自适应差分进化算法优化BP神经网络的时间序列预测

针对BP神经网络学习算法随机初始化连接权值和阈值易使模型陷入局部极小点的缺点,设计了一种自适应差分进化算法优化BP神经网络的混合算法。该混合算法中,差分进化算法采用自适应变异和交叉因子优化BP神经网络的初始权值和阈值,再用预寻优得到的初始权值和阈值训练BP神经网络得到最优的权值和阈值。首先对改进的自适应差分进化算法运用测试函数进行性能测试,然后用一个经典时间序列问题对提出的混合算法进行了检验,并与一般的神经网络、ARIMA预测模型及其它混合预测模型进行了对比,实验结果表明,本文提出的混合算法有效并且明显提高了预测精度。

BP neural network incorporating self-adaptive differential evolution algorithm for time series forecasting

It is easy for a BP neural network (BP NN) to be trapped into a local minimum point for the time series forecasting problem. To improve the forecasting accuracy, we design a hybrid algorithm which combines the self adaptive differential evolution algorithm (SDE) with the BP NN. We adopt the SDE algorithm to search for global initial weights and thresholds of the BP NN. These values are then employed to further search for the optimal weights and thresholds. The performance of the proposed SDE algorithm is verified through benchmark functions and a well known real data set is used to verify the effectiveness of the hybrid algorithm. Compared with general neural network, ARIMA and other hybrid models,experimental results indicate that the proposed algorithm can be an effective way to improve forecasting accuracy.