Year: 2024
Type: article
Abstract: Designing intelligent models for predicting time series have always been a hot topic in AI research fields.The existing work usually follows a "preprocessing-predicting" flow to tackle this issue.In these two procedures, the researchers consider neither the best way to handle preprocessed subseries nor the most efficient artificial neural network.To solve these problems, I have studied the existing models and preprocessing methods, then come up with a novel model called FD3 and have tested its validity on carbon emissions datasets.In my approach, three "FD" procedures are executed: 1) frequency decomposition achieved by using complementary ensemble empirical mode decomposition with adaptive noise (CEEMDAN), an improved version of the famous empirical mode decomposition (EMD); 2) forecasting dendritic neuron model (DNM) that has proved validity on numerous prediction tasks, showing advanced nonlinear fitting ability than traditional network-structured ANNs; and 3) fluctuation density measurement (FD function) that used to regulate the predicting strategy for each decomposed subseries.In experiments, the FD3 framework has shown better performance than seven baseline models in terms of three widely employed time series prediction evaluation metrics.The success of FD3 has confirmed the validity of "preprocessing-predicting" workflow and provides better solutions for carbon emissions prediction.Furthermore, the design of FD function can give more insights for signal analysis that the selection of decomposed subseries can have huge impacts on the original data.
Author: 何, 厚佃
Language: Japanese
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
SDG: Climate action
Open Access status: green
OpenAlex ID: W7053753863
