基于多目标浣熊优化算法的双向长短期记忆神经网络预测

Bidirectional long short-term memory neural network prediction based on multi-objective coati optimization algorithm

摘要:

为了提高双向长短期记忆神经网络(BiLSTM,bidirectional long short-term memory neural network)的预测性能,针对BiLSTM存在的预测精度低、预测结果不稳定的问题,提出了一种新的多目标浣熊优化算法(MOCOA,Multi-objective coati optimization algorithm)。在浣熊优化算法(COA,coati optimization algorithm)的基础上,通过改进探索与开发算子,结合快速非支配排序与拥挤度距离计算方法建立精英浣熊保留策略,实现单目标到多目标的改进。基于所提算法,以预测均方误差(MSE,Mean square error)及预测误差方差为目标函数对BiLSTM超参数进行优化,并建立MOCOA-BiLSTM预测模型,最终实现精确稳定预测。将所提MOCOA-BiLSTM预测模型在变电工程造价数据集上进行了仿真测试,并与其他三种主流算法优化后的模型进行了对比。结果表明,所提MOCOA-BiLSTM的平均百分比误差相比与MOSSA-BiLSTM、NSGAIII-BiLSTM、MOMVO-BiLSTM分别降低了69.59%、58.43%、56.67%。

Abstract:

To enhance the predictive performance of the Bidirectional Long Short-Term Memory neural network (BiLSTM), and address the issues of low prediction accuracy and unstable results in BiLSTM, a novel Multi-Objective Coati Optimization Algorithm (MOCOA) was proposed. Building upon the Coati Optimization Algorithm (COA), the exploration and exploitation operators are improved and establish an elite coati retention strategy using a fast nondominated sorting and crowding distance calculation method, improving from single-objective to multi-objective optimization. Using the proposed algorithm, the hyperparameters of BiLSTM with Mean Square Error (MSE) and prediction error variance as objective functions, led to the establishment of the MOCOA-BiLSTM prediction model, ultimately achieving accurate and stable predictions. The proposed MOCOA-BiLSTM prediction model is simulated and tested on a substation engineering cost dataset, and comparisons are made with three other mainstream algorithm-optimized models. The results indicate that the MOCOA-BiLSTM exhibits a reduction of 69.59%, 58.43%, and 56.67% in average percentage error compared to MOSSA-BiLSTM, NSGAIII-BiLSTM, and MOMVO-BiLSTM, respectively.