Lightweight Edge Stream Processing Framework and Task Scheduling Algorithm for CNN‐Based Distributed PV Output Prediction
An increasing number of distributed photovoltaic systems utilize convolutional neural network (CNN)‐based models for power prediction, yet face computational bottlenecks when deploying these models on resource‐constrained photovoltaic edge computing terminals (PECT). To address this challenge, this...
Gespeichert in:
| Veröffentlicht in: | IET generation, transmission & distribution Jg. 19; H. 1 |
|---|---|
| Hauptverfasser: | , , , |
| Format: | Journal Article |
| Sprache: | Englisch |
| Veröffentlicht: |
01.01.2025
|
| ISSN: | 1751-8687, 1751-8695 |
| Online-Zugang: | Volltext |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| Zusammenfassung: | An increasing number of distributed photovoltaic systems utilize convolutional neural network (CNN)‐based models for power prediction, yet face computational bottlenecks when deploying these models on resource‐constrained photovoltaic edge computing terminals (PECT). To address this challenge, this paper proposes a lightweight edge stream processing framework integrated with a dynamic task scheduling mechanism, comprising three core components: a data receiving module (DRM) implements real‐time task preprocessing with validity screening, a data computing module (DCM) splits and processes sub‐tasks of CNNs in parallel, and realizes distributed node collaboration. and a data summarizing module (DSM) for data aggregation. The scheduling mechanism combines a modified least laxity first (MLLF) algorithm with dynamic priority adjustment and a self‐monitoring allocation (SMA) algorithm enabling local resource‐aware load balancing. Deployed on the iPACS‐5612C1 IoT terminal, experiments show that the proposed framework achieves a 97% average CPU utilization (85% in baseline methods), a 25% reduction in computing time, and a 90% task completion rate, with the best real efficiency. The framework achieves a real efficiency improvement of 40% over cloud batch processing while maintaining prediction accuracy above 90% under dynamic conditions. Experiments also demonstrate that this framework has the potential to be deployed on larger photovoltaic clusters. These results demonstrate the effectiveness and scalability of the edge stream processing framework. |
|---|---|
| ISSN: | 1751-8687 1751-8695 |
| DOI: | 10.1049/gtd2.70057 |