Multi-Step Prediction and Optimization of Harmonic Power Quality in Active Distribution Network of Urban Energy Internet Driven by PatchTST

The integration of high-density distributed energy, complex loads, and multi-energy flows in urban energy internets complicates the dynamic uncertainty of harmonic evolution, driven by high-frequency fluctuations from distributed energy sources and sudden load switching in industrial processes. To address the resulting inaccurate harmonic power quality prediction and optimization in active distribution networks, this paper proposes a multi-step prediction and collaborative optimization method based on PatchTST. The method enhances complex time series modeling via a localglobal feature fusion mechanism. It extracts local temporal characteristics of harmonic distortion using local time block coding, and mines long-range dependencies between multi-channel variables using a global self-attention network. This achieves multi-step prediction of the total harmonic distortion rate (THD) and individual harmonic components. The study then constructs a physical-information-embedded collaborative control model, combining prediction results with dynamic adjustment of inverter reactive compensation and optimization of energy storage strategies to form a closedloop control. Experiments show the PatchTST-based method achieves low MAE for THD multi-step prediction (0.077/0.119/0.140 for 12/24/48 steps), reduced active power loss (13.62kW), fast voltage deviation recovery (2.2–5s), and low control failure (1.2%). Physical constraint embedding improves optimization feasibility to 93.5% and limits SOC violations to 6. This research offers an integrated solution for complex power quality issues through the deep integration of deep learning and power electronic control, providing a framework directly applicable to improving the stable and reliable power supply required for continuous, high-quality production in the textile industry.