Morgan Kelley, University of Texas
Deregulation and increases in electricity generation from wind turbines and solar photovoltaics have transformed U.S. electricity markets. The presence of these renewables has increased variability and uncertainty on the supply side of the grid. Coupled with significant variability on the demand side, this has led to an increasingly challenging environment for balancing the two. Managing demand rather than generation, referred to as demand response (DR), has emerged as an attractive approach for mitigating grid imbalance. Electricity-intensive processes are promising industrial DR candidates. Production can be increased during off-peak hours and excess product stored for use during peak demand when the production rate is lower. Grid-balancing benefits notwithstanding, DR participation has the potential to significantly cut operating costs for the industrial entities due to the inherent fluctuations in electricity prices generated by a mismatch between power supply and demand.
The industry-side benefits of DR increase when multi-day scheduling horizons are considered. Longer time horizons allow the plant to optimize the use of its storage capacity and deploy stored products at times of peak electricity demand. However, considering a longer time horizon presents some disadvantages, notably inaccuracies present in forecasts of electricity prices and product demand over an extended period of time. To account for this uncertainty and generate a robust DR operating schedule, we implement a chance-constrained optimization framework. Chance constraints effectively restrict the feasibility region, thereby increasing the confidence level of the solution.
We represent the applicability of chance constraints to solving DR scheduling problems under uncertainty by applying them to an air separation unit (ASU) capable of producing 50 tons of nitrogen per day. The problems solved are large-scale mixed integer linear programs (MILPs) with around 65,000 variables. Multiple threads were used in parallelizing the program for a max solution time of 15.48 minutes, plus or minus 0.21 minutes.
Abstract Author(s): Morgan Kelley, Ross Baldick, Michael Baldea