Vertical Power Flow Forecast with LSTMs Using Regular Training Update Strategies

TL;DR

This paper proposes a novel LSTM-based method with regular retraining strategies to improve vertical power flow forecasting at transformers, effectively adapting to dynamic grid conditions and outperforming baseline models.

Contribution

Introduces a regular retraining approach for LSTM models to enhance power flow forecasts amidst changing grid characteristics.

Findings

Regular retraining improves forecast accuracy.

Different update strategies impact model performance.

Proposed method outperforms baseline models.

Abstract

The strong growth of renewable energy sources and the high volatility in power generation of these sources, as well as the increasing amount of volatile energy consumption is leading to major challenges in the electrical grid. In order to ensure safety and reliability in the electricity grid, the power flow in the grid needs to be observed to prevent overloading. Furthermore, the energy supply and consumption need to be continuously balanced to ensure the security of energy supply. Therefore a high quality of power flow forecasts for the next few hours within the grid are needed. In this paper we investigate forecasts of the vertical power flow at transformer between the medium and high voltage grid. Forecasting the vertical power flow is challenging due to constantly changing characteristics of the power flow at the transformer. This is mainly a result of dynamic grid topologies, changes in the installed assets, maintenance of the transformer itself as well as the volatile generation. In this paper we present a novel approach to deal with these challenges. For the multi step time series forecasts a Long-Short Term Memory (LSTM) is used. In our presented approach an update process where the model is retrained regularly is investigated and compared to baseline models. The model is retrained as soon as a sufficient amount of new measurements are available. This retraining should capture changes in the characteristic of the transformer that the model has not yet seen in the past and therefore cannot be predicted by the model. For the regular update process we investigate different strategies where especially the number of used epochs are considered, but also different learning rates are used. We show that our new approach significantly outperforms the investigated baseline approaches.