摘要
为了提升海浪波高预测精度,提出了融合多重分解和差值修正的海浪波高预测模型(J-DE-LSTM)。该模型采用自适应噪声完备集合经验模态分解,对波高数据进行一重分解,以及对分解后的残差分量进行二重分解;采用亲和力传播算法进行聚类降维并输人到长短期记忆网络进行预测获取初步预测值。建立波高观测值与初步预测值形成的差值序列进行三重分解,采用样本熵重构为趋势项和周期项并进行权重计算,构建粒子群算法优化极限学习机和LSTM的组合预测模型进行双轨并行预测;最后将预测结果与权重加权融合进行差值修正未来点位波高预测值。实验结果表明J-DE-LSTM模型较LSTM、TCN模型平均绝对误差提升约4.1%~11.5%,均方误差提升6.5%~15.2%。
In order to improve the accuracy of wave height prediction,this paper proposes a wave height prediction model(J-DE-LSTM)integrating fusion multiple decomposition and difference correction.This model utilizes Complete Ensemble Empirical Mode Decomposition with Adaptive Noise(CEEMDAN)to perform a single decomposition on the wave height data,and the residual components after decomposition are double decomposed.Using Affinity Propagation algorithm(AP algorithm)to cluster dimensionality reduction,preliminary prediction values are predicted through inputting it into the Long Short Term Memory Network(LSTM).Then a triple decomposition of the difference sequence are established between observation values of wave height and the preliminary prediction values.The trend term and period term are reconstructed using Sample Entropy(SampEn),and used to calculate the weights.A Particle Swarm Optimization(PSO)is constructed for Extreme Learning Machine(ELM)and LSTM combined prediction model used for dual track parallel prediction.Finally,the predicted results wil be with weighted weights for difference correction of point wave height predictions.The experimental results show that the Mean Absolute Error(MAE)of the J-DELSTM model is increased by about 4.1%~11.5%compared to LSTM and TCN models,and the Mean Square Error(MSE)is increased by 6.5%~15.2%.
作者
卢鹏
姜星竹
王振华
郑宗生
LU Peng;JIANG Xingzhu;WANG Zhenhua;ZHENG Zongsheng(College of Information Technology,Shanghai Ocean University,Shanghai 201306,China)
出处
《海洋测绘》
CSCD
北大核心
2024年第2期36-40,共5页
Hydrographic Surveying and Charting
基金
上海市科委科研计划项目(20dz1203800)。
作者简介
卢鹏(1981-),男,河南洛阳人,副教授,博士,主要从事智能信息处理研究。
