A Dual Sharing Strategy in Multi-Task Learning for Harmonic Signal Decomposition

Abstract

Our work presents the use of neural networks for power grid voltage quality analysis. Power quality parameters are defined by regulations, and therefore knowledge of voltage harmonics is essential for assessing electrical energy quality and connected loads. We propose a new dual multi-task learning approach consisting of soft-parameter sharing for phase prediction and hard-parameter sharing for harmonic amplitude classification. The study is based on realistic synthetic datasets generated strictly according to EU standard limits for harmonics of orders 3, 5, 7, 9, 11, and 13. Rectifier systems supplying DC electric traction generate mainly 5th and 7th harmonics (6-pulse rectifier) as well as 11th and 13th harmonics (12-pulse rectifier). Extensive experiments were performed for different classification variants and phase prediction error margins. The proposed model achieved high performance, detecting harmonics with an average accuracy of 96.41% and predicting disturbance phases with an average accuracy of 94.25% for a 10% error margin.