Session: 04-22-01: Machine Learning-Based Modeling, Prediction, and Optimization of Advanced Manufacturing and Materials System with Multiphysical Phenomena

Paper Number: 165208

Advanced Time Series Forecasting for Extrusion Process Optimization: A Hybrid Sarima and Machine Learning Approach 

Accurate forecasting of Melt Temperature and Melt Pressure is essential in industrial extrusion processes to ensure product quality and operational efficiency. This study investigates various time series forecasting techniques, including Seasonal AutoRegressive Integrated Moving Average (SARIMA), Random Forest (RF), and XGBoost, to forecast these critical process parameters six-time steps (one minute) ahead. Extrusion processes exhibit significant lagged dependencies among process variables, necessitating the use of cross-correlation function (CCF) analysis to properly align time-series data. A naïve persistence model is initially implemented as a baseline, providing a simple reference for evaluating model performance. Additionally, Spectral Fourier Analysis is applied to detect and characterize periodic components within the time series data, ensuring a better understanding of inherent seasonal patterns. The SARIMA models for Melt Temperature and Melt Pressure are developed through an iterative optimization process. Model parameters are selected based on Autocorrelation Function (ACF) and Partial Autocorrelation Function (PACF) analyses, complemented by an Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) grid search. The optimal SARIMA configuration for Melt Temperature is determined to be SARIMA(2,1,2) × (0,0,0,38), confirming a strong seasonal component, whereas Melt Pressure followed a simpler SARIMA(1,0,0) × (0,0,0,0) structure, indicating minimal seasonal influence. To enhance predictive performance, machine learning models, including Random Forest and XGBoost, are trained using a rolling window approach with a window size of 60. This strategy is employed to capture complex nonlinear dependencies between variables. A five-fold expanding cross-validation method is used to ensure model robustness and mitigate overfitting. The models are tested on a dataset comprising 36 test points, allowing for a comprehensive evaluation of forecasting accuracy. Performance comparisons revealed that Random Forest outperformed both SARIMA and XGBoost models in terms of predictive accuracy. RF demonstrated superior capability in capturing nonlinear trends inherent in the extrusion process. However, despite its effectiveness, RF lacks built-in temporal awareness, which can limit its generalizability. Conversely, SARIMA effectively captures linear dependencies but struggles with nonlinear relationships, leading to suboptimal forecasting performance. To address the limitations of individual models, a hybrid SARIMA+RF model is developed. This hybrid approach involved extracting SARIMA residuals, modeling them using RF, and integrating the resulting predictions with SARIMA’s baseline forecasts. This combination leveraged the strengths of both statistical and machine learning methodologies, allowing for enhanced forecasting accuracy. The SARIMA+RF hybrid model achieved the lowest forecasting errors, demonstrating significant improvements in both Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) over standalone models. The results underscore the benefits of integrating statistical and machine learning models to improve forecasting accuracy in industrial time series applications. The proposed SARIMA+RF hybrid model offers a scalable, adaptable, and interpretable solution for real-time process monitoring, making it an effective tool for optimizing industrial operations. Future research could explore further refinements to the hybrid approach, including deep learning methodologies such as Long Short-Term Memory (LSTM) networks, and the integration of additional exogenous variables to enhance predictive performance.

Presenting Author: Mohammad Akram University of New Haven

Presenting Author Biography: Mohammad Basit Akram is a Ph.D. candidate in Mechanical Engineering at the University of New Haven, specializing in machine learning applications in materials science and manufacturing. With an M.S. in Mechanical Engineering from Lehigh University and a B.E. in Power Engineering from Jadavpur University, he has cultivated expertise in statistical modeling, data analytics, and predictive modeling. His research focuses on deep learning-driven optimization of industrial processes, polymer extrusion forecasting, and molecular dynamics simulations. Basit has also presented at leading conferences, including APS Physics and ASME-IMECE. He is also a Six Sigma Green Belt-certified professional, demonstrating his proficiency in process optimization and operational excellence.

Authors:

Mohammad Akram University of New Haven
Ganesh Balasubramanian University of New Haven