CleanStockNet: A Gated Linear Unit Convolution-Enhanced Temporal Attention Fusion Network for International Clean Energy Stock Prediction

Effective inventory management of cleaning resources is crucial for operational efficiency in industries like healthcare and hospitality, where demand fluctuates due to seasonal changes and unexpected events. Traditional forecasting methods often fail to predict these dynamic demands accurately. To address this, we introduce CleanStockNet, a novel forecasting framework combining Gated Linear Unit Convolution-Enhanced and Temporal Attention Fusion Network, designed to enhance prediction accuracy for cleaning resource stocks over a 16-day forecast horizon. CleanStockNet integrates the Gated Linear Unit architecture with a dual-stage temporal attention mechanism, enabling it to capture long-term dependencies in time series data. This capability is crucial for handling irregular inventory demand patterns. The model processes historical data across 1 to 96 window intervals, adapting to long temporal patterns and improving prediction granularity. Our comprehensive evaluation using the SPGTCLEN (Singapore General Hospital Time Series of Cleaning Inventory), CELS (Centralized Environmental Logistics System), and ICLN (Integrated Cleaning Logistics Network) datasets shows that CleanStockNet significantly outperforms traditional methods. The model demonstrated an R-square accuracy above 0.87, with the highest result at the 4th step prediction (0.988). It also achieved a reduction in Mean Square Error (MSE) by over 20% and R² values above 0.9, indicating an exceptional fit to data variability. These results highlight CleanStockNet's potential to revolutionize inventory management by providing highly accurate, reliable forecasts that can significantly enhance decision-making in resource management and planning. This framework sets a new benchmark in predictive analytics for inventory management and offers potential applications across other sectors facing similar challenges.