基于神经网络结构化组合的下行链路SINR预测

Downlink SINR Prediction Based on Structural Combination of Neural Networks

针对无线蜂窝系统下行链路信道SINR预测方法存在的只能对单一信道参数场景进行一步预测、预测误差较大等不足,根据信道参数变化场景下SINR序列相关性的变化,对长短期记忆网络(LSTM)和前馈神经网络(FNN)进行结构化组合,提出一种基于LSTM-FNN预测模型的SINR预测方法,并通过预训练-全局训练策略和迭代调优策略对组合网络进行较好地训练。仿真实验表明,LSTM-FNN模型相比于传统的自回归积分滑动平均模型(ARIMA)和单一FNN、LSTM网络,在信道参数变化场景下具有更好的SINR预测性能,且时间复杂度的增加在可接受范围内。

The Signal Interference Noise Rate(SINR)prediction is an effective means to mitigate the effects of channel feedback delay on the performance of wireless cellular system.However,there are some drawbacks in the existing SINR prediction methods,such as only making one-step prediction for the fixed channel parameter case,larger predictive error and so on.Aiming at overcoming these drawbacks,according to the change of SINR series correlation in different channel parameter case,a prediction method based on LSTM-FNN model is proposed.The LSTM-FNN model is constructed by combining Long Short-Term Memory network(LSTM)and Feedforward Neural Network(FNN).In the process of network training,a pre-training and a global training strategy with an iterative fine adjustment are used,which can better train the combined network.The simulation experiments show that the LSTM-FNN model outperforms the traditional ARIMA model,single FNN and LSTM network in the scenarios of the varying channel parameters,and the increase of time complexity is also within the acceptable range.