Research on Compression Algorithms for Time Series Prediction Models in Source Measure Units Based on Pruning, Quantization, and Knowledge Distillation

The BA-Informer algorithm is a specialized timeseries forecasting model designed to enhance measurement efficiency in source measure unit (SMU). Although it achieves high prediction accuracy, its large-scale architecture presents significant challenges for deployment on resource-constrained FPGA platforms embedded in SMUs. To address this issue, this paper introduces a novel model compression framework incorporating pruning, parameter quantization, and knowledge distillation techniques. The proposed method effectively compresses the original model into a lightweight version with minimal performance degradation. Experimental results demonstrate that this approach significantly reduces computational complexity and memory consumption without sacrificing prediction accuracy, thereby increasing its applicability in practical industrial settings.