Thermocline characteristic parameter forecasting method based on ConvLSTM model

Abstract

The invention aims to provide a thermocline characteristic parameter forecasting method based on a ConvLSTM model, which belongs to the field of ocean and comprises the following steps of: acquiring seawater temperature gridding reanalysis data from a Columbian ocean observation service data source, and performing vertical refinement processing on the data; calculating thermocline features of the corresponding sea area, and obtaining thermocline feature historical data; determining a data set duration, a prediction model data input duration, a prediction model data output duration and other parameters, filtering and dividing a thermocline feature historical data set, and constructing a thermocline prediction sample set; building and training a ConvLSTM thermocline forecasting model; and inputting test set data into the ConvLSTM prediction model obtained by training to obtain thermocline feature data, and evaluating a prediction result. According to the method, the prediction precision of thermocline features is remarkably improved. Accurate prediction of thermocline upper and lower boundary depths and thermocline strength characteristics in a sea area with thermocline is realized.

Description

Technical Field

[0001] The present invention relates to a seawater forecasting method, in particular to a seawater thermocline forecasting method. Background Art

[0002] The marine environment, especially the underwater environment, is complex and changeable. The vertical structure of seawater temperature in different sea areas is often different. Influenced by factors such as sea surface climate and ocean currents, the vertical structure of seawater temperature will be stratified. This seawater temperature stratification phenomenon is called thermocline. Thermocline is a key element in the current marine environmental analysis and an important factor in studying the adaptability of marine organisms to the environment. Thermocline affects the patrol routes of fish. In fishery, the distribution of thermocline is an important reference for improving fishing efficiency.

[0003] At present, in the research on the prediction method of thermocline, it is difficult to accurately calculate the characteristics of thermocline because there is no direct channel to obtain the characteristic parameter data of thermocline, and the existing thermocline characteristic calculation methods have limited applicability. The existing methods all use numerical prediction models to first predict the ocean temperature field data, and then invert the thermocline according to the predicted temperature field, and then predict the characteristics of the ocean thermocline. This prediction model will produce errors in the temperature field prediction link and the thermocline characteristic inversion link. The superposition of the two errors will cause the final prediction result of the entire prediction process to be too low in accuracy and have no reference value. Therefore, how to achieve direct prediction of thermocline characteristics and thus improve its prediction accuracy remains to be solved.

[0004] Existing forecasting methods are divided into two forecasting modes, namely numerical forecasting and statistical forecasting. Numerical forecasting is a method that sets appropriate initial values ​​and boundary values ​​as constraints, constructs a corresponding mathematical model based on the physical laws of the research object itself, and then relies on mathematical theory analysis supplemented by the powerful computing power of computers to meet the forecasting needs. Statistical forecasting refers to the method of applying probability theory and mathematical statistics to the data or elements required for forecasting to analyze and forecast. It is difficult to establish a unified and applicable mathematical model for the calculation of the characteristic parameters of the thermocline, and it is impossible to use numerical forecasting methods to directly forecast the characteristic parameters of the thermocline. Statistical forecasting pays more attention to the statistical significance contained in the constructed forecast object model, does not refer to the physical model of the object, and thus has no physical limit. For extended forecasts of more than ten days, statistical forecasting plays an irreplaceable role. Therefore, the forecasting method of thermocline characteristic parameters based on statistical forecasting methods can directly forecast the characteristics of the thermocline, while avoiding the application of complex physical models of the thermocline, and has good research prospects and application value. Summary of the invention

[0005] The purpose of the present invention is to provide a method for predicting thermocline characteristic parameters based on a ConvLSTM model, which can solve the problem that the current numerical prediction method cannot directly predict the thermocline characteristics, resulting in low prediction accuracy.

[0006] The object of the present invention is achieved in that:

[0007] The present invention provides a method for predicting characteristic parameters of thermocline based on ConvLSTM model, which is characterized by:

[0008] (1) Obtain seawater temperature gridded reanalysis data from the Copernicus Ocean Observing Service (CMEMS) data source and use the Akima interpolation method to vertically refine the data;

[0009] (2) Calculate the thermocline characteristics of the corresponding sea area based on seawater temperature data, i.e., the upper and lower boundary depths of the thermocline and the thermocline strength, and obtain historical data on thermocline characteristics;

[0010] (3) Determine the duration of the data set, the duration of the prediction model input data, the duration of the prediction model output data, and other parameters, filter and divide the historical data set of thermocline characteristics, and construct a thermocline forecast sample set;

[0011] (4) Construction and training of ConvLSTM thermocline forecast model;

[0012] (5) The test set data is input into the trained ConvLSTM forecasting model to obtain the thermocline characteristic data and evaluate the forecast results.

