A method for predicting Argo float trajectory based on convolution and decision tree

Abstract

The present invention discloses a method for predicting the trajectory of an Argo float based on convolution and a decision tree, comprising: obtaining the coordinate data of an Argo float every m hours within a preset time period, and forming a coordinate data set with the obtained coordinate data; preprocessing the coordinate data set to obtain Argo float trajectory feature data; inputting the Argo float trajectory feature data into a buoy trajectory prediction model, and outputting the coordinate position data of the Argo float one day after the preset time period; the method can greatly improve the accuracy of the prediction of the trajectory of the Argo float.

Description

Technical Field

[0001] The invention belongs to the field of marine science and technology, and in particular to an Argo buoy trajectory prediction method based on convolution and decision tree. Background Art

[0002] The Argo program (Array or Real-time Geostrophic Oceanography) deploys a satellite tracking buoy every 300 kilometers around the world, aiming to quickly, accurately and widely collect seawater temperature and salinity profile data in the upper layer of the global ocean. Seawater temperature and salinity profile data play an important role in climate change, ocean heat storage, ecological environment, national defense, military and aquaculture. Increasing the density of observations can improve the accuracy of climate forecasts and effectively prevent the threats posed to humans by increasingly severe climate disasters around the world. The diversity and differences in Argo buoy data lead to a large number of conflicts and inconsistencies in ocean data, and its long-term and short-term spatiotemporal changes and distribution are also different. Accurately predicting the future status of Argo buoys faces many challenges.

[0003] The trajectory of the Argo float not only reflects the ocean environment, but also has a key impact on target tracking and monitoring. Therefore, it is very important to predict the trajectory of the Argo float. Since the trajectory of the Argo float is highly random, and its nonlinear and non-stationary characteristics are very obvious, there is currently no clear, feasible and high-precision method to effectively predict the trajectory of the Argo float. There are three existing methods for predicting the trajectory of the Argo float: (1) Statistical prediction method: using probability theory and mathematical statistics to predict future changes in the trajectory of the Argo float; (2) Numerical model method: based on a series of partial differential equations such as the seawater motion equation and the thermodynamic equation to describe the physical state of the seawater, the equations are solved by given initial conditions, boundary conditions and external forcing factors to obtain future changes in the trajectory of the Argo float.

[0004] Therefore, how to effectively predict the trajectory of the Argo float with high precision has become a key issue in current research. Summary of the invention

[0005] In view of the above problems, the present invention provides an Argo float trajectory prediction method based on convolution and decision tree, which at least solves some of the above technical problems. The method can greatly improve the accuracy of Argo float trajectory prediction.

[0006] The embodiment of the present invention provides an Argo float trajectory prediction method based on convolution and decision tree, comprising:

[0007] S1. Acquire the coordinate data of the Argo float every m hours within a preset time period, and form a coordinate data set with the acquired coordinate data;

[0008] S2, preprocessing the coordinate data set to obtain Argo buoy trajectory feature data;

[0009] S3. Input the Argo buoy trajectory feature data into a buoy trajectory prediction model, and output the coordinate position data of the Argo buoy one day after the preset time period.

[0010] Furthermore, the S2 specifically includes:

[0011] S21, removing abnormal data in the coordinate data set;

[0012] S22, performing standardization processing on the coordinate data set after removing abnormal data;

[0013] S23. Use the CNN convolutional neural network algorithm to select the Argo buoy trajectory feature data from the standardized coordinate data set.

[0014] Furthermore, in S23, the Argo float trajectory characteristic data includes:

[0015] The distance values ​​between different coordinate data, and the direction change data of each coordinate data after being affected by ocean currents and air pressure.

[0016] Furthermore, the buoy trajectory prediction model is trained and generated by the following steps, including:

[0017] Within n days, the coordinate data of the Argo float are obtained every m hours, and the obtained coordinate data are combined into a coordinate data set;

[0018] Preprocessing the coordinate data set to obtain Argo buoy trajectory characteristic data;

[0019] The Argo float trajectory feature data of the first n-1 days within the n days is used as input, and the Argo float trajectory feature data of the nth day is used as output to train and generate a float trajectory prediction model.

