Development of an Air Temperature Observation System Using a Radiation Shield and Neural Network Correction

Accurate air temperature observation requires minimizing solar radiation-induced deviations, which are strongly influenced by radiation shield performance. However, conventional shields often produce significant errors under strong solar radiation or weak ventilation. In this study, an air temperature observation system integrating a radiation shield and a backpropagation (BP) neural network-based correction method is proposed. Computational fluid dynamics (CFD) simulations were conducted to quantify radiation-induced temperature deviations under representative meteorological conditions, and the simulated dataset was used to train and test the neural network model. Initial field comparison experiments were performed using a 076B forced-ventilation system as a reference, where measured differences were treated as experimental deviations and model outputs as predicted deviations. The results show that, before correction, the proposed system exhibited a maximum deviation of 1.05 °C and a mean deviation of 0.26 °C, while the root mean square error and mean absolute error between experimental and predicted deviations were 0.30 °C and 0.23 °C, respectively. The correction significantly reduced temperature deviations, demonstrating the effectiveness of the proposed system in improving measurement accuracy.