High-performance simulation of nation-sized smart grids

In the multi-disciplinary field of smart grids, renewable energy generation technologies have been developed in commercial and residential scale to allow decentralised energy generation and to secure supply. Furthermore, devices and algorithms have been designed to autonomously act in demand response scenarios. Although these advances have been made, their actual impact on large-scale smart grids has not been examined yet. This, however, is tackled in this thesis by facilitating detailed, large-scale smart-grid simulations. The smart grid in this thesis is represented by a multi-agent system associated with a flow network to model the smart-grid structure and capacities. It also features multiple types of interchangeable resources. Each agent determines its optimal device operation over a time interval to meet its electrical and thermal energy demand at minimal costs by employing solar panels, fuel cells and power storages. Thus, the agents may also act as supplier and not only as consumer which allows decentralised energy generation. Moreover, an agent ensures that the temperature within the building does not exceed admissible bounds. An agent is simulated by solving a quadratic programme with linear complementarity constraints which takes the energy prices and the available devices for energy generation and storage into account. The energy prices are related to an abstract economy model in which they are determined with respect to demand and supply. When intending to simulate large-scale smart grids in detail, a single computer is no longer capable to deal with the complex and memory-consuming problem. Therefore, an algorithm is developed to allow distributed, parallel simulation of a smart grid with the goal to simulate nation-sized smart grids. This approach can directly be ported to an actual smart grid. A simulation approach by game-theoretical, cooperative bargaining is followed to determine optimal demand and supply with respect to energy prices to obtain a unique Nash equilibrium. The computed demand and supply comply with the smart grid’s line capacities and structure by solving a modified flow problem in parallel by a two-level approach and forcing the bargaining solution within the capacity limitations. The correctness of the parallelisation is formally derived and experimentally validated. The performance results show that perfect speed-up in strong scaling can be achieved as long as the total time-to-solution per process is approximately equal. The weakscaling experiments exhibited a quadratic increase in complexity, so that the runtime linearly rose with increasing problem size and number of processes. As a proof of concept more than 30 million agents were simulated. Finally, a smallscale and a large-scale simulation with identical setup are compared which showed distinct results. This gives rise to the assumption a smart grid can only be simulated meaningfully by solving the actual problem, not by extrapolating from a small one.