基于改进樽海鞘优化BP神经网络的土壤墒情预测

Prediction of soil moisture based on BP neural network optimized by adaptive salp swarm algorithm

针对传统的土壤墒情预测方法精度较低、训练周期长的问题,本文对BP神经网络预测模型进行研究,提出一种改进樽海鞘算法优化BP神经网络的预测方法。首先,在标准樽海鞘群算法(Salp Swarm Algorithm,SSA)中引入变异算子增强种群的多样性,提高算法的全局探索能力;同时,采用动态权重调整策略增强局部开发性能,改善收敛速度,在位置更新过程中加入动态权重,进一步平衡全局探索和局部开发能力;其次,考虑到BP预测网络收敛精度低、易陷入局部最优等缺点,将改进樽海鞘算法引入到BP中形成ASSA-BP的预测模型算法,该算法缩短了训练时间、提高了预测精度。最后,将ASSA-BP与PSO-BP、BP不同预测模型进行对比,结果表明ASSA-BP的最优预测相对误差平均值3.37%,绝对误差平均值0.0258,比BP模型预测误差有所下降。克服了BP预测模型收敛精度低、易陷入局部最优的缺点,具有更好的鲁棒性和预测精度。

There exists two problems:low accuracy and long training term in the traditional prediction methods of the soil moisture.The BP neural network prediction model is studied,a method of optimizing BP neural network improved from the adaptive salp swarm algorithm is presented.First of all,introducing the mutation operator into the standard salp swarm algorithm(Salp Swarm Algorithm,SSA)to enhance the diversity of population and improve the ability to global search;At the same time,the dynamic weight adjustment strategy is adopted to enhance local development performance and improve convergence speed.Adding dynamic weights in the location updating process to further balance global exploration and local development capabilities.Secondly,considering the shortcomings of BP prediction network such as low convergence accuracy and easy to fall into local optimum,adaptive salp swarm algorithm is introduced into BP to form an ASSA-BP prediction model algorithm,which shortens the training time and improves the prediction accuracy.Finally,the ASSA-BP,PSO-BP,BP are compared with other different prediction models,which shows that the ASSA-BP prediction model of the optimal prediction relative error is 3.37%,the absolute error of 0.0258,down from the BP model prediction error.This method not only is without the shortcomings of the easy to fall into local optimum and slow convergence,but also possesses a better robustness and prediction accuracy.