[0013] The advantages of the present invention are:

[0014] 1. The present invention adopts a statistical forecasting method to directly forecast the characteristics of the thermocline. Compared with the traditional method of first predicting the seawater temperature by numerical forecasting and then calculating the characteristics of the thermocline by the seawater temperature, the prediction accuracy is higher and the error accumulation in multiple processes is avoided.

[0015] 2. The present invention performs standardized processing on the historical data of thermocline characteristics, highlights and amplifies the fluctuation characteristics in the data, and improves the training efficiency of subsequent prediction models.

[0016] 3. The present invention adopts the ConvLSTM model to realize the prediction of thermocline characteristics. The ConvLSTM model simultaneously learns the time and space characteristics of the thermocline characteristics, has a powerful modeling ability in spatiotemporal prediction, and is suitable for processing complex spatiotemporal data scenarios.

[0017] The statistical forecasting method is used to directly forecast the characteristics of the thermocline, avoiding the use of complex physical models of the thermocline like the numerical forecasting method, which significantly improves the forecast accuracy of the characteristics of the thermocline. It realizes the accurate prediction of the upper and lower boundary depths and the intensity characteristics of the thermocline in the sea area with thermocline, provides a reference for the underwater environment for the safe navigation of underwater submersibles, and helps marine fisheries to master fish migration, which has high practical application value.

[0018] The invention was verified using five years of thermocline characteristics data obtained from the Copernicus Ocean Observing Service (CMEMS) in the regional range of 128°E~133°E, 10°N~15°N, with a horizontal resolution of 0.25°×0.25°. The forecast accuracy exceeded 94%, exceeding the forecast accuracy of less than 80% of existing numerical forecast methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a schematic diagram showing the difference between the functional structures of ConvLSTM and LSTM;

[0020] Figure 2 This is a working diagram of the encoding module and prediction module of the ConvLSTM model;

[0021] Figure 3 This is a schematic diagram of the network structure of the ConvLSTM model;

[0022] Figure 4 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0023] The present invention is described in more detail below with reference to the accompanying drawings:

[0024] Combination Figure 1-4 , the forecasting method based on the ConvLSTM model adopted by the present invention belongs to a supervised deep learning method. The ConvLSTM model can process the time and space dimensions of data simultaneously. The traditional LSTM can only capture the long-term and short-term dependencies in the time series, while the ConvLSTM can extract the spatial features of the input data by introducing the convolution operation, and can retain and transfer these spatial features in the time dimension, so that the model is more effective in processing spatiotemporal data. Therefore, for tasks that need to consider the relationship between time and space at the same time (such as video processing, weather forecasting, etc.), ConvLSTM has significant advantages. In view of this, the present invention proposes a thermocline feature forecasting method based on the ConvLSTM model.

[0025] The ConvLSTM network structure used in the present invention is as follows: Figure 4As shown in the figure, the ConvLSTM network is divided into an encoding module and a decoding module according to its function. The encoding module consists of a structure connected by two layers of convolution and ConvLSTM, which can efficiently extract spatiotemporal information from the input thermocline feature data. At the same time, the ConvLSTM unit can ensure the long-term spatiotemporal correlation of the input and prevent the occurrence of gradient vanishing and gradient explosion during training. The decoding module structure is symmetrical to the encoding module, consisting of a structure connected by two layers of deconvolution and ConvLSTM, which can further predict the thermocline feature data in combination with the spatiotemporal features extracted by the encoder. At the same time, deconvolution converts the predicted output of the network model into a data structure with the same format and dimension as the input.

