Development of a high-accuracy temperature sensor for meteorological observations based on computational fluid dynamics and neural networks

Global temperatures are rising at approximately 0.1 °$C$ per decade. Existing air temperature observation sensors often measure temperatures higher than the actual values due to solar radiation effects, leading to errors of up to 1 °$C$, which significantly affects the accuracy of meteorological observations. Traditional naturally ventilated temperature sensors exhibit significant limitations in reducing radiation errors, making it challenging to meet the precision requirements of 0.05 °$C$ or even higher in atmospheric science research. To address this challenge, this paper proposes and designs a novel naturally ventilated temperature sensor. The core sensing element of the sensor employs a Pt100 thin-film platinum resistor, and its performance is optimized through computational fluid dynamics (CFD) methods and a multi-layer perceptron (MLP) network to reduce radiation errors. External field comparative experiments with the 076B artificially ventilated temperature monitoring device, using its measurements as temperature references, have validated the effectiveness of the new sensor in reducing radiation errors. Experimental results indicate that the new sensor has a root mean square error ($RMSE$) of 0.047 °$C$, a mean absolute error ($MAE$) of 0.039 °$C$, and a correlation coefficient ($r$) of 0.999. The average radiation error of the calibrated sensor is 0.03 °$C$. These findings fully demonstrate the significant advantages of this sensor in improving the accuracy of temperature measurements.