An arctic buoy track prediction method based on deep learning

Abstract

The application belongs to the technical field of track prediction, and relates to an Arctic buoy track prediction method based on deep learning. The Arctic sea ice drift data is first collected and then screened. The Arctic sea ice drift data is then subjected to missing value filling, anomaly removal, standardization and feature selection preprocessing to generate a standardized multi-source feature set. Then, a track prediction analysis model is constructed and subjected to forward propagation to output a prediction result. The BuoyModel neural network model is used to fit the buoy track law, effective information is obtained from the historical track sequence, and the buoy position after multiple time steps in the future is predicted, thereby avoiding the error accumulation problem caused by the traditional LSTM model single-step prediction, directly predicting the multi-time step track, and the model structure is simple, the calculation efficiency is high, the model is suitable for regions with large changes in space and time, and the prediction accuracy is improved through data enhancement and feature selection.

Description

Technical Field

[0001] This invention belongs to the field of trajectory prediction technology and relates to a deep learning-based method for predicting the trajectory of Arctic buoys. Background Technology

[0002] Ocean buoys are complex systems involving multiple disciplines, and the level of ocean buoy monitoring technology is a comprehensive reflection of a country's marine equipment manufacturing capabilities. Currently, general-purpose ocean data buoy observation technology has matured, resulting in a diverse product series and the formulation of corresponding marine environmental monitoring specifications and standards. Drifting ocean buoys provide valuable data for ocean and atmospheric research, playing a crucial role in short-term weather forecasting and long-term climate prediction. To ensure the normal operation of the observation array, it is necessary to constantly monitor the buoys' operational status to maintain a certain number and density of buoys within the study area. However, buoy deployment is expensive; deploying too early leads to resource waste, while deploying too late results in missing data. One significant issue in the maintenance and management of existing GNSS buoys is the need to monitor their drift trajectories in real time. Failures in the GNSS buoy's positioning, power, and communication systems can lead to a loss of accurate buoy location, posing serious challenges to subsequent repair and recovery. Furthermore, buoys are crucial navigational aids, and in complex waterways, drifting and loss are frequent occurrences. Therefore, to ensure the safety of buoys, it is essential to rationally plan buoy deployment and recovery, and accurately predict their drift trajectories.

[0003] Existing buoy trajectory prediction methods, such as CN112364119A which discloses an LSTM encoding / decoding model-based method for ocean buoy trajectory prediction, include: acquiring the trajectory sequence of ocean buoys, including the longitude and latitude coordinates of the ocean buoys, with all ocean buoy trajectory sequences constituting a dataset, and dividing the dataset into a training set and a test set; the encoder periodically encodes and normalizes the data of one trajectory point in the training set at each time step, and combines it with the hidden layer output of the previous time step to generate the output of that step, continuously reading along the time axis until finally converting it into a fixed-length global vector representation; the decoder initializes its internal state based on the global vector representation until the model converges; the test set is input into the trained model to obtain the predicted trajectory of the ocean buoy; however, the LSTM encoding / decoding model described in this method is a cyclic single-step prediction, which is prone to error accumulation during the prediction process; CN117972315A discloses a... A method for predicting the trajectory of a drifting buoy based on a hybrid neural network prediction model includes: S1, acquiring simultaneous and spatiotemporal marine environmental data and historical trajectory data of the drifting buoy; S2, preprocessing the data acquired in S1 to calculate the input data required for predicting the northward and eastward velocities of the drifting buoy; S3, injecting the input data required for predicting the northward and eastward velocities of the drifting buoy into the hybrid neural network prediction model to obtain the predicted values ​​of the northward and eastward velocities of the drifting buoy; S4, calculating the latitude and longitude coordinates of the trajectory points of the drifting buoy based on the predicted values ​​of the northward and eastward velocities of the drifting buoy; S5, using steps S1-S4, predicting the latitude and longitude coordinates of the trajectory points of the drifting buoy at multiple future time points in chronological order to obtain the trajectory point coordinate sequence, thus realizing trajectory prediction. The hybrid neural network prediction model used in this method has a complex structure and a large computational load, and is only applicable to simultaneous and spatiotemporal ranges, and is not applicable to areas with large spatiotemporal variations. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and to design a deep learning-based method for predicting the trajectory of Arctic buoys. Based on deep learning, a neural network model is used to establish a buoy drift tracking and prediction analysis model to predict the buoy drift trajectory and the optimal search area.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A deep learning-based method for predicting Arctic buoy tracks includes the following steps:

[0007] S1. Data Collection: Collect Arctic sea ice drift data and filter the Arctic sea ice drift data, selecting the time, location, temperature and air pressure of sea ice drift as known data and training and testing data for neural network models.

