Efficient Multivariate Time-Series Forecasting via Latency-Constrained Distillation and Dynamic Pruning

Multivariate time-series forecasting models often achieve improved prediction accuracy at the cost of increased model complexity and inference latency, limiting their deployment on resource-constrained computing platforms. In addition, lower parameter counts or theoretical computational costs do not necessarily lead to faster execution on actual hardware. To address this issue, we propose a Latency-Constrained Distillation and Dynamic Pruning (LCDP) framework for efficient multivariate time-series forecasting. LCDP introduces a latency-constrained adaptive optimization strategy that jointly considers forecasting loss, knowledge-distillation loss, local feature-alignment loss, and hardware inference latency during student-model training. Local feature-guided knowledge transfer is used to enhance the student model's ability to capture fine-grained temporal patterns, while deployment-oriented dynamic structural pruning progressively adjusts the group-wise sparsity pattern of the student model according to latency feedback from the target hardware while retaining the original dense layer dimensions. Experiments on the ETTh1, ETTm1, and ETTm2 datasets using two CPU platforms show that LCDP achieves competitive forecasting accuracy with a lightweight student inference backbone containing 7336 active parameters and requiring 50,848 dense multiply-accumulate operations. The saved student checkpoint additionally contains 2080 parameters in a training-time feature-alignment projection that is not invoked during inference. LCDP provides an effective balance between forecasting accuracy and practical inference efficiency, demonstrating its potential for low-latency multivariate time-series forecasting on CPU-based resource-constrained platforms.