Neural networks for forecasting irregular demand in an Automotive Diagnostic Centre
محورهای موضوعی : Design of Experiment
Sonia Isabel Polo Triana
1
(Industrial Engineering Program, Faculty of Engineering, Universidad de Investigación y Desarrollo, Bucaramanga, Colombia)
Juan Camilo Gutierrez
2
(Industrial Engineering Program, Faculty of Engineering, Universidad de Investigación y Desarrollo, Bucaramanga, Colombia)
کلید واژه: demand forecasting, neural networks, Automotive Diagnostic Centre, demand variability, model optimization.,
چکیده مقاله :
This study investigates the prediction of demand in an Automotive Diagnostic Centre (ADC) in Bucaramanga, Colombia, applying advanced neural network techniques to three services: technical inspections for motorcycles, private cars, and public service vehicles. Utilising daily data from December 2019 to December 2021 (628 observations), the squared coefficient of variation of demand (CV²) and the average demand interval (ADI) are employed to classify demand, following the methodology of Syntetos & Boylan (2005b). Five Deep Learning models were evaluated: RNN, Bidirectional LSTM, CNN, GRU, and MLP, adopting a recursive approach for prediction. The findings reveal that despite the promises of forecasting algorithms and their success in other sectors, their performance in the context of ADCs is limited. Evaluation metrics, including RMSE, MSE, MAE, r2_score, and MAPE, reveal that although the LSTM model exhibits the best overall performance, no model achieves precise and reliable prediction due to the complexity and irregularity of demand. This finding underscores the need for continued research in demand forecasting and suggests exploring hybrid approaches or more specialised models. The variability in performance across models and services reflects the importance of tailoring predictive approaches to the specific characteristics of each demand segment. The principal contribution of this study is the innovative approach in applying neural networks to irregular demand in ADCs, an area hitherto little explored. It highlights the need to integrate external variables and develop adaptive management practices. Future research should focus on the integration of external factors and the development of hybrid models and mitigation strategies for efficient management of demand uncertainty.
This study investigates the prediction of demand in an Automotive Diagnostic Centre (ADC) in Bucaramanga, Colombia, applying advanced neural network techniques to three services: technical inspections for motorcycles, private cars, and public service vehicles. Utilising daily data from December 2019 to December 2021 (628 observations), the squared coefficient of variation of demand (CV²) and the average demand interval (ADI) are employed to classify demand, following the methodology of Syntetos & Boylan (2005b). Five Deep Learning models were evaluated: RNN, Bidirectional LSTM, CNN, GRU, and MLP, adopting a recursive approach for prediction. The findings reveal that despite the promises of forecasting algorithms and their success in other sectors, their performance in the context of ADCs is limited. Evaluation metrics, including RMSE, MSE, MAE, r2_score, and MAPE, reveal that although the LSTM model exhibits the best overall performance, no model achieves precise and reliable prediction due to the complexity and irregularity of demand. This finding underscores the need for continued research in demand forecasting and suggests exploring hybrid approaches or more specialised models. The variability in performance across models and services reflects the importance of tailoring predictive approaches to the specific characteristics of each demand segment. The principal contribution of this study is the innovative approach in applying neural networks to irregular demand in ADCs, an area hitherto little explored. It highlights the need to integrate external variables and develop adaptive management practices. Future research should focus on the integration of external factors and the development of hybrid models and mitigation strategies for efficient management of demand uncertainty.
