Data-Driven Power System Operation

Most decision makers understand that climate change origins from the large use of carbon-heavy energies. To jointly stop this climate change was contracted by over 190 countries in the Paris Agreement in 2015, a historical key success. To reduce carbon-heavy energy generation we will move to renewables, as of their incomparably low (zero) carbon emissions. In the future, this high generation-share of renewables will rise dramatically challenges across all Energy Networks, from its generation to consumption over transmission and distribution. The key challenge is the fundamental different operational characteristics to conventional power generation. Conventionally, power is generated steadily and can be called on-demand, whereas renewables power generation, such as Wind Power, is extremely uncertain and cannot be easily controlled. To remain controllability of the system in the future, renewable energy may be (distributionally) stored (Energy Storage), or the power demand can be managed accordingly as well (increasing Energy Efficiency). Consequently, power generation will be much more dynamic, distributed and uncertain.

Current tools to operate the system fail to handle this new dynamic, distributed and uncertain power generation. In the past, when generators were able to generate power on-demand, the power system operation involved matching generation to the demand, while retaining its physical feasibilities. In the future, when generation and consumption are extremely uncertain, more phenomena will arise that corresponds to the dynamics of the system. These dynamic phenomena can lead to instabilities in power system operation and, therefore, increases the danger that the system collapses (power blackouts). The blackouts across multiple counties (e.g., half Europe in 2006) were eye-openers for many politicians and operators. Therefore, to avoid power blackouts operators must operate the system in a very conservative way. This conservatism is extremely cost-inefficient but necessary as, currently, no appropriate approaches exist to account for dynamic phenomena and the uncertainty in operation-planning.

In the context of this new reality, this project aims to develop radically new approaches that can use all assets in the system harder at its physical limitations. We aim to address the following three research questions:
(i) How to model the uncertainty in an automatic operation-planning model?
(ii) How to efficiently predict the impact of dynamic phenomena using the operating state of the power system?
(iii) How to derive optimal preventive and corrective actions in a highly uncertain operating environment?

Our approaches to these questions use data-driven. We start by using methods from Statistics and Applied Probability, learning a statistical model from historical observations. Subsequently, we generate a large training database form this model to extrapolate from these observations. When we model the uncertainty of the system to address (i), we investigate learning suitable operational constraints to describe the uncertainty. When we deal with dynamic phenomena to address (ii) and (iii), we continue with simulating dynamic phenomena to enrich the training database by the simulation outputs, for either suffering or withstanding the dynamic phenomena. By using this enriched training database, we investigate different high-performance classifiers from Artificial Intelligence, in (ii) to predict how a new operating point performs under the dynamic phenomena or in (iii) to be used in an operating planning model, utilizing methods from Operations Research, to research how optimal preventive and corrective actions can be derived. These actions ensure that all operating points withstand the dynamic phenomena.