[0020] Furthermore, the buoy trajectory prediction model includes an SRU deep learning prediction model and a decision tree model; the SRU deep learning prediction model and the decision tree model are effectively combined through a weighted average method.

[0021] Furthermore, the SRU deep learning prediction model is a multi-layer SRU parallel optimization deep learning prediction model based on convolution, and the corresponding optimization steps include:

[0022] Performing linear transformation on multiple neural networks of unequal dimensions, and merging the multiple neural networks of unequal dimensions into one matrix;

[0023] All element-by-element multiplications are performed through the kernel function;

[0024] The loop statements in the kernel function are processed based on the sequence order; and the loop statements outside the kernel function are processed in parallel through gridding and parallel computing methods.

[0025] Furthermore, the buoy trajectory prediction model uses a hyperbolic tangent function as an activation function and has two hidden layers and one output layer.

[0026] Compared with the prior art, the Argo float trajectory prediction method based on convolution and decision tree described in the present invention has the following beneficial effects: The method provided in the embodiment of the present invention realizes high-precision prediction of Argo float trajectory data. The method provided in the embodiment of the present invention only needs to use the long-term series data of Argo float trajectory, and does not need to use other data. The method provided in the embodiment of the present invention occupies less resources and has a fast calculation speed.

[0027] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description, claims, and drawings.

[0028] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0030] Figure 1 A flow chart of an Argo float trajectory prediction method based on convolution and decision tree provided in an embodiment of the present invention.

[0031] Figure 2 A flow chart of the buoy trajectory prediction model training provided in an embodiment of the present invention.

[0032] Figure 3 A schematic diagram of the SRU deep learning network structure provided in an embodiment of the present invention.

[0033] Figure 4 A schematic diagram of the training results of Argo float trajectory time series data using an ensemble method provided in an embodiment of the present invention.

[0034] Figure 5A training loss function diagram of the SRU deep learning prediction model provided in an embodiment of the present invention.

[0035] Figure 6 A comparison chart of the predicted results and true values ​​of the Argo float trajectory time series provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0036] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0037] See also Figure 1 As shown, the embodiment of the present invention provides an Argo float trajectory prediction method based on convolution and decision tree, which specifically includes the following steps:

[0038] S1. Acquire the coordinate data of the Argo float every m hours within a preset time period, and form a coordinate data set with the acquired coordinate data;

[0039] S2, preprocessing the coordinate data set to obtain Argo buoy trajectory feature data;

[0040] S3. Input the Argo buoy trajectory feature data into a buoy trajectory prediction model, and output the coordinate position data of the Argo buoy one day after the preset time period.

[0041] The above steps are described in detail below.

[0042] In the above step S2, the coordinate data set is preprocessed, mainly including a series of preprocessing work such as filling in missing, removing anomalies (i.e., removing data of types such as data format errors, trajectory points located on land or islands, and the distance between two trajectory points exceeding the maximum preset value), standardization and feature selection, etc., in order to provide more effective feature information for model prediction; in an embodiment of the present invention, a CNN convolutional neural network algorithm is used to select Argo buoy trajectory feature data from the coordinate data set after standardization, and based on the CNN's ability to capture global and local features, in-depth mining of the change rules and connections of multiple trajectory points is achieved; thereafter, a variety of classic deep learning methods are used to fit the Argo buoy trajectory feature data respectively; wherein, the Argo buoy trajectory feature data includes the distance values ​​between different coordinate data, and the change data of the direction of each coordinate data after being affected by ocean currents, air pressure, wind, etc.

[0043] In the above step S3, the buoy trajectory prediction model is trained and generated by the following steps, referring to Figure 2 As shown, it includes: first, within n days, the coordinate data of the Argo float is obtained every m hours, and the obtained coordinate data is formed into a coordinate data set; the embodiment of the present invention adopts a coordinate data set from January 1, 2008 to December 30, 2019, and the coordinate data set is based on the Argo global observation plan and is formed by receiving and processing the French Argos satellite; in the coordinate data set, the coordinate data is collected every 6 hours, and a total of 32984 sets of data;

[0044] Secondly, a series of preprocessing tasks such as missing filling, anomaly removal, standardization and feature selection are performed on the generated coordinate data set to obtain Argo float trajectory feature data; in the embodiment of the present invention, after quality control of the 32984 sets of data generated, 24200 sets of valid data are screened out; 60% of them are used for training, 20% for verification, and 20% for testing.

