Behind the Paper

Voltage regulation in PV-rich distribution networks: an edge pipelined intelligent computing approach

We propose a deep neural network–based voltage regulation method for distribution grid edge nodes, using a unified-weight model and pipelined parallel computation to achieve compact and efficient inference, demonstrating strong suitability for edge deployment.
Yong Li Dec 12, 2025

As distributed photovoltaic (PV) capacity grows, distribution networks increasingly face voltage violations when PV generation exceeds local demand, causing reverse power flow. PV inverters can mitigate this via reactive power control. Traditional system-level optimization relies on a central server to collect network-wide data and solve power flow problems, which is costly and introduces systemic risks if communication links fail. The rise of distributed resources motivates decentralized, autonomous operation, where fast response from edge computing nodes is critical. However, conventional numerical algorithms are computationally intensive, and deep reinforcement learning (DRL), while “model-free”, typically requires large models and extensive training data. Deploying DRL on edge devices often involves manual compression, increasing development cost and potentially degrading performance.

To address this, we propose NN4VVC, a unified-weight Neural Network for Volt-Var Control that combines regionally distributed strategies with FPGA pipeline hardware for high-efficiency edge inference. Key design features include:

  • Unified weights with regionally distributed VVC: The distribution network partitions the distribution system into regions. All inverters share a single network, using agent IDs as input to produce differentiated actions, significantly reducing model size and storage requirements.
  • Pipeline FPGA hardware: Forward propagation is decomposed into matrix multiplication, accumulation, and nonlinear modules, executed in a streaming pipeline without intermediate buffering, improving speed and energy efficiency.
  • Queue scheduling and parallel execution: A queue buffer algorithm allows multiple inverters to execute inference in different pipeline stages concurrently, achieving high-throughput edge computation.

We validated NN4VVC on the IEEE 33-bus distribution system (6 PVs, 32 loads) using real PV and load data at 3-minute resolution, with full nonlinear AC power flow. Key findings include:

  1. Significant voltage compliance: All nodes remained within 0.95–1.05 p.u., even under large PV fluctuations (midday overvoltage, evening undervoltage).
  2. Near-optimal control actions: NN4VVC learned selective reactive power control, involving only key PVs and avoiding the “one-size-fits-all” approach of traditional droop control.
  3. Ultra-fast edge inference: Custom hardware pipelines complete neural network forward computation within  sub-microsecond, supporting parallel processing across multiple nodes.
  4. Resilience under communication failure: NN4VVC operates stably using only local observations, without relying on full communication connectivity.

The results show that NN4VVC can run efficiently and reliably on edge devices. By enabling computation close to the data sources, distribution networks can transition toward truly distributed and autonomous energy systems.