Advances in the Optimization of Energy Systems and Machine Learning Hyperparameters
Abstract
Intensifying public concern about climate change risks has accelerated the push for more tangible action in the transition toward low-carbon or carbon-neutral energy. Concurrently, the energy industry is also undergoing a digital transformation with the explosion in available data and computational power. To address these challenges, systematic decision-making strategies are necessary to analyze the vast array of technology options and information sources while navigating this energy transition. In this work, mathematical optimization is utilized to answer some of the outstanding issues around designing cleaner processes from resources such as natural gas and renewables, operating the logistics of these energy systems, and statistical modeling from data.
First, exploiting natural gas to produce lower emission liquid transportation fuels is investigated through an optimization-based process synthesis. This extends previous studies by incorporating chemical looping as an alternative syngas production method for the first time. Second, a similar process synthesis approach is implemented for the optimal design of a novel biomass-based process that coproduces ammonia and methanol, improving their production flexibility and profit margins.
Next, operational difficulties with solar and wind energies due to their temporal intermittency and uneven geographical distribution are tackled with a supply chain optimization model and a clustering decomposition algorithm. The former describes power generation through energy carriers (hydrogen-rich chemicals) connecting resource-dense rural areas to resource-deficient urban centers. Results show the potential of energy carriers for long-term storage. The latter is developed to identify the appropriate number of representative time periods for approximating an optimization problem with time series data, instead of using a full time horizon. This algorithm is applied to the simultaneous design and scheduling of a renewable power system with battery storage.
Finally, building machine learning models from data is commonly performed through k-fold cross-validation. From recasting this as a bilevel optimization, the exact solution to hyperparameter optimization is obtainable through parametric programming for machine learning models that are LP/QP. This extends previous results in statistics to a broader class of machine learning models.
Subject
process systems engineeringmathematical optimization
process synthesis
capacity expansion
unit commitment
design & scheduling
supply chain optimization
parametric programming
machine learning
hyperparameter tuning
natural gas
biomass
solar
wind
energy storage
battery
hydrogen
ammonia
methanol
liquid transportation fuels
chemical looping
Citation
Tso, William Weikang (2020). Advances in the Optimization of Energy Systems and Machine Learning Hyperparameters. Doctoral dissertation, Texas A&M University. Available electronically from https : / /hdl .handle .net /1969 .1 /191896.