well-known and widely used hybrid methodologies, including the preprocessing, the series, and the parallel methodologies, are combined together by incorporating the linear/nonlinear and certain/uncertain components. In addition, in the proposed model, a new process is proposed based on the complex/uncertain modelling to model the preprocessing phase residuals, which have been ignored in the modelling procedures. In this way, in the first stage of the proposed model, the data is preprocessed by the Kalman filter as a preprocessing approach in order to divide data into two groups of trend and residual patterns. The trend data provided in the previous step, with the original data, are simultaneously considered input data of an autoregressive integrated moving average as the certain linear model and a multilayer perceptron as the certain nonlinear model for certain linear and nonlinear modeling of patterns. This step is repeated for the residual data by the series hybridization of models in the previous stage by the fuzzy models for the uncertain linear and nonlinear modelIng of patterns. Finally, each component's weight is optimally calculated by the least square algorithm, and then the results are combined together in a parallel process. Empirical results of two benchmarks of wind domain indicate that the proposed method has averagely improved the performance of its component used separately, parallel-based hybrid models, and series-based hybrid models 46%, 22%, and 19%, respectively, for predicting wind power time series.