This study is essentially aimed at improving adaptive model predictive control of batch and semi-batch chemical reactor temperature. To this end, it involves a ’software sensor’, able to estimate the rate of chemical heat production in a reliable and easy way, thanks to the use of an extended Kalman filter. Temperature measurements are used to determine the unknown parameters of on-line energy balance and kinetic equations. Thus, the instantaneous heat release rate is figured out and forecasted along the prediction horizon. The robustness and flexibility of the software sensor are examined on simulated reaction runs. The outcoming inferential controller is tested on a complex industrial protocol carried out in a pilot reactor. A comparison is made with the ’basic’ digital regulator in the same working conditions.