ZEBRA-SPARSEECHONET: A HYBRID TRANSFORMER–BASED ECHO STATE NETWORK OPTIMIZED WITH THE ZEBRA ALGORITHM FOR SMART LOGISTICS DEMAND FORECASTING

Abstract

Background: Smart logistics has emerged as a transformative approach to supply chain management, leveraging advanced information technologies and intelligent systems to enhance operational performance. Within this context, accurate demand forecasting is critical, as it shapes inventory planning, resource allocation, cost control, and customer satisfaction. However, demand is inherently volatile—driven by stochastic fluctuations, seasonal patterns, and sudden disruptions arising from external events or shifts in consumer behavior. These characteristics make traditional forecasting models insufficient for capturing the complexity of real-world demand dynamics, underscoring the need for more intelligent and adaptive forecasting approaches. Methods: This study introduces Zebra-SparseEchoNet, a novel hybrid AI forecasting framework designed for high-accuracy prediction in dynamic logistics environments. The methodology begins with extensive data preprocessing, including cleaning, exploratory analytics, and visualization to ensure reliable inputs. Feature transformation is subsequently applied through categorical encoding and numerical normalization to enhance model interpretability, stability, and overall predictive performance. The curated dataset is then divided into training and testing sets to enable unbiased evaluation. The core forecasting engine merges two complementary components. First, a Sparse Self-Attention Transformer (SAT) captures long-range temporal dependencies while reducing computational load by selectively attending to the most informative interactions within the sequence. Second, an Echo State Network (ESN) offers lightweight, memory-efficient modeling of short-term dynamics using its fixed recurrent reservoir. To further strengthen forecasting capability, the Zebra Optimization Algorithm (ZOA) is utilized for intelligent hyperparameter tuning, ensuring optimal architectural configuration and accelerated convergence across components. Results: Experimental results demonstrate that Zebra-SparseEchoNet outperforms conventional deep learning models across multiple forecasting error metrics. The model also achieves higher R² values and explained variance, indicating a better fit to underlying demand structures and more reliable forecasts. Conclusion: For smart logistics applications, the Zebra-SparseEchoNet framework delivers significant gains in both computational efficiency and forecasting precision. By effectively capturing intricate temporal patterns and maintaining scalability through the integration of Zebra Optimization, Echo State Networks, and Sparse Self-Attention Transformers, the model demonstrates strong potential as a dependable and sophisticated solution for contemporary supply chain and inventory management challenges.