The CEEMDAN-EWT-CNN-GRU-SVM model: A robust framework for decomposing non-stationary time series, extracting data features, and predicting solar radiation

Abstract

• Developing an advanced data processing technique for learning complex patterns in non-stationary time series of input variables. • Developing a new deep learning model for extracting data features. • Developing a novel SVM model for predicting (SR) data. Solar radiation (SR) prediction is needed for energy management, climate modeling, and agricultural planning. As a result, modelers have developed various models such as support vector machine (SVM) models, to address this need. However, support vector machines (SVMs) and other machine learning models cannot effectively extract data features and learn complex patterns in nonstationary input time series. Thus, our study proposes a new model to address these limitations. The new model is a combination of five algorithms and models, including Complete ensemble empirical mode decomposition adaptive noise: (CEEMDAN), the empirical wavelet transform (EWT), the convolutional neural network (CNN), the gated recurrent unit network (GRU), and the support vector machine (SVM) model and named CECGS. First, CEEMDAN and empirical wavelet transform decompose nonstationary input time series into subseries with simpler patterns. The CNN and GRU models then extract important features from these sub-series. Finally, the SVM model uses the extracted features to predict SR data in the Sefidrood basin, Iran. Our research employs various performance metrics to assess the effectiveness of the new model in predicting outcomes. These performance metrics include explained variance value, Kling Gupta efficiency (KGE), mean absolute error (MAE), and uncertainty at 95% (U95). The new model is compared with the other 34 models to evaluate its effectiveness and robustness in predicting SR data. The CECGS model enhanced the KGE, MAE, U95, and Evar of the eight groups of different models by 1–43%, 4.4%- 74%, 1–66%, and 1.01- 33%, respectively. The CECGS model also improved the testing KGE, MAE, U95, and Evar of the eight groups of different models by 1.01–41%, 3.2%- 74%, 1.17–64%, and 1.01- 33%, respectively. The CECGS model improved the R 2 values of the other models by 1.0- 7.03%. The KGE and MAE of the new model are 0.97 and.45. The results of this study indicate that the CECGS model is a reliable tool for predicting solar radiation data. Moreover, the model effectively addresses the main drawbacks of the SVM model. The new model also effectively processes time series with complex patterns. Using multiple components, the new model makes accurate predictions that can be effectively used for energy management. The accurate predictions of the CECGS model make it suitable for use in various fields. However, it should be taken into account that the training time of the new model is longer than that of the other models, which may be one of its disadvantages.

Topics & Concepts