Abstract: In order to obtain the prediction model of compressive strength of ultra-low temperature frozen soil and explore the changes of physical and mechanical properties of frozen soil under ultra-low temperature, the uniaxial compressive strength test of −180 °C ~ −10 °C was carried out on the low liquid limit clay soil samples with water content of 19%, 22%, 25% and 28%, and the unfrozen water content of −80 °C ~ −10 °C soil samples was measured. Using the above data, a prediction model based on WOA-BP neural network and BP neural network was established to explore the relationship between moisture content, temperature, unfrozen water content and compressive strength of ultra-low temperature frozen soil. The prediction results show that there is a complex nonlinear relationship between moisture content, temperature, unfrozen water content and the compressive strength of ultra-low temperature frozen soil, especially in the range of −180 °C ~ −80 °C, the existing linear fitting formula can not accurately predict the compressive strength of frozen soil in this range. The overall prediction effect of the prediction model based on WOA-BP neural network is good. The average absolute error is 1.167 MPa and the average relative error is 7.62%. The average absolute error of BP neural network prediction model is 8.462 MPa and the average relative error is 47.99%. The prediction error of BP neural network prediction model based on whale optimization algorithm is significantly less than that of BP neural network prediction model and linear fitting value, and is closer to the measured value. The prediction model has high accuracy and can effectively solve the complex nonlinear relationship between the compressive strength of ultra-low temperature frozen soil and its influencing factors. It can provide a reference for the application of artificial freezing technology in stratum emergency engineering.