Time–Frequency Domain Joint Noise Reduction Multi-Resolution Power System Time Series Prediction Network

Abstract

This paper proposes an innovative time series prediction method designed for power systems to overcome the shortcomings of existing deep learning techniques in complex noise environments. The method, called Time–Frequency Domain Joint Noise Reduction Multi-Resolution Power System Time Series Prediction Network (TDRSE), investigates the impact of noise on prediction results and proposes a complete solution. TDRSE consists of two key components Exponential Decay-based Denoising Network (EDnet) and Dynamic Frequency-Domain Signal Enhancement Network (FDse). EDnet achieves dynamic attention to different time points in the time dimension by introducing exponential decay units to cope with the volatility of power loads and noise disturbances. At the same time, FDse employs frequency-domain enhancement techniques and adaptive thresholding strategies to remove the noise components in the frequency domain, thus further improving the model’s power data prediction accuracy. The experimental results show that TDRSE performs well on real data sets of multiple power systems, significantly improves the prediction accuracy under complex noise conditions, and reaches the industry-leading level.