Development and Application of a Machine Learning-Based Load Flow Forecast

Inhalt:
Supply and demand in an electric power system must be balanced at all times. Ensuring this in a transmission system is one of the tasks of the transmission system operator (TSO). To meet this requirement, TSOs must constantly determine future load flow and take action as needed prior to real time. European TSOs established several forecast processes, with forecast horizons ranging from a few hours to seven days ahead, which predict the load flow on all power lines and help to determine potential redispatch demand. These processes require a variety of input variables from multiple internal and external data sources, including load and generation forecasts as well as the grid topology. As some of the required input variables cannot be explicitly obtained beforehand, assumptions and simplifications have to be made. To reduce the associated uncertainty, this paper presents a new holistic approach to develop a forecast model for load flow based purely on empirical and forecasted data. The approach presented in this work can be applied with state-of-the-art machine learning (ML) and deep learning (DL) algorithms such as Light Gradient Boosting Machine (LGBM), Long Short-Term Memory (LSTM) or Temporal Fusion Transformers (TFT) models. Training and evaluation is performed using weather forecast data from the DWD ICON-EU model, publicly available historical data from the ENTSO-E transparency platform, as well as realized data gathered at Amprion, a German TSO. Corresponding results are presented demonstrating reasonable prediction accuracy of the approach.