[0026] The present invention is achieved through the following steps:

[0027] Step 1: Obtain seawater temperature gridded reanalysis data from data sources such as the Copernicus Ocean Observing Service (CMEMS), and use the Akima interpolation method to vertically refine the data;

[0028] Step 2: Calculate the thermocline characteristics of the corresponding sea area based on the seawater temperature data, i.e., the upper and lower boundary depths of the thermocline and the thermocline strength, and obtain the historical data of the thermocline characteristics;

[0029] Step 3: Determine the duration of the data set, the duration of the prediction model input data, the duration of the prediction model output data and other parameters, filter and divide the thermocline characteristic historical data set, and construct a thermocline forecast sample set;

[0030] Step 4: ConvLSTM thermocline forecast model construction and training.

[0031] Step 5: Input the test set data into the trained ConvLSTM forecasting model to obtain the thermocline characteristic data and evaluate the forecast results.

[0032] The factors that affect the accuracy of the classifier are as follows:

[0033] (1) Model architecture: The number of ConvLSTM layers and the number of units in each layer need to be adjusted according to the actual training prediction accuracy; the convolution kernel size affects the extracted spatial features, which in turn affects the prediction progress.

[0034] (2) Hyperparameters: A learning rate that is too high or too low will affect the convergence speed and final model performance; the choice of batch size will affect the stability and convergence speed of training.

[0035] (3) Training strategy: Sufficient training rounds can improve model performance, but too many rounds may lead to overfitting; choosing a suitable optimization algorithm (such as Adam, SGD, etc.) will affect the training effect.

[0036] (4) Regularization: Dropout can prevent overfitting and improve the generalization ability of the model; L2 regularization helps control model complexity.

[0037] Example:

[0038] Step 1: Obtain three thermocline characteristic data of the upper boundary, lower boundary, and thermocline intensity of the specified area for a certain period of time to create a training set. Assuming that the regional grid size is 20×20 and the historical data duration is 5 years, a total of 1827 days, the historical data size is 1827×20×20.

[0039] Step 2: Create training and test sets based on the input and output data duration of the forecast model. Assuming that the model inputs 7 days of data and outputs 3 days of data, that is, the data for the next 3 days is forecasted based on the 7 days of data, the input size of a single sample is 7×20×20, and the sample label size is 3×20×20. According to the sample format, data samples are created in chronological order, and 1817 samples can be obtained from 1827 days of historical data. The sample set is used to create training and test sets in a 4:1 ratio, so the training set has 1454 samples and the test set has 363 samples.

[0040] Step 3: Put the training set and test set into the ConvLSTM model for training. In each training iteration, calculate the Loss based on the network prediction results and sample labels, update the network parameters based on the Loss, and calculate the Loss and ACF of the model prediction of the validation set after the updated parameters to evaluate the model performance.

[0041] The Loss uses the L2 norm, and the calculation method is the square root of the sum of the squares of the difference between the predicted value and the true value.

[0042]

[0043] ACF can characterize the accuracy of the prediction results of the network model. The closer the ACF value is to 1, the higher the prediction accuracy of the model. ACF comes from the correlation between the prediction results and the corresponding prediction model input time series. The prediction model with a small prediction loss may not have a corresponding ACF close to 1, but when the ACF predicted by the network model is closer to 1, it means that the training effect of the network model is better. Suppose the time series data matrix X t , lag step length k, then the ACF calculation formula is as follows:

[0044]

[0045] For a prediction model, the goal of updating the network parameters is to make the model prediction result Loss approach 0 and ACF approach 1.

Claims

1. A method for predicting characteristic parameters of thermocline based on ConvLSTM model, characterized by: (1) Obtain seawater temperature gridded reanalysis data from the Copernicus Ocean Observing Service (CMEMS) data source and use the Akima interpolation method to vertically refine the data; (2) Calculate the thermocline characteristics of the corresponding sea area based on seawater temperature data, i.e., the upper and lower boundary depths of the thermocline and the thermocline strength, and obtain historical data on thermocline characteristics; (3) Determine the duration of the data set, the duration of the prediction model input data, the duration of the prediction model output data, and other parameters, filter and divide the historical data set of thermocline characteristics, and construct a thermocline forecast sample set; (4) Construction and training of ConvLSTM thermocline forecast model; (5) The test set data is input into the trained ConvLSTM forecasting model to obtain the thermocline characteristic data and evaluate the forecast results.

Images