OVER VOLTAGE SITUATIONS FORECAST AND ACTIVE CONTROL IN A MEDIUM-VOLTAGE POWER GRID

The thesis will focus on how to increase wind energy production using machine learning algorithms, in particular reinforcement learning algorithms. The recent aspirations for a more sustainable energy system and the reduction of CO2 have started a transformation in the power networks. Traditionally considered as passive systems, the power grids are undergoing a rapid change with the introduction of more distributed energy resources. Their introduction requires a better control of the networks to ensure reliability and avoid energy losses. A technical consequence of these devices is the increased number of network’s problems, like over voltages of the lines. This could damage the devices connected to the grid, with social and economic consequences. This thesis intends to investigate the possible solutions when dealing with the issues introduced by these grid changes. It also suggests different techniques to address the challenges of network forecasting and controlling. In particular, to test whether it is possible to predict and respond in time to solve the voltage problems in the network system, some machine learning models are implemented to forecast and to control the network’s devices. Two main learning algorithms are used: supervised learning, for the forecast part; and reinforcement learning, for the control part. The thesis focuses on a medium-voltage network and the analysis of one-year time series measurements of its devices. The time series are built using the Simbench dataset and are adapted to the MV Oberrhein network in order to have a real network with realistic time series. The methods’ results revealed that, starting from the network’s devices measurements, it is possible to forecast the over voltage problems with a certain level of accuracy and in a similar way control these devices to reduce the number of voltage issues.