Research On Temperature Deviation Correction Of FIO-COM Forecasting System Based On LSTM Series Neural Network

In recent decades,with the deepening of people’s understanding of the ocean and the rapid development of computer science,the ability of ocean numerical model simulation and prediction has made great progress.However,compared with the actual observation,there is still some deviation between the simulation and prediction results and observation.It is still a very important work to correct the deviation of ocean numerical model prediction results based on actual measurement.Based on the forecast results of the global high-resolution Coupled wave and tidal current forecast system(FIO-COM)independently developed by the First Institute of Oceanography,Ministry of Natural Resources,Optimum Sea Surface Temperature(OISST)for multisource satellite fusion is studied for Optimum sea surface temperature.Statistical model and long and short term memory network were used to correct the prediction results of sea surface temperature(SST),and then the 3D prediction results were corrected according to the vertical distribution characteristics of temperature.Main work and conclusion of this paper:(1)By investigating the variation rule of sea surface temperature deviation over time,a single point deviation prediction model was established at each space point respectively by linear fitting and LSTM(Long Short-Term Memory)neural network method.The established single point deviation prediction model was used to forecast the deviation at the next moment.Then the sea surface temperature predicted by the model is corrected.The experiment shows that the deviation prediction models established by zero-order linear fitting,first-order linear fitting and LSTM neural network can achieve the effect of improving the forecast results of sea surface temperature,and the single point deviation prediction model established by LSTM neural network has the most stable correction effect.LSTM neural network can reduce the deviation of model sea surface temperature forecast results about 60%～70%.The comparison of the probability density distribution of temperature deviation shows that the deviation of model temperature corrected by LSTM method is concentrated around0,and the deviation decreases overall.(2)In order to reduce the calculation amount of error correction method,ConvLSTM neural network was used to establish a two-dimensional deviation prediction model of sea surface temperature deviation,which fully considered the spatial continuity of physical quantity,and realized the spatial distribution prediction of sea surface temperature deviation.The results show that ConvLSTM neural network can achieve better results than LSTM method,and can greatly reduce the training time of the model.Using the 2017-2020 data as the training set,a ConvLSTM deviation prediction model was established.Satisfactory results were obtained in the correction of sea surface temperature.After the correction,the absolute mean difference of temperature error in the test set was reduced from 0.344℃ to 0.116℃ by 66%.(3)In order to realize the error correction of the three-dimensional temperature field,based on the ConvLSTM two-dimensional error prediction model,the relationship between the temperature correction of each layer and the sea surface temperature correction was established by taking advantage of the fact that the temperature variation of the upper mixing layer was consistent with that of SST,and then the three-dimensional correction of FIO-COM temperature prediction results was realized.OISST data and Argo temperature profile observations were used to compare and analyze the temperature prediction results before and after correction.The results show that not only the deviation of sea surface temperature is greatly reduced after correction,but also the prediction results of upper ocean temperature are significantly improved after correction.