遗传算法优化BP神经网络的大功率LED结温预测

Prediction of junction temperature for high power LED by optimizing BP neural network based on genetic algorithm

将遗传算法(GA)与BP神经网络相结合,对研发的120W LED双进双出的射流冲击水冷散热系统中LED阵列的结温进行预测。采用GA优化BP网络的权值和阈值,利用BP算法训练网络,改善了单独使用BP网络容易陷入局部极小值和收敛速度慢的缺点。并且在训练过程中为了使网络输出有足够长的空间,改进了GA的数据处理。结果表明,经GA优化的BP神经网络较使用Levenberg-Marquardt(LM)算法优化的BP神经网络的大功率LED结温预测精确度提高了14.14%,且预测效果较稳定。GA和BP神经网络相结合的结温预测模型较传统的结温测量方法更能掌握散热结构设计的主动性,对大功率LED寿命的延长有较高的实用价值。

In this paper,the junction temperature of high-power LED array of th e LED 120W double inlet and outlet jet impingement water cooling system developed by the research group is predicted by combining t he genetic algorithm （GA） with BP neural network. Taking advantage of genetic algorithm to optimize the weights and threshold of B P network and BP algorithm for training the network can reduce the shortcomings of local minimum value and slow convergence spe ed of using BP network alone.And in the training process,in order to make the network output have space a long enough on ge netic algorithms for data processing,we make some improvements to the original data normalized to [0.050.95].The collected data is studied,trained and forecast by the model and the results show that the model can reflect the junction temperature of LED better.T he prediction accuracy is improved by 14.14% using the GA-optimized BP neural n etwork than that using the LM-optimized BP neural network.and the predicted effect is more stable.The junction temperature prediction model o f the BP neural network combined with genetic algorithm is more able to grasp the initiative of the heat dissipation structure design than the traditional measurement,and it has high practical value to extend the life of the high power LED.