(373ak) Learning Predictive Classifier Models of the Correlating Nature of Cost and Emissions Objectives in Moving-Horizon Demand Response

Demand response is an important process operations paradigm for enabling sustainability whereby electricity users modify their electricity use over time based on signals obtained by the power grid. Both the price and associated carbon dioxide emissions of purchasing power from the grid are known driving forces for intermittent demand response operation of flexible chemical processes [1]. These parameters exhibit temporal variability and have significant influence over cost and carbon emissions objectives [2-3]. As demand response decisions are typically made in a moving horizon fashion whereby the optimal scheduling problem is solved repeatedly over time, it is useful to understand how the nature of the cost and emissions signals to correlate (point towards similar optimal schedules) or compete (point towards different optimal schedules) also changes in time. Recent efforts in our group have developed methods to systematically identify the competing vs. correlating nature of objectives [4] and studied the correlating and competing nature of cost and emissions for chemical demand response [1]. However, to make use of this knowledge in online demand response scheduling, we require a model that can quickly predict when these objectives will be correlating or competing.

In contemporary process systems engineering, machine learning (ML) has attracted considerable attention and found widespread application [5]. Through the utilization of ML models, this work aims to build classifier models capable of predicting the correlating versus competing nature of the cost and emissions objectives and use this prediction to determine when to expend computational effort to find compromise solutions between the two. In this study, we leverage historical hourly data encompassing local marginal prices and corresponding carbon dioxide emissions of power generation from diverse geographic locales, including California, the UK, Tokyo, and so on, as features in the machine learning model. The objective partitions derived from prior work analyzing the demand response of an ammonia production process [1] serve as training data for the ML model. Subsequently, the trained model is applied to predict objective groupings for untrained datasets, thereby validating its accuracy and generalizability. Therefore, optimization can be applied in a real chemical engineering scheduling process since correlation versus competing nature of cost and emissions is determined by the obtained classifier model.

The original scheduling problem operates within 48-hour horizons, incorporating power price and carbon emission data. Consequently, each data point comprises 96 features. However, empirical observations indicate that the inclusion of all features adversely impacts the model's accuracy. To mitigate this issue, principal component analysis is employed to reduce feature dimensionality, thereby enhancing model performance. Additionally, various ML models including support vector machines and decision trees are explored to ascertain the optimal performance. To demonstrate the utility of the proposed method, we apply the trained classifier model to a year-long moving horizon simulation of ammonia production using power grid data not in the original training set. The classifier model is used to determine when we can optimize for cost and expect good emissions performance, versus when more computational effort needs to be taken to seek a compromise solution between the two goals. Results of the simulation demonstrate the benefits of this approach for achieving low-emissions operating schedules with only small increases in cost and computational time in comparison to a traditional cost-optimal process schedule.

[1] Wang, H. and Allman, A., 2024. Analysis of the correlating or competing nature of cost-driven and emissions-driven demand response. Computers & Chemical Engineering, 181, p.108520.

[2] Allman, A., Palys, M.J. and Daoutidis, P., 2019. Scheduling‐informed optimal design of systems with time‐varying operation: A wind‐powered ammonia case study. AIChE Journal, 65(7), p.e16434.

[3] Kelley, M.T., Baldick, R. and Baldea, M., 2018. Demand response operation of electricity-intensive chemical processes for reduced greenhouse gas emissions: application to an air separation unit. ACS sustainable chemistry & engineering, 7(2), pp.1909-1922.

[4] Russell, J.M., Allman, A., 2023. Sustainable decision making for chemical process systems via dimensionality reduction of many objective problems. AIChE Journal, 69(2), e17692

[5] Daoutidis, P., Lee, J.H., Rangarajan, S., Chiang, L., Gopaluni, B., Schweidtmann, A.M., Harjunkoski, I., Mercangöz, M., Mesbah, A., Boukouvala, F. and Lima, F.V., 2023. Machine learning in process systems engineering: Challenges and opportunities. Computers & Chemical Engineering, p.108523.