Memory-Efficient Imagification for Light-weight Prediction Model of Multivariate Time-Series Data

This paper addresses the challenge of memory-efficient time-series forecasting in resource-constrained environments. To this end, an imagification method is proposed that enables lightweight convolutional neural network CNN-based prediction by transforming multivariate time-series data into image representations. The method consists of three steps: rearranging features using the Pearson correlation coefficient to enhance local associations, generating images through a sliding window technique along the time axis, and applying multivariate data interpolation to improve smoothness across the feature axis. The proposed method is evaluated using real-world traffic speed data collected from highway sections in Seoul, South Korea. Compared to benchmark imagification methods (Recurrence Plot, Gramian Angular Field) and a conventional multivariate LSTM model, the proposed approach achieves competitive prediction accuracy with significantly reduced training time. These results suggest that the method is well-suited for deployment in embedded or low-memory systems requiring efficient time-series prediction.