[0008] S2. Data preprocessing: The Arctic sea ice drift data is preprocessed by missing data filling, outlier removal, standardization and feature selection to generate a standardized multi-source feature set;

[0009] S3. Buoy Track Prediction and Analysis Model Construction: A buoy track prediction and analysis model is constructed based on the BuoyModel neural network model. The buoy track prediction and analysis model includes an input layer, two hidden layers, and an output layer. The input layer takes in the feature vectors from the multi-source feature set generated in step S2. The feature vectors include historical trajectories of latitude and longitude, historical drift speeds (magnitude and direction), temperature, air pressure, wind speed, sea ice density, and periodic encoding of time variables (such as daily periodic sine and cosine encoding). These variables are windowed and concatenated to form a fixed-length one-dimensional feature vector, which serves as the input data for the input layer. The first hidden layer consists of a fully connected layer (fc1), which maps the input feature vectors to hidden representations and introduces a nonlinear transformation using the ReLU activation function. The second hidden layer further processes the hidden representations through another fully connected layer (fc2), also using the ReLU activation function to enhance the model's expressive power. The output layer (fc3) converts the hidden representations into the target output.

[0010] S4. Forward Propagation Outputs Prediction Results: Forward propagation includes three fully connected layers (first fully connected layer fc1 → first hidden layer, second fully connected layer fc2 → second hidden layer, third fully connected layer fc3 → output layer. These layers sequentially perform layer-by-layer nonlinear mapping on the input feature vector, ultimately outputting the predicted result y (i.e., the longitude and latitude of the buoy at the next moment) and the ReLU activation function. For the input vector x, the model output is: ,in , , , and These are the weight matrices for the first, second, and third fully connected layers, respectively. , and These are the bias vectors for the first, second, and third layers, respectively. ReLU(x) = max(0,x) is the ReLU activation function; where, during the forward propagation, the input vector x represents the set of features extracted and standardized from the original buoy observation data, including variables such as latitude and longitude, temperature, air pressure, speed, and direction. After standardization, x∈R7 is fed into the neural network model BuoyModel, which passes through three fully connected layers and the ReLU activation function in sequence, finally outputting the predicted value y^, i.e., the latitude and longitude coordinates of the buoy at the future time. The weights Wi and bias bi are parameters automatically learned during model training, aiming to approximate the nonlinear mapping relationship between input and output. The training process uses the MSE loss function and the Adam optimizer to achieve error backpropagation and parameter updates, thereby continuously improving the prediction accuracy of the model.

[0011] In this invention, the mean squared error (MSE) is used to evaluate the quality of the prediction results during the training process. The MSE is calculated using the MSE as the loss function. Where N is the number of samples, This is the predicted value of the buoy track. This represents the actual value of the buoy's trajectory.

[0012] When updating parameters, this invention first uses the chain rule to calculate the gradient of the loss function with respect to each parameter, and then uses the Adam optimizer to update the parameters. The update formula of the Adam optimizer is as follows:

[0013] ,

[0014] ,

[0015] ,

[0016] ,

[0017] ,

[0018] in and It is an estimate of momentum and acceleration from Arctic ocean ice drift data. and It is the rate of decay of the momentum term. This is the current gradient. It is the learning rate, used to prevent division by zero for decimals.

[0019] This invention analyzes and optimizes the model input structure by calculating feature importance. First, it calculates the MSE (Mean Sequence Equation) on the original dataset. Then, it performs permutation on each feature obtained in step S2 (the variables obtained after data preprocessing in S2 and used as model input; these features are standardized (mean 0, standard deviation 1) and concatenated into an input vector x, which becomes the model input). (Feature permutation is a method to evaluate the impact of a feature on model performance; it artificially disrupts the order of a feature, breaking its correspondence with the label. If model performance significantly decreases, it indicates that the feature is important for the prediction task.) The MSE after permutation is then calculated. The importance of feature i is defined as: ,in It is the MSE on the original dataset.

[0020] This invention improves the accuracy of buoy trajectory prediction by quantitatively evaluating the prediction error and continuously optimizing the model parameters based on error backpropagation and gradient optimization mechanisms. At the same time, it introduces feature permutation analysis to quantify the impact of each input variable on the prediction results, assisting in the design and optimization of the model feature structure, and further enhancing the prediction stability and interpretability.

[0021] Compared with the prior art, the present invention has the following advantages: The present invention uses the BuoyModel neural network model to fit the trajectory pattern of the buoy, obtains effective information from the historical trajectory sequence and predicts the position of the buoy multiple time steps in the future, avoids the error accumulation problem caused by the traditional LSTM model's cyclic single-step prediction, directly predicts the trajectory of multiple time steps, and has a simple model structure, high computational efficiency, adapts to areas with large spatiotemporal changes, and improves prediction accuracy through data augmentation and feature selection. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the workflow of the present invention.

[0023] Figure 2 The above are the buoy tracking prediction results described in Embodiment 2 of the present invention, where (a)-(b) are the prediction results for buoys 2014C1, 2014C2, 2014C5, and 2014C7, respectively.

[0024] Figure 3 The image shows the trace prediction error diagram (a) and the trace prediction error histogram (b) of the buoy 2014C1 described in Embodiment 2 of the present invention.

