Table 4.
Run times and overall prediction errors of three configurations (number of neurons of the neural networks) of the forecasting models. Only one preprocessing algorithm of the hybrid models is presented that provided the lower errors because the differences in performance between EMD and CEEMD were very small. The best performances are shown in boldface type.
| Model | No. of Hidden Neurons | Elapsed Training Time (s) | Overall MAPE (%) |
|---|---|---|---|
| 128 | 139 | 5.43 | |
| LSTM | 256 | 202 | 5.29 |
| 512 | 268 | 5.37 | |
| 32 | 43 | 5.36 | |
| GRU | 128 | 98 | 5.11 |
| 256 | 171 | 5.20 | |
| 100 per IMF | 1251 | 3.87 | |
| CEEMD–LSTM | 1000 per IMF | 3033 | 4.09 |
| 1000-100 decay | 2197 | 3.51 | |
| 100 per IMF | 915 | 3.88 | |
| EMD–GRU | 1000 per IMF | 2623 | 4.19 |
| 1000-100 decay | 1894 | 3.68 |