The three-dimensional (3D) acoustic intensity probe is not popularly used for the source localization of acoustic source because of the large error even though the technique has been well known. In particular, the measured direction-of-arrival strongly depends on the frequency, which is a detrimental drawback that hampers the use of this method. Such spectral bias error, due to the discrete positioning and spacing of sensors surrounding the acoustic center, fluctuates in a nearly regular spectral pattern when a 3D acoustic intensity probe configured in a tetrahedral shape is used. In this work, two error compensation methods are tried: The first method is to adopt a double-module configuration, which is formed by the connected assembly of two intensity modules, but with different locations of acoustic centers for a spatial averaging effect. The second one is the 1/3-octave band averaging of the measured intensity spectrum within the effective measurement range. Because two probe modules are attached, they concatenate 1-3 sensors in the overlap position. Four double-module probes different in assembly configuration of involved two single-modules are tested: modules having a same acoustic center but twisted by 60 degrees, inverted modules sharing one microphone and twisted by 60 degrees, symmetric modules having the joint angle of 60 degrees in the sharing edge, symmetric modules to a common face thus sharing three microphones. Measurements are conducted in an anechoic chamber by changing the source positions for 4 azimuth and 3 elevation angles. The results reveal that the maximum mean bias error of about 4 degrees can be obtained by the double-module 3D acoustic intensity method, depending on the module configuration and angular source position. This contrasts to the fluctuating direction-of-arrival within about +/- 10 degrees error range, which is obtained by each single-module in a concatenated double-module. In particular, when the band averaging technique is applied to the measured data by the probe module composed of the inverted single modules, the maximum bias error can be reduced to less than 1 degrees in both azimuth and elevation angles. (C) 2019 Elsevier Ltd. All rights reserved.