基于逆模型区间优化的神经网络预测控制

针对神经网络预测控制中，在线滚动优化计算量大、算法稳定性难以保证的问题，提出一种确定黄金分割优化算法初始搜索区间的方案，即初始搜索区间的宽度与神经网络逆模型输出和上一时刻系统输入的误差成正比，二者越接近，搜索宽度就越小，从而黄金分割优化算法的在线计算量就越小；该方案有效地降低了在线滚动优化计算量，同时又使控制系统具有神经网络预测控制和神经网络逆控制的双重特性，在模型匹配稳态工况下，神经网络预测控制转化为神经网络逆控制，具有逆控制快速性的优点，而在模型失配或动态过程中，神经网络预测控制起主导作用，具有模型的宽容性和鲁棒性强的特点。采用区间套定理对该算法的收敛性给予了严格的数学证明。通过对某300MW机组仿真表明，提出的方案在控制品质和降低计算量方面均获得满意的效果。

Neutral Networks Predictive Control Using Optimizing Intervals of Inverse Models

A scheme is proposed to determine initial detecting interval by means of the golden section algorithm in order to reduce the amount and ensure the stability of optimizing algorithm in the neural networks predictive control. Namely, the width of initial detecting intervals is in proportion to the error between the output of neutral networks inverse model and the foregoing system input. The less the error is, the smaller the detecting width is and so is the on-line computing amount. The scheme effectively reduces the on-line rolling optimizing algorithm and enables the control system to feature the dualism of both neutral networks predictive control (NNPC) and neutral networks inverse control (NNIC). In steady states, NNPC is transformed to NNIC with the rapid characteristics of inverse control. In the process of model mismatch or dynamic state, NNPC plays dominant role with characteristics of model tolerance and robustness. The interval theorem offers the strict calculating proof to this algorithm. Manifested by the simulated experiment of 300MW unit, the scheme proposed yields satisfactory performance in controlling quality and reducing the amount of calculation.