An Ensemble Framework for Daily Carbon Emissions Forecasting Based on an Ireformer Modeland Multimodel Optimization Piecewise Error Correction

Abstract

The effective prediction of carbon emissions is crucial to enable governments to formulate energy saving and emission reduction policies. Since carbon emission data are complex, volatile and contain a large amount of noise, in this paper, a novel ensemble framework is proposed to address these challenges. First, the framework proposes an iReformer model with a Mish activation function and an Adam weight decay (AdamW) optimizer to make initial predictions on the original data. Second, for the error sequences generated by the prediction, decomposition is performed using time-varying filter empirical mode decomposition (TVF-EMD), and the quality of the decomposition is improved by parameter optimization of the key parameters of TVF-EMD, bandwidth thresholding and B-spline curves using the human evolutionary optimization algorithm (HEOA). The decomposed sequence is analyzed in the frequency domain by a Fourier transform and reconstructed into high-frequency, mid-frequency and low-frequency subsequences. Since the error sequence itself contains considerable noise the decomposition process may not be able to remove the noise completely, the reconstructed subsequence is further denoised using a wavelet soft threshold denoising technique to make the sequence more regular. Finally, the decomposed and reconstructed sequences are used as inputs to the validation set error sequences, the optimal model is selected using the principle of multimodel prediction optimization to obtain the error sequence predictions, and the proposed framework predictions are obtained from the preliminary predictions and the error sequence predictions. The ensemble framework is called iReformer-multimodel error correction. In this study, the proposed framework is tested on three datasets with different complexities, and by evaluating different performance metrics, the proposed framework is demonstrated to have superior prediction accuracy and robustness.

Keywords: iReformer model, Ensemble framework, Error correction, Carbon emissions forecasting