(383q) From the Lab to Production: Control-Informed Machine Learning for Reliable Industrial Forecasting

Data-driven modeling techniques such as Neural Networks show promise in their ability to extract information from data and perform symbolic regression. Chemical Engineers are not strangers to machine learning and neural network models; However, implementation was initially hampered by conceptual, technological, and organizational challenges [1]. Advances in training algorithms, computational parallelization, and success of machine learning models in areas such as consumer interest and large language models have increased organizational trust in data-driven methods. Applications to industrial systems have faced additional challenges due to significant safety and economic consequences model errors may cause. In addition, non-stationary environments and opaque model interpretability add additional complexity to model creation and operation respectively.

We present work using techniques common to controls and process systems engineering such as closure modeling, noise models, and Bayesian parameter estimation to help-enable trained neural network models to better accommodate the complexity of realistic implementation. Key technical contributions include: (1) embedding known physical patterns (such as diurnal cycles) and persistence assumptions into solar energy forecasting models, achieving improved accuracy over purely data-driven approaches; (2) incorporating control engineering techniques and noise modeling frameworks to enhance machine learning prediction performance; and (3) implementing Extended Kalman Filter methodologies to enable real-time neural network model adaptation in changing operational environments.

Research Interests

I am interested in roles where I can deploy data-driven and hybrid modeling and control techniques to solve real operational challenges.

References

[1] Venkatasubramanian, V. (2019). The promise of artificial intelligence in chemical engineering: Is it here, finally? AIChE Journal, 65(2), 466–478. https://doi.org/10.1002/aic.16489