In this paper, a capacity-achieving or lower-bound achieving constellation optimization algorithm is proposed for multi-input single-output (MISO) visible light communication (VLC) systems to minimize the power consumption under the intensity constraints and dimming demands for individual light emitting diode (LED). The capacity of the MISO VLC systems is upper bounded by the capacity of a corresponding single-input single-output (SISO) VLC system, which is achievable unconditionally if all LED groups are confined by the same dimming coefficient. For inconsistent dimming demands of LEDs, a necessary condition is put forward to reach this upper bound, while it can not be meet, a capacity lower bound and lower bound achieving signal input can be found. Since the capacity-achieving or lower-bound-achieving input distributions are generally not unique, the extra electrical power consumption for communication can be minimized among all capacity-achieving or lower-bound-achieving constellation distributions. Finally, simulation results show that the capacity lower bounds approaches their upper bounds and communication power saving by the constellation optimization algorithm gets larger as SNR increases.