A method to improve the real time predictions of ocean currents on the basis of a machine learning technique called model tree is proposed. It consists of forming an error time series obtained as the difference between the numerical prediction and the actual measurement of the current at a given time step, carrying out time series prediction as per the technique of model tree and predicting the error for a future time step. Subtraction of such error from the numerically predicted current produces the improved current magnitude for the next time step. The suggested procedure is applied at two deepwater locations in the Indian Ocean. The numerical current model under investigation is code named: HYCOM, while corresponding current observations are those coming from a measurement program called: RAMA. It was found that such method of error subtraction yielded more accurate predictions than those based only on the numerical modelling. This is judged from analysing certain error statistics as well as by comparison with the random walk time series prediction method. The predictions up to five days in advance are satisfactorily done in this manner.

中文翻译：

---

使用模型树改善数值电流预测

提出了一种基于称为模型树的机器学习技术来改进洋流实时预报的方法。它包括形成一个误差时间序列，该误差时间序列是作为给定时间步长的数值预测与实际电流测量之间的差而获得的，根据模型树的技术执行时间序列预测，并预测未来时间步长的误差。从数字预测的电流中减去这种误差会为下一时间步提高电流幅度。建议的程序适用于印度洋的两个深水位置。正在研究的数字电流模型的代号为：HYCOM，而相应的电流观测值则来自于名为RAMA的测量程序。已经发现，这种误差相减方法比仅基于数值建模的方法能产生更准确的预测。这是通过分析某些错误统计数据以及与随机游走时间序列预测方法进行比较来判断的。以这种方式令人满意地完成了最多五天的预测。