The key issue of three-dimensional active noise control (3D ANC) problems is that global control is generally difficult, given limited number of discrete sensors. In this paper, feedforward multi-channel ANC approach is proposed to circumvent this difficulty. In view of the model-matching principle and multiple secondary sources, an underdetermined multi-channel inverse filtering (UMIF) system is formulated. With this UMIF system as a design constraint, a cost function is introduced to minimize the noise energy at a large number of control points. This linearly constrained minimum variance (LCMV) proves effective in broadening the controlled area in a 3D space. Optimal deployment of control points and the regularization terms of LCMV approach are also examined. To implement the proposed ANC system in a non-freefield environment, sensor interpolation can be used to find the frequency response between control points and loudspeakers, with plane wave decomposition and some room response measurements. The proposed ANC system has been implemented on a six-element linear loudspeaker array. Simulation and experiment results have demonstrated that the propose approach has yielded significant noise reduction performance in a large control area.