Partitioning the Data Domain of Combinatorial Problems for Sequential Optimization

Abstract

Abstract-Following the long-term goal of substituting conventional power generation with cleaner energy will lead to an integration of a large share of small energy generation units imposing large problem sizes for coordination. The expected huge number of entities leads to a need for new techniques reducing the computational effort for coordination. Predictive scheduling is a frequent task in energy grid control. For a number of energy resources, schedules have to be found that fulfill several objectives at the same time. Considering day-ahead scenarios with 96-dimensional schedules imposes additional challenges to this already hard combinatorial problem. We explore the effects of reducing complexity by partitioning the data domain of the optimization problem for a sequential approach that integrates energy models for constraint handling directly into the optimization process. We explore the effects of different partitioning schemes and evaluate the trade-off between accuracy and effort with several simulation studies.

Towards fully Decentralized Multi-Objective Energy Scheduling

“…For our evaluation we simulated different virtual power plants consisting of different co-generation plants. The model has already been used and evaluated in different projects, e. g. [13], [46], [50], [51]. We started with a rather small setting of four agents and 96-dimensional schedules resulting in a 384-dimensional search space which has already been evaluated to be highly multi-modal and ragged [52].…”

“…For the single objective case several solutions exist. In [13] an example for a centralized approach can be found, examples for decentralized approaches are given in [14]- [16]. A centralized multi-objective variant based on parallel tempering that harnesses a decoder extension to co-encode different key performance indicators can be found in [17].…”