The improved self-tuning regulator is successfully developed here for the water level control in the steam generators of the nuclear power plant. It uses a recursive least-squares estimator with a variable forgetting factor and the bounded covariance matrix to obtain the estimated parameters for both slow and sudden changes of the steam flow rates. The present methodology enables us to avoid one of the major difficulties associated with the constant exponential weighting of past data, namely, ``wind-up`` of the covariance matrix of the estimates, and ``bursting`` phenomenon in the covariance values when the input is not persistently excited. The experimental evaluation of the self-tuning regulator reveals the simplicity of the adaptive algorithm and its excellent performance. Experimental results confirm that the present self-tuning regulator outperform a well-tuned PI controller when a setpoint is changed and the mismatch or degradation of sensors happens.