Data-driven models used for predicting soil temperature usually have increasing errors with increasing depth. By exploring the integration of knowledge-based and machine learning approaches, this study used a novel transformation of meteorological variables to increase prediction accuracy of soil temperature with increasing soil depth. Using datasets for two soil textures (silty clay loam and loamy coarse sand) at two locations with different climates, predictive models were developed for five depths (5, 10, 20, 50, and 100 cm) as a function of meteorological features using an adaptive neuro-fuzzy inference system (ANFIS). For each depth, soil temperature was predicted with nontransformation (NT), autocorrelation (AC), moving average (MA), and a combination of transformations (NT-ACMA) of meteorological features. Across all depths, the predictive accuracy of NT-ACMA models was significantly higher than that of NT, AC, and MA models for both soil textures (R 2 = .99, RMSE = 1˚C for silty clay loam; R 2 = .99, RMSE = 1.2˚C for loamy coarse sand), with increasing prediction accuracy as soil depth increases. Results for different soil textures and climates in 11 locations across the contiguous United States show that, except for 100-cm depth, there seems to be significant positive linear relationships between the best moving average and the solar inclination. This makes our NT-ACMA technique transferable to any location in the contiguous United States irrespective of the location, climate, and soil texture. We conclude that integrating lag times and moving averages of meteorological features can lead to better prediction of soil temperature at different soil depths.