Full Data-Processing Power Load Forecasting Based on Vertical Federated Learning

Abstract

Power load forecasting (PLF) has a positive impact on the stability of power systems and can reduce the cost of power generation enterprises. To improve the forecasting accuracy, more information besides load data is necessary. In recent years, a novel privacy-preserving paradigm vertical federated learning (FL) has been applied to PLF to improve forecasting accuracy while keeping different organizations’ data locally. However, two problems are still not well solved in vertical FL. The first problem is a lack of a full data-processing procedure, and the second is a lack of enhanced privacy protection for data processing. To address it, according to the procedure in a practical scenario, we propose a vertical FL XGBoost-based PLF, where multiparty secure computation is used to enhance the privacy protection of FL. Concretely, we design a full data-processing PLF, including data cleaning, private set intersection, feature selection, federated XGBoost training, and inference. Furthermore, we further use RSA encryption in the private set intersection and Paillier homomorphic encryption in the training and inference phases. To validate the proposed method, we conducted experiments to compare centralized learning and vertical FL on several real-world datasets. The proposed method can also be directly applied to other practical vertical FL tasks.

Advancing Power System Services With Privacy-Preserving Federated Learning Techniques: A Review

“…Vertical FL is particularly valuable in environments where the parties hold the same samples but with scattered features. For example, vertical FL is employed in scenarios for accurate load forecasting using different features located in different data partitions for model training [96]. For example, in [38], an innovative learning framework integrating vertical federated learning and horizontal learning with an XGBoost-based learning framework is proposed to address the distributed features of datasets for utility power prediction in China.…”