This paper deals with the design of a temperature controller for a greenhouse using a wood pellet ]heating system. This heating system is carbon neutral; however it has three undesirable characteristics in control: 1) an On/Off control signal, 2) a longer time constant and 3) constraints on operation. The last property, in particular, is essential for achieving better fuel consumption and for achieving a longer lifetime of the heater. However the conventional PWM control also gives good control performance, the properties 1) and 2) block its effectiveness. In addition, the property 3) leads to performance degradation in the designed controller. To compensate for these effects, in this paper the model predictive control is applied to achieve the temperature control. The required specification is to keep the temperature error within 2 degrees Celsius in peak-to-peak with sufficient robustness for the parameter change in greenhouse. Experimental results in greenhouse give the effectiveness of proposed controller. Moreover the idea for computational load relaxation is also discussed.