[0045] Finally, the Argo buoy trajectory feature data of the first n-1 days within the n days is used as input, and the Argo buoy trajectory feature data of the nth day is used as output to train and generate a buoy trajectory prediction model; in an embodiment of the present invention, the buoy trajectory prediction model includes an SRU deep learning prediction model and a decision tree model; wherein, the SRU deep learning prediction model and the decision tree model are effectively combined by weighted average method to generate the final buoy trajectory prediction model; in this process, the buoy information is predicted by combining the deep learning network structure of the SRU model, which can effectively avoid the problems of the traditional RNN model prediction taking too long and the model being difficult to train; at the same time, considering that the generalization ability of deep learning is insufficient and lacks interpretability, a decision tree ensemble model is constructed based on the ensemble prediction theory, which has better prediction ability and generalization ability. Based on the buoy trajectory prediction model, the prediction results of the time series data can be obtained separately, and the various prediction results are synthesized to obtain the final prediction result y(t); the buoy trajectory prediction model uses the hyperbolic tangent function as the activation function, and has two hidden layers and one output layer, with 2000 neurons in each layer. The ratio of input and output data in training is 100:10, that is, training is performed by using 10 days of data to predict 1 day of data.

[0046] Next, the SRU deep learning prediction model and decision tree model are explained respectively.

[0047] For the SRU deep learning prediction model:

[0048] The SRU deep learning prediction model is constructed based on the time series coordinate data set. As a variant model of the Recurrent Neural Network (RNN), SRU makes up for the problems of gradient vanishing and gradient explosion, insufficient long-term memory capacity, etc. of RNN, so that the recurrent neural network can truly and effectively utilize long-distance time series information. The SRU deep neural network uses a special SRU structure to replace the hidden layer neurons of the general recursive neural network. The SRU structure contains input gates, output gates, forget gates and memory cells (Cell); the input gate, output gate, and forget gate are all logic units. They will not send their outputs to other neurons, but are responsible for setting weights at the edges where other parts of the neural network are connected to the memory cells, which are used to selectively memorize the correction parameters of the feedback error function as the gradient descent. Its specific structure is as follows Figure 3 shown.

[0049] The input gate and output gate in the SRU recurrent neural network are used to receive and output parameters and modify parameters, denoted by i and h respectively. The forget gate indicates whether to retain the historical information stored in the current hidden layer node, denoted by f. The memory cell represents the memory of the neuron state, denoted by c. The design of the gated cell unit enables the SRU unit to save, read, reset and update long-distance historical information. The training loss function diagram of the SRU deep learning prediction model is shown in Figure 1. Figure 5 As shown; the calculation process of the SRU network structure is as follows:

[0050]

[0051] f t =σ(W f x t +b f )

[0052] r t =σ(W r x t +b r )

[0053]

[0054] h t =r t ⊙g(c t )+(1-r t )⊙x t

[0055] in, Represents the current input x tThe intermediate state after transformation with the neural network weight matrix W; the previous neuron state is C t-1 ; ⊙ represents matrix product operation; f t represents the forget gate at time t; r t represents the reset gate at time t (used to determine whether to combine the current state with the previous information); h t represents the output gate at time t; c t represents the internal state at time t; x t Represents the input at time t; W and b are the corresponding weight coefficient matrix and bias term respectively; σ represents the sigmoid activation function.

[0056] The SRU deep learning prediction model used in the present invention is a multi-layer SRU parallel optimization deep learning prediction model based on convolution, which specifically implements the construction of multi-layer SRU; since the activation functions of matrix addition and sigmoid need to call their own independent functions respectively, and add additional operation delays and data movement overhead. Therefore, the present invention optimizes, transforms and upgrades the multi-layer SRU model based on gridding, parallel computing and other technologies, so that its computing efficiency and GPU parallel computing performance are significantly improved, and the specific steps are as follows:

[0057] 1. For neural networks with unequal dimensions, the three weight matrices of the neural network are merged into one large matrix through linear transformation.

