Knowledge-guided hybrid deep learning framework for robust early warning of food temperature deviations in dynamic urban delivery

Temperature fluctuations are one of the primary factors threatening food quality during dynamic urban delivery. However, frequent operations in cold chains of perishable foods (e.g., loading and unloading stages) can destabilize the thermal environment, resulting in reduced prediction accuracy and unreliable early warnings. To overcome the limitation, this study proposes a knowledge-guided hybrid deep learning framework that integrates physical insights of cold chain stages with a Convolutional Neural Network–Bidirectional Long Short-Term Memory (CNN–BiLSTM) model. The framework utilizes prior domain knowledge to guide multi-source data integration, feature optimization, and model design, thereby improving predictive robustness under dynamic urban delivery. Experimental validation across seven loading/unloading scenarios demonstrated that the proposed model achieved high accuracy (average RMSE of 0.31 °C) with a 30-min early warning horizon, using only one ambient temperature sensor and one door magnetic sensor. Compared with conventional models, i.e., radial basis function neural network, LSTM, CNN, the proposed framework reduced prediction errors by 71%, 22%, and 18%, respectively. Additionally, the model exhibits high computational efficiency, with an average inference time of only 66 ms per prediction. The results indicate that this hybrid knowledge-driven approach can serve as a cost-effective and reliable tool for predictive temperature management and early warning in urban delivery environments, contributing to improved food safety and reduced product losses.