Article: Gradient sparsity compression algorithm for incremental pre-training of large models in power communication scenarios Journal: International Journal of Information and Communication Technology (IJICT) 2026 Vol.27 No.78 pp.30 - 51 Abstract: In power communication scenarios, incremental pre-training of large models faces challenges such as limited bandwidth and limited edge device resources, and existing gradient compression methods are difficult to adapt to dynamic network environments. To this end, this paper proposes a power-aware adaptive gradient sparsity compression (PAGSC) algorithm, which achieves efficient training by fusing dynamic sparsity, hierarchical quantisation and channel adaptive scheduling mechanisms. The test results show that on data sets such as CIFAR-100 and UCI power, PAGSC reduces the communication volume to 12% of the baseline, with only 2% accuracy loss, reduces training time by 19%, increases bandwidth utilisation to 85.6%, and significantly optimises resource use. To sum up, the algorithm effectively balances accuracy and efficiency. However, further research on hyperparameter adaptive optimisation and noisy channel robustness is needed. The innovation of this paper lies in providing a deployable lightweight training solution for the power internet of things. Inderscience Publishers - linking academia, business and industry through research

Title: Gradient sparsity compression algorithm for incremental pre-training of large models in power communication scenarios

Authors: Xiangcheng He; Qun Xie; Xin Yang; Lin Guo; Xuedong Sun; Jing Zhang

Addresses: State Grid Yibin Electric Power Supply Company, Yibin, Sichuan, China ' State Grid Yibin Electric Power Supply Company, Yibin, Sichuan, China ' State Grid Yibin Electric Power Supply Company, Yibin, Sichuan, China ' State Grid Sichuan Electric Power Company, Chengdu, Sichuan, China ' State Grid Yibin Electric Power Supply Company, Yibin, Sichuan, China ' State Grid Yibin Electric Power Supply Company, Yibin, Sichuan, China

Abstract: In power communication scenarios, incremental pre-training of large models faces challenges such as limited bandwidth and limited edge device resources, and existing gradient compression methods are difficult to adapt to dynamic network environments. To this end, this paper proposes a power-aware adaptive gradient sparsity compression (PAGSC) algorithm, which achieves efficient training by fusing dynamic sparsity, hierarchical quantisation and channel adaptive scheduling mechanisms. The test results show that on data sets such as CIFAR-100 and UCI power, PAGSC reduces the communication volume to 12% of the baseline, with only 2% accuracy loss, reduces training time by 19%, increases bandwidth utilisation to 85.6%, and significantly optimises resource use. To sum up, the algorithm effectively balances accuracy and efficiency. However, further research on hyperparameter adaptive optimisation and noisy channel robustness is needed. The innovation of this paper lies in providing a deployable lightweight training solution for the power internet of things.

Keywords: power communication; large model incremental pre-training; gradient sparse compression.

DOI: 10.1504/IJICT.2026.155025

International Journal of Information and Communication Technology, 2026 Vol.27 No.78, pp.30 - 51

Received: 14 Jan 2026
Accepted: 30 Apr 2026
Published online: 22 Jul 2026 *

Title: Gradient sparsity compression algorithm for incremental pre-training of large models in power communication scenarios

Keep up-to-date