【摘要】 为了解决常用时序预测算法精度不高和调参困难的问题,提出基于多层双向长短期记忆(BiLSTM)神经网络的负载预测方法,包括网络模型设计、自适应参数设置和改进粒子群算法优化等步骤.将数据输入网络模型中进行训练,使用自适应算法进行自动调参;采用基于基准模型的多指标融合的模型评价方法,计算改进粒子群算法的适应度;使用改进粒子群算法优化模型的预测结果.通过与多种典型时间序列预测算法的实验对比,方法的预测平均绝对百分比误差减小3.6%～7.2%,训练时间缩短10%以上,实验结果验证了方法在时间序列预测中具有更高的准确性和很强的适用性,为使用负载预测结果进行弹性扩缩容提供了重要的科学依据.更多还原

【Abstract】 In order to solve the problems of low accuracy of commonly used time series prediction algorithms and difficulty in tuning parameters, a load prediction method was proposed based on a multilayer bidirectional long short-term memory(BiLSTM) neural network. The method includes network model design, adaptive parameter setting, and improved particle swarm optimization(PSO). The data was input into the BiLSTM network model for training and uses an adaptive algorithm for automatically parameter adjustment. The method uses a model evaluation method based on the benchmark model to calculate the fitness of the improved particle swarm optimization algorithm. The method uses improved particle swarm optimization to optimize the prediction results of the model.The average absolute percentage error of this method was reduced by 3.6% to 7.2% compared with several typical forecasting algorithms, and the training time was reduced by more than 10%, Through experimental comparison with a variety of typical time series prediction algorithms. Experimental results show that the method has higher accuracy and stronger applicability in time series forecasting, and provide an important scientific basis for the use of load forecast results for elastic expansion and contraction.更多还原