A Novel Multi-Scale and Adaptive Multi-Period Deep Learning with Compression-Fusion Attention for Cold Storage Load Prediction

Accurate load forecasting is essential for energy-efficient scheduling in cold storage facilities, where cooling demand is shaped by strong periodicity, nonlinear temporal dynamics, and irregular operational disturbances. Traditional statistical and machine-learning models struggle with these multi-scale variations, and existing deep learning approaches often rely on fixed receptive fields or fail to extract adaptive periodic structures. This study introduces MA-CFAN, a multi-scale and adaptive multi-period forecasting framework that integrates temporal decomposition, dynamic frequency-period identification, and a newly designed Compression-Fusion Attention Block (CFABlock) for cross-period representation learning. The architecture leverages FFT-derived adaptive periods to capture seasonal-trend components and employs compression-fusion attention to enhance feature discrimination across temporal scales. Furthermore, this work provides the first systematic evaluation of state-of-the-art forecasting models, including Informer, Autoformer, iTransformer, TimesNet, DLinear, and TimeMixer, to the domain of cold storage load prediction. Experiments on real operational data from a logistics center in Jinan, China, demonstrate that MA-CFAN consistently outperforms all baselines, reducing average MSE and MAE by up to 19.3% and 14.8%, respectively.