3rd ICAI 2024

International Conference on Automotive Industry 2024

Mladá Boleslav, Czech Republic

Table 2: Comparison of the top 5 ML models – shift average load forecasting

AVG_POWER_SOURCE_1

AVG_POWER_SOURCE_2

MAPE MAE MSE R 2

MAPE

Model

MAE MSE R 2

53840 0.862

23330 0.804

Extra Trees Regressor

168.5

0.147 109.8

0.146

Random Forest Regressor

171.2 57758 0.853 0.153 114.8 22937 0.818 0.152

CatBoost Regressor 171.3 49788 0.871 0.153 110.4 21562 0.846 0.159 Extreme Gradient Boosting 183.7 63371 0.841 0.159 112.0 22801 0.820 0.158 Gradient Boosting Regressor 176.8 59917 0.849 0.162 119.2 26645 0.785 0.165 Source: Own elaboration After selecting Extra Trees Regressor as the target model, the optimisation of the hyperparameters was undertaken. The detailed data of hyperparameters is not the subject of the study, for which reason the hyperparameters of the model will not be published in this publication. From the analysis of the model training data shown in figure 5, it can be seen that the learning curve of the validation set, is significantly below that of the training set. This indicates that the data in the validation set are underrepresented in relation to the training set. In the case of this study, this phenomenon is not an error, but the result of a sampling scheme selected from an inhomogeneous dataset, with the aforementioned dynamic character. The dynamics in time, aggregated to days of the week, can be observed in Figure 6(where 1 – Monday, 7 – Sunday), which shows the profiles of power use on individual days of the week for both power sources. The analysis of the graph allows to conclude that on days 1 and 7 there is, on average, a lower energy demand, which is due to the discontinuity of selected operations carried out in the company where the research was carried out.

88