Optimizing Microgrid Energy Scheduling - Comparing performance of Genetic algorithms and Mixed Integer Linear Programming

Abstract

This thesis deals with the comparison of Genetic algorithms and Mixed Integer Linear Programming on the microgrid energy scheduling problem. The focus lies on genetic algorithms and improvements to problems experienced by the GA with MILP being used as a baseline for comparison. The GA and MILP are being compared on the unit commitment / economic load dispatch problem within a microgrid setting. The objective is to investigate under which circumstances the individual optimizers are preferable. In the GA approach constraints were handled using penalty terms added in the fitness function encoding physical constraints such as battery storage limits. Additional penalties were used to enforce objectives such as temperature ranges and binary statuses of components. In the MILP approach constraints were handled in the model creation with objectives encoded as terms in the objective function scaled by individual weight factors. The research was carried out on a simulation tool where two scenarios were modeled. One scenario represents a microgrid in a remote setting without external grid connection where energy consumption minimization was of importance to prolong the useful life of components. The second scenario represents a grid connected microgrid where precise temperature management and energy cost minimization through strategic energy market interactions was the main objective. The results suggest that the MILP outperforms the GA on most of the problems considered with the exception of certain cases of uncertainty in the renewable energy generation. Results also show that improvements could be made to the genetic algorithm by seeding the initial population. Furthermore, the gentic algorithm performance and computation time improved on the mountain hut scenario when a timestep of 8-16 hours was used. This result is unexpected as a lower temporal resolution would make the problem harder to handle, however in the case of maximizing up time of components a larger timestep duration seems preferable.