Multi-Step Sea Surface Wind Speed Prediction Based On Signal Decomposition And Deep Learning

With the vigorous development of ocean shipping and tourism in China,the demand for safety and economy is increasing.With the development of meteorological detection means and advanced prediction technology,it is possible to accurately predict marine meteorology and sea conditions.Accurate meteorological and hydrological forecast can provide meteorological navigation for ships,make ships avoid typhoons and bad weather areas as much as possible,and reduce the damage of wind and waves to the hull.It can also effectively use the beneficial meteorological and hydrological conditions in the navigation area,shorten the navigation time and improve the economic benefits.Among the three meteorological and hydrological factors affecting navigation,sea breeze,ocean wave and ocean current,sea breeze has the characteristics of nonlinearity,uncertainty,high volatility and complex periodicity,so it is difficult to predict.Therefore,improving the accuracy of meteorological and hydrological prediction is very important for the safety and economic guarantee of ocean going ships.Aiming at the problem of single location and regional prediction of sea surface wind speed,a multi-step prediction model based on signal decomposition and depth learning is designed in this paper.The specific contents are as follows:Firstly,the paper summarizes the current situation of wind speed prediction at home and abroad,and makes it clear that physical methods,classical time series prediction methods and artificial intelligence methods are the main means of wind speed prediction.Considering the highly nonlinear and fluctuating characteristics of sea surface wind speed data and the demand for time resolution for meteorological and hydrological support,the 5th European reanalysis data set is selected as the data source,and 24 h is selected as the input step according to the correlation analysis results.This paper analyzes the differences of various multi-step prediction strategies and neural networks in time series prediction,designs a wind speed prediction model based on gated cyclic network(Gru),and preliminarily realizes the 12 hour prediction with an interval of 1 hour.Secondly,aiming at the complex periodicity of single site sea surface wind speed data,the signal decomposition algorithm is used to improve the prediction model.Because the decomposition effect of variational modal decomposition(VMD)algorithm is easily affected by parameter selection,a VMD parameter selection method is designed,which uses center frequency and sample entropy to quantitatively evaluate the advantages and disadvantages of decomposition.The frequency domain image of VMD component shows that this method can avoid the possible negative phenomenon of VMD.Finally,the Gru prediction model is improved by using the VMD algorithm with correct parameters,and compared with the decomposition prediction model based on different decomposition algorithms and component processing mechanism.The experimental results show the superiority of the proposed method.Finally,aiming at the difficulty of multi-step pixel level prediction of regional wind speed,a spatio-temporal prediction model based on depth U-shaped convolutional neural network(resunet)is designed.The model uses the multi-step prediction strategy of leading time to input the wind speed and the known time difference information into the model at the same time,so that the prediction time of the model can be controlled.The model uses the exogenous variable of wind direction to further reduce the error.The ablation experiment proves that the resunet model,multi-step strategy and other improvement measures improve the prediction accuracy,and the comparative experiment proves the superiority of the model.The prediction results in space and time show that the model can effectively extract the spatial and time correlation,and get good multi-step wind speed prediction results.