PackMLP: A Patch Based MLP Model for Intelligent Inventory Forecasting and Automation with A-LIQ QR Scanning Integration

Abstract

Traditional inventory management systems are often constrained by manual inefficiencies, inaccurate forecasting, and poor scalability, particularly in dynamic retail environments. This paper introduces a novel, intelligent inventory management solution that integrates a high performance PackMLP model for time series sales forecasting with a modified A-LIQ (Advanced Laser Inertial QR) code scanning framework for real time inventory tracking. The scanning module leverages inertial correction and visual calibration to enhance recognition robustness, achieving 99.2% accuracy with an average latency of 0.9 seconds, even under suboptimal scanning conditions. For forecasting, we propose PackMLP, a lightweight, attention free deep learning model that combines multi-scale patch embedding, trend residual decomposition, and cross channel MLP blocks to capture temporal and inter-variable dependencies. Evaluated on the M5 Walmart dataset, PackMLP outperforms leading state-of-the-art (SOTA) models, including Transformer, Informer, PatchTST, PatchMLP, and TSMixer, achieving the lowest NWRMSLE (Normalized Weighted Root Mean Squared Logarithmic Error) of 0.0811, MAE (Mean Absolute Error) of 0.358, and MAPE (Mean Absolute Percentage Error) of 14.6%, while also offering the smallest model size (39 MB) and fastest inference time (12.2 ms/sample). A one-month deployment in a real retail setting demonstrated a 4.4% improvement in inventory accuracy, a 25% reduction in stockouts, and a 15% decrease in holding costs compared to manual systems. With its modular architecture, web-based dashboard, and real time low stock alerts, the system offers a cost effective, scalable alternative to complex enterprise platforms. These results confirm PackMLP’s potential as a practical and high accuracy forecasting engine for next generation inventory management systems.