A new hybridization filtering-based linear-nonlinear models for time series forecasting
The prediction is one of the most influential factors in management and efficient utilization in various sciences as well as economic planning. Since there is a direct relationship between the accuracy of predictions and the quality of the decisions made, today, despite the numerous prediction methods and the achievement of accurate predictions, most researchers still try to combine different methods in order to obtain more accurate results. In order to improve the accuracy of predictions, the present study introduces a new parallel hybrid methodology based on trend-residual data preprocessing to time series predict. In the proposed model, different patterns and structures of trend and residual as well as linear and nonlinear, are simultaneously modeled. In the first stage of the proposed method, the data is analyzed by the Kalman filter (KF) method and divided into two groups of trend and residual patterns. Then, the trend patterns from the previous step, with the original data, are simultaneously considered as input of the autoregressive integrated moving average with explanatory variable (ARIMAX) and multilayer perceptron (MLP), as linear and nonlinear forecasting models, respectively. Then, this step is repeated for residual patterns. In this way, the proposed model can model four types of patterns, including linear-trend, linear-residual, nonlinear-trend, and nonlinear-residual. Finally, the results of these patterns, along with the Kalman trend, are combined together in a parallel hybridization process, and final forecasts of the proposed model are obtained. Numerical results of wind power forecasting indicate that the proposed model can approximately improve 68.44% and 56.80% the performance of its single linear and nonlinear components, respectively. Furthermore, the proposed model can yield more accurate results than traditional series-based components combination hybrid models, parallel-based components combination hybrid models, preprocessing-based linear models, and preprocessing-based nonlinear models by the same components. The proposed method can roughly improve 17.80%, 40.93%, 56.27%, and 23.47% the performance of these hybrid models, respectively. Therefore, the proposed model may be a suitable alternative for single as well as hybrid models for wind power forecasting, especially when more accurate results and/or more quality decisions are required.