Cold chain logistics require strict transport time control due to the high requirement of freshness preservation. It needs high accuracy in forecasting. To improve the accuracy of cold chain logistics demand forecasting, this paper proposes a fusion model PCA-Xavier-BPNN. At first, PCA is applied to reduce dimensionality, eliminate redundant information, and enhance feature representation. Secondly, Xavier initialization is adopted to strengthen the stability of BPNN network training. At the same time, given the small size of the training dataset, a data augmentation method is implemented to make it suitable for small sample sizes. The relationship between cold chain logistics demand and key factors, such as population size and per capita income, is also analyzed. The proposed method is validated using a cold chain logistics demand dataset for fresh agricultural products, specifically fruits and vegetables. Experimental results show that: (1) The fusion effectively improves the accuracy of logistics demand forecasting, with the Mean Absolute Percentage Error (MAPE) remaining below 2%; (2) The data augmentation method significantly improves forecasting accuracy by more than 20%.