Evolutionary-based neurofuzzy model with wavelet decomposition for global horizontal irradiance medium-term prediction
This study investigates the influence of wavelet decomposition on standalone adaptive neurofuzzy inference system (ANFIS) and its hybrid with evolutionary algorithms [genetic algorithm (GA) and particle swarm optimization (PSO)] at a medium-term temporal scale. Time series data of global horizontal...
Gespeichert in:
Veröffentlicht in: | Journal of ambient intelligence and humanized computing 2023-08, Vol.14 (8), p.9793-9805 |
---|---|
Hauptverfasser: | , , , |
Format: | Artikel |
Sprache: | eng |
Schlagworte: | |
Online-Zugang: | Volltext |
Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
Zusammenfassung: | This study investigates the influence of wavelet decomposition on standalone adaptive neurofuzzy inference system (ANFIS) and its hybrid with evolutionary algorithms [genetic algorithm (GA) and particle swarm optimization (PSO)] at a medium-term temporal scale. Time series data of global horizontal irradiance used as inputs was decomposed according to 5-level wavelet sub-signals and modeled independently. Statistical metrics were used to evaluate the models, and these results were compared for all models. From our findings, the wavelet-based models incurred more computational time and lesser accuracy compared to the ANFIS and its hybrid models. ANFIS-GA and its wavelet hybrid had a high computational time of 7 s and 15.2 s, respectively., Among the six models, the standalone ANFIS model outperformed its hybrids with a root mean square error (RMSE) of 13.07, normalized root mean square error (NRMSE) of 0.11%, mean absolute percentage error (MAPE) of 4.57%, variance accounted for (VAF) of 92.03%, relative coefficient of variation (RCoV) of 0.24, and computational time (CT) of 3 s. The integration of wavelet decomposition reduced the performance of the standalone ANFIS and its hybrids with wavelet-based PSO-ANFIS recording the least model accuracy (RMSE = 31.02, NRMSE = 0.19, MAPE = 15.31, VAF = 55.89, RCoV = 0.21, CT = 8.4 s). Overall, standalone machine learning models could be more efficient, computationally less intensive, and computationally less costly than their hybrids if their hyperparameters are appropriately tuned. |
---|---|
ISSN: | 1868-5137 1868-5145 |
DOI: | 10.1007/s12652-021-03639-2 |