Abstract:

In order to improve the prediction accuracy of time series, a time series prediction method based on improved fruit fly algorithm to optimize connect inputs and outputs in long short-term memory network is proposed. Firstly, the input and output of time sets in the memory system is fully connected（Connect Inputs and Outputs in Long Short-Term Memory, CIAO-LSTM）, thus, the characterization of linear components in the target system is enhanced. Secondly, an Improved Fruit fly Optimization Algorithm（IFOA） is proposed. By dynamically changing the search radius of Drosophila and increasing the escape coefficient to the fitness function, the global optimization ability and local convergence speed of Drosophila optimization algorithm are improved. Finally, IFOA is used to optimize the CIAO-LSTM network parameters and build a predictive model（IFOA_CIAO-LSTM）. The experimental results show that the optimized time series prediction method is more generalized and more accurate than traditional long-and short-term memory networks, and can be better fitted to the data with large fluctuations.

Key words: improved fruit fly optimization algorithm, connect inputs and outputs in long short-term memory network, time series prediction, prediction accuracy

摘要：

为了提高时间序列的预测精度，提出了一种基于改进果蝇算法优化直连长短期记忆网络的时间序列预测方法。将长短期记忆网络的多个时间步输入与输出进行全连接（CIAO-LSTM，直连长短期记忆网络），增强了对目标系统中线性成分的表征。提出了一种改进的果蝇优化算法（IFOA），通过动态改变果蝇的搜索半径和对适应度函数增加逃脱系数，提高了果蝇优化算法的全局寻优能力和局部收敛速度。使用IFOA优化CIAO-LSTM网络参数并构建预测模型（IFOA_CIAO-LSTM）。实验结果表明，优化后的时序预测方法相比传统的长短期记忆网络泛化能力更强、预测精度更高，对于波动较大的数据可以实现更好的拟合。

关键词: 改进果蝇优化算法, 直连长短期记忆网络, 时序预测, 预测精度