Deep Learning
Advanced State Estimation for Distribution Grids Based on Deep Learning
State estimation techniques establish a link between the measuring devices of an electrical network and the control centers to present and monitor all the electrical magnitudes of the network itself. State estimation approaches have been applied in transmission networks for many years, but these techniques are not directly applicable in distribution networks.
In distribution systems, State estimation algorithms aim to obtain the state of the system, i.e. the voltages at all buses and connection points. It is a necessary process before calculating the power flow and safety analysis. In addition, it should be added that traditional state estimation only takes into account statistical distributions to model uncertainties in measurements (e.g., power distribution among single-bus feeders).
Development of Deep Learning Algorithms
The development is based on a multipurpose prediction software module. This module includes a time series decomposition model with three main components: trends, seasonality (in different periods, varying from daily period to an annual period), and exogenous variables (holidays, weather, etc.). The modeling of these components can be summarized as follows:
- Trends can be modeled using two internal models: A saturated nonlinear growth model and a trend model with change points. The saturated nonlinear growth model captures the initial acceleration and subsequent deceleration as a limit is reached. The trend model with change points identifies specific moments where the trend changes direction or speed, allowing a better adaptation to market dynamics
- Seasonality in energy consumption, such as increased demand in summer due to the use of air conditioning, can be modeled using Fourier series. By breaking down demand into sinusoidal components, seasonal patterns can be identified, and deep learning models can be adjusted to forecast these periodic variations. This improves the accuracy of the state estimation by taking into account regular fluctuations in energy demand.
- Exogenous variables are added to the overall model as additive linear regressions. These variables, which are factors external to the system (but influence its behavior) are incorporated into the model to improve the accuracy of the predictions. In the context of state estimation, exogenous variables can include factors such as temperature, time of day, day of the week, special events, or even energy prices. By adding these variables using additive linear regressions, their direct and linear impact on the variable of interest can be captured, allowing the model to adjust its predictions more accurately and better reflect the reality of the electricity system. These results in a more robust and reliable.
For the generation of the models, historical data by the project Pilot Sites have been used.
Integration into the ETRA I+D ÉTER software tool
ETRA I+D has its own tool called ÉTER, developed and validated in real conditions in the different European projects in which it has participated. The purpose of ÉTER is to control, manage and monitor an electricity distribution network, improving its flexibility, stability and security; being particularly relevant in scenarios with a high penetration of renewable energies and that require better demand management.
ÉTER currently integrates a State estimation module, but based mainly on probabilistic methods. The ultimate goal of ETRA is to integrate this new State estimation module based on Deep Learning into ÉTER once it has been validated in the OPENTUNITY pilots.
The following image shows how the power flows calculated with the current State estimation module are displayed in ÉTER. The visualization once the new module is integrated will be similar, but the calculations are expected to be more accurate and reliable.