Multi-Year Time-Series-Based power system planning with hybrid optimization and supervised learning methods

The increasing share of renewable energy sources (RES) in the power system necessitates new planning methods for power systems systems. On the one hand, flexible operational measures must be included in planning. On the other hand, conventional measures have to be considered. An integrated optimization of both measures is needed. This integrated optimization requires time series simulations of the power system in comparison to traditional worst case planning approaches. The high voltage (HV) level has the highest security requirements of all distribution levels. Typically, HV systems have a meshed topology, and the single contingency policy (SCP) is additionally considered in planning. Finding a trade-off between computational effort and solution quality is the main challenge when considering time series simulations and the SCP. A power system planning strategy is needed, which is able to find an investment decision for multiple years within a realistic simulation time. In this thesis, a multi-year planning strategy for meshed HV systems is proposed considering operational flexibility as well as conventional planning measures. The defined optimization problem is solved by a hybrid optimization algorithm combining the advantages of heuristic and mathematical programming approaches. A reduction of the high computational effort of time series simulations is achieved by several strategies, including a custom Newton-Raphson power flow implementation and an efficient time series module, which are integrated into the open-source tool pandapower. Furthermore, several machine learning algorithms are implemented and compared to approximate bus voltages and power line loadings. Operational simulation models of two curtailment strategies and two storage system operational models are implemented. An exhaustive evaluation of four optimization metaheuristics, three mathematical programming approaches, and the developed hybrid approach is shown. Results are validated on four realistic benchmark systems and a real power system model. Several benchmarks show that the implemented methods significantly reduce the calculation time of time series simulations. The Newton-Raphson implementation is up to 30 times faster than comparable open-source versions. A further reduction of the simulation time is possible with the implemented regression method based on an artificial neural network (ANN). The ANN correctly identifies more than 99.4% of all critical time steps for the benchmark cases, including contingency situations when trained with 10% of the time series data of one year. The developed hybrid optimization method is a combination of the Iterated Local Search metaheuristic and a linear optimization model. This combination increases convergence while reducing simulation time in comparison to the existing methods.The hybrid strategy is the only method of all compared algorithms, which finds valid solutions in large optimization space, including replacement and switching measures. Additionally, the simulation time is reduced by up to 80% for the benchmark cases. Finally, two case studies show the applicability of the developed planning framework for a real HV power system model. In these case studies, curtailment of energy from RES and line replacement measures are regarded for a planning horizon of 12 years. The results show that a reduction of up to 60% of the total expenditures, compared to the worst case method, is possible by combining the optimization of RES curtailment and line replacement measures. Expenditures can further be reduced by 4% when using operational flexibility from storage systems.