These results demonstrate that SW-SVR predicts complicated micrometeorological data with the best prediction performance and the lowest computational complexity compared with standard algorithms. In particular, we found that dynamic aggregation of models built from very little extracted data by D-SDC is effective for compatibility of high prediction performance and low computational complexity. However, there are problems to be solved in SW-SVR. Firstly, the prediction performance of SW-SVR sometimes deteriorates despite an increase of training data. In particular, this problem occurred under the conditions that prediction horizons are 6 h as shown in Fig. 3. This is because data extracted by D-SDC involves unnecessary training data for highly accurate prediction. If D-SDC extracts the same data as the extracted data when training periods are shorter, the prediction performance of SW-SVR never deteriorates due to an increase of training data. Therefore, we must review both feature mapping and algorithms of D-SDC so as to avoid extracting unnecessary training data. Meanwhile, SW-SVR is based on a combination of several algorithms: kernel approximation, PLS regression, k-means, D-SDC, and linear SVR. Moreover, each algorithm has several parameters. Therefore, SW-SVR has more varied parameters, and it takes more time to tune the parameters. In this experiment, we used a grid search roughly so as to decide the parameters in a certain time. However, there is still room for improvement in the prediction performance by using other approaches such as a genetic algorithm instead of a grid search (Huang & Wang, 2006).