[0058] 2. By putting the element-by-element multiplication operation into a kernel function, the matrix multiplication of all time steps can be processed in parallel, which significantly improves the computing efficiency and GPU utilization.

[0059] 3. Through gridding, parallel computing and other operations, the two outer fors in the kernel function can realize parallel operations. The innermost for is based on the sequence order, and only this dimension requires front-and-back association. The minibatch dimension and the hiddenState dimension can be separately calculated in parallel. Finally, the sequence dimension that needs to be managed front and back can be put into the register to maintain the front-and-back relationship. The minibatch and hiddenState dimensions can be regarded as the x and y axes of the grid.

[0060] For the decision tree model:

[0061] The input of the decision tree algorithm is the training set D, the threshold of the Gini coefficient, and the threshold of the number of samples. The output is a decision tree T. The algorithm starts from the root node and recursively builds the tree using the training set.

[0062] 1) For the data set D of the current node, if the number of samples is less than the threshold or there is no feature, the decision subtree is returned and the recursion of the current node stops.

[0063] 2) Calculate the Gini coefficient of sample set D. If the Gini coefficient is less than the threshold, return the decision tree subtree and stop recursion at the current node.

[0064] 3) Calculate the Gini coefficient of each feature value of each feature of the current node for the data set D.

[0065] 4) Select the feature A and the corresponding feature value a with the smallest Gini coefficient from the calculated Gini coefficients of each feature value for the data set D. Based on the optimal feature and optimal feature value, divide the data set into two parts D1 and D2, and establish the left and right nodes of the current node. The data set D of the node is D1, and the data set D of the right node is D2.

[0066] 5) Recursively call steps 1-4 on the left and right child nodes to generate a decision tree.

[0067] After the decision tree is established, the prediction method is to use the mean or median of the final leaf to predict the output result.

[0068] For the regression model, the embodiment of the present invention uses a common sum variance measurement method. The measurement goal of the CART regression tree is to find the feature and feature value partition point corresponding to the minimum mean square error of each set of D1 and D2 and the minimum sum of the mean square errors of D1 and D2 for any partition feature A and the corresponding data sets D1 and D2 on both sides of any partition point s. The expression is:

[0069]

[0070] Where A represents the partition feature (such as time information); s is the partition point used to divide the data sets D1 and D2; x i represents the x-axis variable in data set D1 or D2; y i and t They represent the y-axis variables in data sets D1 and D2 respectively; c1 and c2 represent the means of data sets D1 and D2 respectively.

[0071] In an embodiment of the present invention, the buoy trajectory prediction model is specifically constructed using a three-layer SRU and decision tree algorithm, with 256 neurons in each layer, a dropout of 0.2, a learning rate of 0.001-0.1, and multi-year climatological ocean currents as the background field. The background field at the next moment is used as an input parameter to perform a more reliable trajectory prediction at the next moment, and the comprehensive buoy trajectory data and ocean current seasonal characteristics are used as model hyperparameter inputs.

[0072] The decision tree algorithm and the SRU model are organically combined in a weighted average manner to ultimately achieve effective prediction of the buoy trajectory. Figure 4 As shown; the comparison diagram of the prediction results of the Argo float trajectory time series of the embodiment of the present invention and the actual value is shown in FIG. Figure 6 shown.

[0073] In the embodiment of the present invention, the root mean square error (RMSE) and the mean absolute error (MAE) are used to evaluate the prediction results, and the calculation formulas are as follows.

[0074]

[0075]

[0076] Among them, l represents the number of sequence data; i represents the current sequence point; y i obs and i pre represent the observed true value and the predicted value respectively.

[0077] RMSE is a measure of the deviation between the observed value and the true value; MAE is the average of the absolute errors, which can better reflect the actual situation of the predicted value error. The smaller the RMSE and MAE, the better the prediction effect.

[0078] The following table shows multiple buoy trajectory prediction evaluation indicators.

