基于深度学习的组合体航天器模型预测控制

利用模型预测算法先预测控制结果后控制的类人行为特点，借助深度学习在多参数寻优上的优势，提出了一种基于卷积神经网络的模型预测控制算法，满足航天工程低硬件需求，实现组合航天器多场景下姿态控制律的重构。该算法首先利用模型预测控制将组合航天器从初始状态控制到预期状态，然后将控制过程中状态量用于3层3核卷积神经网络的训练，训练完成后，用该卷积神经网络代替模型预测对组合航天器进行控制，从而降低计算资源需求。仿真校验表明：该算法可预测5个控制周期内的控制参数，相比传统模型预测算法所需硬件计算时间降低约5倍，在一般硬件环境下30 s内即可完成各场景下的组合航天器姿态控制，控制精度在10 -4 量级。

Model Predictive Control of Combined Spacecraft Based on Deep Learning

 To solve the problem of the attitude control reconstruction in the combined spacecraft, a predictive model control algorithm based on the convolution neural network is proposed. By the advantage of deep learning in the multi-parameter optimization, this algorithm has the characteristic of human-like behavior of prediction before control. It is suitable for the aerospace application scenarios with low hardware requirements. Firstly the algorithm uses the predictive model to control the combined spacecraft from the initial state to the expected state. Then the state variables in the process are used in the training of the three-layer convolutional neural network. When the training is completed, the convolutional neural network is used to replace the model prediction algorithm to control the combined spacecraft. This reduces the need for hardware performance. The simulation results show that the algorithm can predict the control parameters within five control cycles and can be reduced by about 5 times of hardware calculation time compared with the traditional model prediction algorithm. The attitude control of the combined spacecraft is completed within 30 seconds, and the control accuracy is about 10 -4 order of magnitude.