Accurate and Explainable Electrical Load Forecasting Using Time-Series Transformers

Accurate electrical load forecasts are essential for the operation of an increasingly complex energy system, including for maintaining the balance between supply and demand, for preventing grid congestion, and for operating energy management systems (EMS).
In addition to forecast accuracy, explainability is crucial for fostering user trust and meeting regulatory transparency requirements. This dissertation evaluates and extends time-series Transformers for accurate and explainable electrical load forecasting.

First, different strategies for training time-series Transformers on data from multiple metering units are compared. The results show that a global Transformer model generalizes well across time series and achieves lower forecast errors than local, multivariate, and cluster-specific models.

Second, a flexible Transformer architecture is introduced, that integrates architectural modifications from previous work and allows for an automated architecture optimization.
This Transformer is benchmarked against established forecasting methods on three electrical load datasets representing the TSO level, the low-voltage feeder level and the client level.
... mehr
Transformer-based models outperform established methods across grid levels, reducing the forecast error by 6.6-10.7% compared to the next-best method.
The Time-Series Foundation Model (TSFM) Chronos-2 is competitive with specialized load forecasting models on datasets from low aggregation levels, but it is inaccurate for special events in the TSO data.

Third, the potential of the Transformer’s attention mechanism for explainable time-series forecasting is investigated. An analysis of individual attention heads reveals distinct functional patterns; however, their interpretability and the insights into feature dependencies remain limited. To address these limitations, a novel Transformer variant, termed SHAPformer, is proposed. SHAPformer uses attention manipulation to efficiently compute Shapley Additive Explanations (SHAP). The resulting explanations provide insights into the trained model, including seasonal patterns and dependencies on covariates. Furthermore, an adaptation of the SHAPformer approach to TSFMs is presented.

Fourth, the effectiveness of Transformer-based forecasts is evaluated in two downstream applications. In the BigDEAL peak load forecasting competition, a conditional Invertible Neural Network (cINN) outperforms the Transformer. In contrast, in a peak shaving application for an energy community comprising heat pumps and battery storage, Transformer-based forecasts enable more effective peak load reduction compared to alternative approaches.

Overall, this dissertation demonstrates the effectiveness of Transformers for electrical load forecasting and introduces methods to enhance model explainability based on the attention mechanism. These findings contribute to both the methodological advancement and the practical deployment of electrical load forecasting models in modern energy systems.