*5.2. Experimental Results*

According to above-mentioned relevant experimental parameters shown in Table 2 to Table 5 and experimental layouts, the visualized experimental results within half a year output by six PGPMs mentioned above are shown in Figure 6.

**Figure 6.** Prediction results of different algorithms. (**a**) PGPM based on SVR; (**b**) PGPM based on Decision Tree; (**c**) PGPM based on Random Forest; (**d**) PGPM based on LSTM; (**e**) PGPM based on Bi-LSTM; (**f**) Ours.

As shown in Figure 6, it can be found that the deviations between the true data and experimental results of PGPMs based on SVR, Decision Tree, and Random Forest were more obvious than that generated of PGPMs based on LSTM, Bi-LSTM, and Attention-Bi-LSTM. Summarily, the LSTM-based PGPMs are very suitable for power generation forecasting scenarios. However, according to Figure 6d–f, it can be seen that from the visualization point of view, the performance of Attention-Bi-LSTM PGPM proposed in this paper is basically the same as that of the other LSTM-based PGPMs. Therefore, to further illustrate the advantages of the proposed PGPM, this paper evaluates the performance of above-mentioned PGPMs from a quantitative perspective.

Besides, in the training procedure of the proposed PGPM, the model converges very quickly, as presented in Figure 7.

**Figure 7.** Convergence curve of the proposed PGPM.

As Figure 7 shows, the loss function of the model decreased quickly, and converged nearly to zero within the first 100 epochs, which means in the practical training procedure it could be finished very fast.
