A Depth-averaged Current Velocity Time Series Prediction Method For Underwater Glider Course Correction

Underwater glider is a new type of unmanned underwater vehicle,which can float and dive and adjust its attitude by changing the buoyancy and the position of the center of mass.Underwater glider can only communicate with satellite to determine its position information when surfacing,while its position underwater can only be determined by course projection.Due to the influence of ocean currents,the intended motion of the underwater glider will drift and it is difficult to maintain the planned course,resulting in the deviation between the projected course and the actual position.this paper takes "Petrel-L" underwater glider as the research object to study the deviation and correction of its course when the ocean environment changes.In order to ensure the accurate course keeping capability of the underwater glider and to serve the needs of ocean observation,the paper adopted machine learning and other related methods to model and predict the local depth-averaged current and to study the course correction and local path planning methods.The main research contents of the paper are as follows.Based on the motion characteristics of the underwater glider,a short-data streamlined deep-averaged flow model was established by combining the shore-based operation and control conditions.The model took into account the motion principle of the underwater glider and embeds the commands of the shore-based control platform in the face of the quasi-real-time requirements of shore-based control,which effectively retains the construction accuracy of the traditional depth-averaged current,simplified the operation and improves the calculation speed.It is of reference value for obtaining the operation status of underwater gliders,analyzing the characteristics of marine environment and regional ocean modeling.Based on the short data of underwater gliders to streamline the depth-averaged current time series data,two different machine learning methods,least squares support vector machine(LSSVM)and long-short time memory neural network(LSTM),were used to predict the future depth-averaged current data.The impact of machine learning models with different dimensional input matrices on the prediction results was considered to determine the correlation dimension of depth-averaged current on the time scale,and predictions based on variational mode decomposition(VMD)of LSSVM and LSTM were performed,respectively.The predictions were compared with those of empirical mode decomposition(EMD)prediction methods,and the results showed that the LSSVM methods based on VMD have better prediction results,and further obtained the optimal profile association dimension.A study related to the adaptive VMD based on the prediction method of LSSVM considering the optimal profile correlation dimension was conducted.Three methods of VMD layer determination were involved,namely,empirical method,VMD layer determination based on EMD,and VMD layer determination based on permutation entropy(PE-VMD),which were validated by sea trial data.The results showed that the use of PE-VMD reduced the instability of the sequence to a greater extent and further improved the accuracy of prediction.Using the above-mentioned method of determining the number of layers of VMD based on the permutation entropy(PE),the research of underwater glider course correction was carried out in combination with the method of LSSVM.In the face of further meeting the requirements of underwater glider course correction and local path planning,active course adjustment was carried out based on the predicted depthaveraged current,combined with the working operation mode of the underwater glider;sea trial tests were carried out to verify the accuracy of the algorithm.