[0079] Prediction results RMSE(°) MAE(°) Buoy 1 0.42 0.36 Buoy 2 0.45 0.35 Buoy 3 0.46 0.38 Buoy 4 0.48 0.4 Buoy 5 0.47 0.39 Buoy 6 0.53 0.43 Buoy 7 0.57 0.46 Buoy 8 0.65 0.48 Buoy 9 0.57 0.45 Buoy 10 0.59 0.46 average 0.51 0.45

[0080] It can be seen that the average absolute error of the prediction model is 0.45°, and the overall state of the model is stable without drastic changes. Moreover, when using the method provided by the embodiment of the present invention to predict the trajectory of the Argo float, it is only necessary to load the data of the past 10 days into the trained model, and the prediction result can be quickly obtained, avoiding the long-term calculation of traditional numerical prediction.

[0081] The embodiment of the present invention provides an Argo float trajectory prediction method based on convolution and decision tree. By acquiring Argo float trajectory data and using the data information obtained by time series mining, a float trajectory prediction model based on CNN convolutional neural network, SRU deep learning and decision tree is proposed, and finally the predicted value of Argo float trajectory is obtained, which improves the prediction accuracy and efficiency of Argo float trajectory. The method provided by the embodiment of the present invention realizes high-precision prediction of Argo float trajectory data, with a prediction data time interval of 6 hours and a prediction average absolute error of 0.45°, reaching the international advanced level. The method provided by the embodiment of the present invention only needs to use the long time series data of Argo float trajectory, and does not need to use other data. The method provided by the embodiment of the present invention occupies less resources and has a fast calculation speed. The method provided by the embodiment of the present invention can be used for but not limited to Argo float trajectory prediction, and can also be used for target tracking prediction other than Argo float trajectory.

[0082] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for predicting Argo float trajectory based on convolution and decision tree, characterized in that: include: S1. Acquire the coordinate data of the Argo float every m hours within a preset time period, and form a coordinate data set with the acquired coordinate data; S2, preprocessing the coordinate data set to obtain Argo buoy trajectory feature data; S3, inputting the Argo buoy trajectory feature data into a buoy trajectory prediction model, and outputting the coordinate position data of the Argo buoy one day after the preset time period; The buoy trajectory prediction model includes an SRU deep learning prediction model and a decision tree model; the SRU deep learning prediction model and the decision tree model are effectively combined by a weighted average method; The SRU deep learning prediction model is a multi-layer SRU parallel optimization deep learning prediction model based on convolution, and the corresponding optimization steps include: Performing linear transformation on multiple neural networks of unequal dimensions, and merging the multiple neural networks of unequal dimensions into one matrix; All element-by-element multiplications are performed through the kernel function; The loop statements in the kernel function are processed based on the sequence order; and the loop statements outside the kernel function are processed in parallel through gridding and parallel computing methods.

2. The Argo float trajectory prediction method based on convolution and decision tree according to claim 1, characterized in that: The S2 specifically includes: S21, removing abnormal data in the coordinate data set; S22, performing standardization processing on the coordinate data set after removing abnormal data; S23. Use the CNN convolutional neural network algorithm to select the Argo buoy trajectory feature data from the standardized coordinate data set.

3. The Argo float trajectory prediction method based on convolution and decision tree as claimed in claim 2, characterized in that: In S23, the Argo float trajectory characteristic data includes: The distance values ​​between different coordinate data, and the direction change data of each coordinate data after being affected by ocean currents and air pressure.

4. The Argo buoy trajectory prediction method based on convolution and decision tree according to claim 1, characterized in that: The buoy trajectory prediction model is trained and generated by the following steps, including: Within n days, the coordinate data of the Argo float are obtained every m hours, and the obtained coordinate data are combined into a coordinate data set; Preprocessing the coordinate data set to obtain Argo buoy trajectory characteristic data; The Argo float trajectory feature data of the first n-1 days within the n days is used as input, and the Argo float trajectory feature data of the nth day is used as output to train and generate a float trajectory prediction model.

5. The Argo float trajectory prediction method based on convolution and decision tree according to claim 4, characterized in that: The buoy trajectory prediction model uses a hyperbolic tangent function as an activation function and has two hidden layers and one output layer.

Images