The current ammonium perchlorate water content detection methods are difficult to meet the rapid and safe detection requirements. The time domain spectra of samples were first measured by the terahertz time domain spectroscopy technique and then the refractive index and the absorption coefficient were extracted. According to the Lambert-Beer law and the effective medium theory, the water content prediction model based on absorption coefficients was constructed. Simultaneously, according to the effective medium theory and the equal optical paths, the water content prediction model based on refractive indexes was constructed. These two prediction models both need to weigh the samples. In addition, this paper proposed another new water content prediction model based on absorption coefficients combined with refractive indexes, which does not need to weigh the samples under the condition of ensuring the measurement accuracy. The experimental results showed that when the water mass fractions of the samples in the same group were below 1%, the determination coefficients between water content and absorption coefficient or refractive index were both greater than 0.995. In contrast, when their water mass fractions were below 0.1%, the determination coefficients were around 0.95. The root mean square errors of prediction for these three models were all within 0.05%. The baseline shift between the predicted and measured water contents for different groups of samples is probably caused by the difference in sample preparation. The study shows that the proposed method provides a rapid, nondestructive, safe and high precision detection method for the detection of water contents in ammonium perchlorate and similar solid powder.