Methods of filtering and regression for forecasting noisy timeseries based on machine learning

Abstract

Predicting parameters in industrial processes is significantly complicated by the presence of noise in sequential measurements, which reduces the effectiveness of technological process control. The aim of the research is to develop an integrated model that combines adaptive noise filtration methods and regression to improve the accuracy of forecasting noisy time series using machine learning algorithms. During the research, a comprehensive database of time series with various levels and types of noise was created, providing a thorough verification of the effectiveness of the proposed methods. The datasets were developed considering the specifics of technological processes and the diversity of noise patterns, which allowed for an accurate evaluation of the developed methods under different conditions. As part of the development of adaptive noise filtration methods, the Kalman filter and wavelet filtration were implemented and optimized. The relationship between the effectiveness of filtration methods and temporal patterns was established: for rapidly changing parameters, wavelet filtration provides higher smoothing efficiency, whereas the Kalman filter better preserves signal characteristics for more stable sequences. To solve the time series forecasting problem, two regression algorithms were implemented and tested – Support Vector Regression and Multilayer Perceptron. It was proven that Support Vector Regression demonstrates better results with low-noise data, while Multilayer Perceptron shows higher stability under significant noise conditions, especially after preliminary filtration. To evaluate the effectiveness of the proposed solutions, a comprehensive quality assessment system was developed that simultaneously considers forecasting efficiency, temporal aspects, noise characteristics, and computational complexity. Experimental confirmation demonstrates that the developed approach improves forecasting accuracy compared to machine learning methods without preliminary filtration, while maintaining acceptable computational complexity. The developed approach is promising for industrial applications, including modeling iron ore enrichment processes, where noise-resistant forecasting is important for process control. The proposed methods can be extended to various industrial processes with similar temporal data and noise characteristics, especially in metallurgical, chemical, and food industries.