3rd ICAI 2024

International Conference on Automotive Industry 2024

Mladá Boleslav, Czech Republic

chosen as the optimisation criterion in 10% of the models and RMSE in 10%. Similar conclusions were reached in the analysis of (Rodrigues et al., 2023), which showed the same relationship, i.e. MAPE was optimised for 65,8 % of the studies and MAE for 44% as a secondary metric. Despite the critical opinions on the use of the MAPE indicator highlighted by (Davydenko & Fildes, 2016) , considering the widespread usage of the STLF problem by researchers and the lack of model prediction negative values, it has been decided to use the MAPE as the basis for model optimisation in the solution presented in next paper section. As the secondary metric MAE was chosen. The formulae for the calculation of the chosen metrics are shown below.

3. Problem Solution The development of the active power prediction model in this paper was based on commercial and widely available solutions for designing applications based on machine learning methods. In the research, the ‘Pycaret’ library and the statistical analysis software ‘Statistica’ were used. Modelling was performed on a PC using GPU calculations. The data used for modelling were taken from the energy meters stored in the SCADA system. Detailed data on the operations of each group of machines and their load/utilisation was developed using historical data from the MES class system. Data of recorded orders start times; end times and the content of operations were used. To ensure high quality data for the modelling, periods with evident errors were excluded from the data set. The following modelling steps did not define the need for data cleaning, so only a few periods were removed based on expert assessments. The data preparation step was followed with a coarse model selection without extensive optimisation of the hyperparameters. Table 2 presents a summary of the top five ML models for load prediction for two power sources. It is evident from the comparison that the best model is the ETR model, which achieves the best results for both power source i.e. MAPE – 0.147(SOURCE_1), 0.146(SOURCE_2) and MAE – 168.5(SOURCE_1), 109.8(SOURCE_2) respectively.

87