Extending Privacy-preserving Regression Trees to a Value-oriented Forecasting Approach for Trading DER Resources - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2024

Extending Privacy-preserving Regression Trees to a Value-oriented Forecasting Approach for Trading DER Resources

Résumé

In the state of the art, it has been shown that in several applications like spatiotemporal forecasting of wind or solar power, accuracy in the next hours can be improved if data from neighbor plants are used as input. However, this is not always feasible due to confidentiality constraints that prevent data sharing. This required privacy-preserving forecasting methods which are also applicable to other use-cases like EV-load demand. We apply vertical federated learning, leveraging different sources of information. The Energy forecasts are used to optimize decisions in applications. In cases like energy trading, recent advances have shown that decisions and the resulting value (i.e., revenue on energy markets) can be enhanced if forecasting models are optimized for their value in the application instead of accuracy. We combine these two aspects in order to maximize the potential of forecasting models: 1) confidentiality preserving data sharing and 2) a forecasting model optimization based both on accuracy and value. We focus on the EV charging stations' demand forecasting. We will focus on tree-based models for a combined approach for the following reasons. First, as shown in real-world applications, trees with their non-parametric design are suited to complex tasks with unclear relationships. For vertical federated learning, gradient-boosted trees fit the design naturally. For value-oriented approaches, single-decision trees or Random Forests are more favored due to their natural fit. In previous works, we developed a privacy-preserving multivariate tree model. We propose its adaption with a loss function surrogate of SPO+. Additionally, we propose adapting SPO-trees to a multivariate data-sharing setting. We evaluate both approaches and apply secret-sharing to the better-performing approach. Secret-sharing is a lossless, efficient encryption technique. We benchmark the models on two European EV charging station data sets, Dundee and Paris. We evaluate the revenue of the single-price market mechanism, where each charging station owner acts as a buyer. We use a symmetric approach as if an owner was a seller. We evaluate the benefits of data-sharing vs. a non-data-sharing application. The results indicate that the combination of approaches outperforms each individual approach and higher efficiency.
Fichier non déposé

Dates et versions

hal-04637097 , version 1 (05-07-2024)

Identifiants

  • HAL Id : hal-04637097 , version 1

Citer

Lukas Stippel, Simon Camal, Georges Kariniotakis. Extending Privacy-preserving Regression Trees to a Value-oriented Forecasting Approach for Trading DER Resources. 44th International Symposium on Forecasting, Jun 2024, Dijon, France. ⟨hal-04637097⟩
21 Consultations
0 Téléchargements

Partager

Gmail Mastodon Facebook X LinkedIn More