Main steam temperature is a key parameter for the boiler of coal-fired power plants. It is difficult to remain stable under extreme load changes such as deep peak-shaving. To solve this problem

a predictive control model is added to the existing temperature loop. The proposed control system targets an ultra-supercritical boiler. A hybrid long short-term memory(LSTM) network forms the core predictor of the predictive model. A hyper parameter transfer method speeds up global optimization

avoids local optima and cuts optimization calculation amount by 88%. The predictive model gives a root-mean-square error of 0.495 ℃ and a mean absolute percentage error of 0.082%. MATLAB simulations show that the predictive control system reduces peak overshoot by 50% under extreme conditions

while preserving the control-loop stability during normal operation through multicondition piecewise control. These results demonstrate that the proposed model predictive control system satisfies the main steam temperature regulation requirements across all operating conditions.