Currently, most autonomous underwater gliders (AUGs) operate on primary lithium batteries. As the state of charge of a primary lithium battery and the influence of marine environment on the glider are difficult to measure, it is hard to forecast the energy consumption of a glider accurately, which has caused the failure of many glider missions. For the purpose of safely deploying the AUG mission and effectively optimizing the motion parameters to increase the endurance, it is very important to make an accurate energy consumption prediction model of the AUG. In this paper, a novel model based on the least squares support vector machine (LSSVM) and particle swarm optimization (PSO) algorithm, namely the LSSVM-PSO model, is proposed to forecast the energy consumption of the AUG. Considering that the kernel function and the LSSVM related parameters have a great influence on the performance of the prediction model, several LSSVM models based on different kernel functions for energy consumption prediction are established, and the parameters are optimized by the PSO algorithm. The performance of LSSVM-PSO models with different kernel functions are compared based on the sea trial data. The results indicate that the LSSVM-PSO model with a radial basis kernel function has a higher accuracy than other models for energy consumption prediction. Moreover, the performance of the LSSVM-PSO model trained by different sample sizes and that of the conventional mathematical energy consumption prediction model are compared. The results demonstrate that the LSSVM-PSO model is superior with a large enough training sample size.

当前，大多数自主水下滑翔机（AUG）都使用一次锂电池工作。由于难以测量一次锂电池的充电状态以及海洋环境对滑翔机的影响，因此难以准确预测滑翔机的能耗，这导致许多滑翔机任务失败。为了安全地部署AUG任务并有效地优化运动参数以增加耐力，建立精确的AUG能耗预测模型非常重要。本文提出了一种基于最小二乘支持向量机（LSSVM）和粒子群优化（PSO）算法的模型，即LSSVM-PSO模型，以预测AUG的能耗。考虑到核函数和与LSSVM有关的参数对预测模型的性能影响很大，建立了几种基于核函数的LSSVM模型进行能耗预测，并通过PSO算法对参数进行了优化。基于海试数据，对具有不同内核功能的LSSVM-PSO模型的性能进行了比较。结果表明，具有径向基核函数的LSSVM-PSO模型比其他模型的能耗预测具有更高的准确性。此外，比较了由不同样本量训练的LSSVM-PSO模型的性能和常规数学能耗预测模型的性能。结果表明，LSSVM-PSO模型在具有足够大的训练样本量的情况下是优越的。