ZHOU Yufeng, PENG Jing, BAI Yun
The study aims to develop a high-performance prediction model tailored to the blood collection and supply scenarios unique to China, considering its specific national conditions. It begins by analyzing seven factors: Workdays, holidays, weekdays, months, seasons, winter and summer vacations, and the blood collection volume from the previous day. Statistical analyses confirm that all these factors significantly influence daily blood collection volumes. Subsequently, the study proposes a CNN-LSTM hybrid model that integrates convolutional neural networks (CNN) and long short-term memory (LSTM) networks. The CNN component extracts periodic and local features from the data, while the LSTM component captures long-term temporal dependencies, enhancing feature representation capabilities. Experimental results demonstrate that the CNN-LSTM model outperforms other models, including CNN, LSTM, generalized regression neural network (GRNN), back propagation neural network (BPNN), extreme learning machine (ELM), seasonal autoregressive integrated moving average (SARIMA) and linear regression (LR). The CNN-LSTM model achieves the most comprehensive extraction of time series features across multiple factors and delivers the highest prediction accuracy. Specifically, its normalized mean absolute error (NMAE) and normalized root mean square error (NRMSE) are reduced by up to 25.80% and 26.54%, respectively, while the coefficient of determination ($R^2$) improves by up to 320.85%. The prediction results provide more precise decision-making references for blood collection and supply institutions, enabling better adjustment of collection plans and inventory management strategies.
