Assessing the Impact of Temporal Resolution Using BSM1 on the Performance of Machine Learning

Abstract

Objectives : This study aims to establish efficient strategies for data-driven operational management by examining the variations in machine learning modeling outcomes and data characteristics based on data acquisition intervals and methods.Methods : The BSM1 was used to simulate wastewater treatment facilities and to generate influent and effluent water quality data at 15-minute intervals. The generated data was processed by volume reduction through down sampling and data characteristic observation via resampling techniques, including up sampling through interpolation. Subsequently, the study involved a comparative analysis of the performance of 30 machine learning models built with the down sampled data.Results and Discussion : As data acquisition interval increased (i.e., down sampling progressed), <i>R</i>² decreased and RMSE increased. When using the mean value as a representation, data accuracy was high, and error loss was minimal. Utilizing the maximum value as a representation helped maintain data characteristics and reduce information loss. Simple interpolation methods did not yield improved data accuracy. Furthermore, with wider data acquisition intervals, the practical predictive performance of machine learning models decreased, and the models experienced a sharp decline in performance when data became insufficient.Conclusion : For models requiring the ability to detect changes rather than accuracy, utilizing the maximum value over a specific period proves to be effective. The measurement interval of data emerges as a significant factor affecting the performance of machine learning models, with models developed under different measurement intervals often failing to demonstrate the expected performance. In this study, we have implemented all stages of data preprocessing, classification, training, and validation using LabVIEW, confirming the potential for integrating data analysis processes into LabVIEW, a widely used platform in the fields of control and measurement.