Incorporating Grid Effects into Bottom-Up Load Models
Background and motivation for the project
Due to the energy transition and ongoing digitalization of the society, the Dutch power grid has become a bottleneck for further developments. Many initiatives are undertaken to relieve the pressure on the grid, both by extending the grid faster and by making more efficient use of the grid over time. Essential for both are fast and reliable calculations of the grid. Bottom-up load models, which add up the consumption and generation of individual customers up to a certain point in the grid offer a granular and fast approach to understanding and predicting demand (Peter van de Sande, 2017) (Roel Bouman, 2024). However, these models often overlook the nuanced effects of the grid, such as grid losses, cable reactance and capacitance. Integrating these elements can improve the accuracy and reliability of load forecasts, thereby enhancing grid capacity planning and operational efficiency. To do this a detailed load flow calculation is required (Power-flow study wikipedia page, 2025), which makes the calculation much slower and even unstable. Also it might be even impossible because smart meter data can often only be used in an aggregated way and not coupled to their individual locations on the grid. (J. Heres, 2023). Finally, a method that is 100% physically accurate is often not needed, because there is a substantial uncertainty in the customer loads but more importantly the grid properties. It would therefore be helpful if there would be a method that gives a reasonable estimate of the grid effects, but much faster and without the need to disaggregate the bottom-up loads into individual ones.
The research goal
The primary goal of this project is to develop a method, using machine learning methods, to include an estimate of effects of the grid on the active and reactive power, without performing a full load flow calculation. It can be using the physical information of the grid in a graph structure, which includes the electrotechnical properties such as the impedance and capacitance and locations of the individual loads. Efficient load flow models such as in the open source power-grid model can be used for training and benchmarking. (Power Grid Model project page, 2025).
The method you envision
A relatively straightforward method would be to model the whole distribution grid and all it’s customer’s as a single line connecting a source (the high voltage grid) to the customers (a single load object). The goal of the machine learning model would then be to calculate an effective resistivity, reactance and capacitance of the whole grid from the input-graph. Probably a graph neural network (GNN) is able to do this, because it can map the varying data structure of different input graph to a standardized output. The effective grid properties can then be used to do a load flow on this very simple grid, which is much faster then performing a load flow on a grid with around a hundred thousand of components.
Maybe it is however even more effective to directly train a GNN on the output of a loadflow calculation, so that it can predict the total grid losses and effect on the reactive power when applying a given bottom-up load to the grid.
What also could be beneficial is to split up the bottom-up load into groups of several customers, as e.g. bulk customers might have a different effect on e.g. the grid losses then residential customers, because they are located at different places in the grid.
Contact: Yuliya Shapovalova.
Supervision: Yuliya Shapovalova (Radboud University), Jacco Heres (Alliander).
References
J. Heres, M. v. (2023). Creating bottom up load profiles using disaggregation, clustering and supervised machine learning on large smart meter dataset. CIRED. Rome.
Peter van de Sande, M. D. (2017). ANDES: grid capacity planning using a bottom-up, profile-based load forecasting approach. CIRED. Glasgow.
Power Grid Model project page. (2025). Opgehaald van LF energy: https://lfenergy.org/projects/power-grid-model/
Power-flow study wikipedia page. (2025, 04 21). Opgehaald van Wikipedia: https://en.wikipedia.org/wiki/Power-flow_study
Roel Bouman, L. S. (2024). Acquiring better load estimates by combining anomaly and change point detection in power grid time-series measurements. Sustainable Energy, Grids and Networks. doi:doi.org/10.1016/j.segan.2024.101540