4.1.4. Linear Regression

The linear regression (LR) method can be improved based on the choice of the attribute selection method. We tested the LR algorithm without any attribute selection method, as well as with the M5 and greedy attribute selection method. The results obtained are presented in Table 5 with little or no difference between the different selection methods. Using an attribute selection method achieved only about 0.764% reduction in the error rate over the use of the no-selection method. Thus, for power demand prediction purposes, it is sufficient to apply the LR method without any attribute selection method. This is expected since there exist only the day and time as the main input attributes for forecasting purposes, thus attribute selection introduces no significant performance advantage.

**Table 5.** Performance of different LR parameter settings.


#### 4.1.5. Random Forest

The random forest (RF) algorithm has a few parameters to be fine tuned, namely, the maximum depth and the number of iterations. In our study, a combination of three parameters were examined with progressively increasing values and the results obtained are presented in Table 6. We found that increasing the maximum depth and number of iterations barely resulted in 0.409% and 0.378% decreases in the RAE and RRSE error rates, respectively. This insignificant difference in the error rate implies that using low maximum depth values and number of iterations will be suitable in using the RF algorithm for power demand prediction purposes. It also may present faster computational time since fewer iterative steps are considered during the algorithmic process.

**Table 6.** Performance of different RF parameter settings.

