Optimize electricity considering installation costs and carbon emissions: A comparison between reinforcement learning and linear programming

Considering the trade-off between installation costs and carbon emissions, this research presents a novel framework to achieve effective energy optimization: (1) bi-objective linear programming (LP) is used to derive the optimal allocations of energy resources, (2) reinforcement learning (RL) is applied to conduct supply optimization and compare to static and dynamic LP, and (3) the impacts of varying the weight of installation cost, capacity expansion, and increasing installation costs are quantitatively assessed. Experimental results show that static LP based on one-shot data splitting performs the worst because it cannot catch up with the most recent cost parameters and capacity limitations. Across all scenarios, RL significantly performs the best in minimizing total objective cost, while dynamic LP is good at minimizing carbon emissions. To reduce carbon emissions, capacity expansion of solar power is the most effective, followed by capacity expansion of nuclear power, and capacity expansion of wind energy ranks third. Increasing installation costs of natural gas, wind energy, and solar power result in reduced consumption of these energies because they are substantially replaced by thermal coal. This research provides a quantitative basis for practitioners to conduct capacity planning and optimization for energy resources. • Static/dynamic linear programming (LP) and reinforcement learning (RL) are compared. • Installation costs and carbon emissions are incorporated to optimize electricity supply. • Capacity expansion of solar power is the most effective to reduce carbon emissions. • Capacity expansion of nuclear performs the best to minimize total objective cost. • RL is powerful in minimizing total objective cost, while dynamic LP is good at minimizing carbon emissions.