Research of ALSTM-GPC in Coordinated Control System of Nuclear Power Plant

Aiming at solving the problem of poor control effect of conventional proportion integration differentiation (PID) controller when dealing with complex system, combining the advantages of deep learning in feature extraction, regression prediction and predictive control in dealing with multivariable and strong coupling, a predictive model controller is built by ALSTM deep networks, the model takes one-dimensional time series signal of object as its input, completes feature extraction through LSTM. Then, by introducing the attention mechanism, the weight parameters of LSTM extraction features are optimized, the target features are screened and retained, and the redundant features are filtered to achieve accurate extraction of effective time series features. And then a generalized predictive control (GPC) is used as a rolling optimization controller, thus building attention long short-term memory-generalized predictive controller (ALSTM-GPC). Then, the simulation experiment was carried out in the nuclear power plant's double-input and double-output multivariable coordinated control system. Through a series of simulations such as set-point disturbance, internal disturbance and external disturbance, ALSTM-GPC controller is proven having better control effect than conventional PID control.