基于频分神经网络和预测控制的PID参数整定研究

Research of PID Tuning Based on Frequency_Divide Neural Network and Model Predictive Control

为了得到精确的泛化性较高的缓变非线性对象的可离线在线模型,提出了频分时滞回归径向基神经网络(FTRR)算法。此算法基于频谱分析,先把信号分解出数个频带,再构建神经网络模型。该模型用于改进的单步模型预测控制中离线求得控制输出,由此,再依据有约束线性最小二乘优化算法对PID参数进行离线整定,使其PID输出与单步模型预测控制输出相似。仿真结果表明,FTRR模型精度高且泛化性好,PID整定后的系统调节品质较高,适用于缓变控制系统。

In order to get the very generalized accurate on_off_line model of slow changing non_line object, an algorithm of frequency_divide time_delay regress RBF（FTRR） neural network was discussed. With the different frequency bands of signal by power-density spectrum analysis, a neural network model was built. The FTRR model was used in improved model predictive control, and an off_line control output was got. From this output, the PID parameter was tuned with linear least-squares in off_line status, to ensure that the two control output was very similar. Simulation shows that FTRR model is accurate and generalized, and system adjusting quality is increased after PID tuning. The model can be used in slow changing control system.