[0025] Figure 4 The image shows the trace prediction error diagram (a) and the trace prediction error histogram (b) of the buoy 2014C2 described in Embodiment 2 of the present invention.

[0026] Figure 5The image shows the trace prediction error diagram (a) and the trace prediction error histogram (b) of the buoy 2014C5 described in Embodiment 2 of the present invention.

[0027] Figure 6 The image shows the trace prediction error diagram (a) and the trace prediction error histogram (b) of the buoy 2014C7 described in Embodiment 2 of the present invention. Detailed Implementation

[0028] The present invention will be further described below with reference to the embodiments and accompanying drawings.

[0029] Example 1:

[0030] like Figure 1 As shown, this embodiment provides a deep learning-based method for predicting Arctic buoy tracks, specifically including the following steps:

[0031] S1. Data Collection: Collect Arctic sea ice drift data and filter the Arctic sea ice drift data, selecting the time, location, temperature and air pressure of sea ice drift as known data and training and testing data for neural network models.

[0032] S2. Data preprocessing: Perform missing data filling, outlier removal, standardization, and feature selection preprocessing on Arctic sea ice drift data;

[0033] S3. Buoy Track Prediction and Analysis Model Construction: A buoy track prediction and analysis model is constructed based on the BuoyModel neural network model. The buoy track prediction and analysis model includes an input layer, two hidden layers, and an output layer. The input layer takes in the feature vector. The first hidden layer consists of a fully connected layer (fc1), which maps the input feature vector to the hidden representation and introduces a nonlinear transformation using the ReLU activation function. The second hidden layer further processes the hidden representation through another fully connected layer (fc2), and also uses the ReLU activation function to enhance the model's expressive power. The output layer (fc3) converts the hidden representation into the target output.

[0034] S4. Forward Propagation Output Prediction: Forward propagation includes three fully connected layers and a ReLU activation function. For the input vector x, the model output is: ,in , , , and These are the weight matrices for the first, second, and third fully connected layers, respectively. , and These are the bias vectors for the first, second, and third layers, respectively. ReLU(x) = max(0,x) is the ReLU activation function; where, during the forward propagation, the input vector x represents the set of features extracted and standardized from the original buoy observation data, including variables such as latitude and longitude, temperature, air pressure, speed, and direction. After standardization, x∈R7 is fed into the neural network model BuoyModel, which passes through three fully connected layers and the ReLU activation function in sequence, finally outputting the predicted value y^, i.e., the latitude and longitude coordinates of the buoy at the future time. The weights Wi and bias bi are parameters automatically learned during model training, aiming to approximate the nonlinear mapping relationship between input and output. The training process uses the MSE loss function and the Adam optimizer to achieve error backpropagation and parameter updates, thereby continuously improving the prediction accuracy of the model.

[0035] In this embodiment, the mean squared error (MSE) is used as the loss function during training, and the mean squared error is calculated as follows: Where N is the number of samples, This is the predicted value of the buoy track. This represents the actual value of the buoy's trajectory.

[0036] In this embodiment, backpropagation and gradient descent are used to update the parameters. Specifically, the chain rule is first used to calculate the gradient of the loss function with respect to each parameter, and then the Adam optimizer is used to update the parameters. The update formula of the Adam optimizer is shown below:

[0037] ,

[0038] ,

[0039] ,

[0040] ,

[0041] ,

[0042] in and It is an estimate of momentum and acceleration from Arctic ocean ice drift data. and It is the rate of decay of the momentum term. This is the current gradient. It is the learning rate, used to prevent division by zero for decimals;

[0043] This embodiment calculates the MSE (Mean Sequence Size) of the original dataset by performing feature importance analysis and optimizing the model input structure. Then, it permutes each feature and calculates the MSE after the permutation. The importance of feature i is defined as:

[0044] ,in It is the MSE on the original dataset.

[0045] Example 2:

[0046] This embodiment uses the method described in the previous embodiment to collect Arctic sea ice drift data from March 18 to May 18, 2014 (https: / / doi.pangaea.de / 10.1594 / PANGAEA.941334). The data contains drift data for each Arctic sea ice floe, including the collection date, time, and geographic coordinate system (GCS). The latitude and longitude coordinates in WGS-84 are used in this embodiment. The data time is filtered to select sufficiently long and similar time periods. The drift time, drift location, temperature, and air pressure of sea ice are used as known data, and as training and testing data for the neural network model. Specifically, time is used as the standard. The neural network model is trained using eight Arctic sea ice drift data points: 2014C1, 2014C2, 2014C4, 2014C5, 2014C7, 2014C11, 2014C12, and 2014C13. The experimental results are then used for verification, with the specific prediction results as follows: Figure 2 As shown in Table 1.

[0047] Table 1:

[0048] Data Volume 7 - Buoy 1 173 1220 1232 Data Volume 6 - Buoy 1 853 3034 3152 Data Volume 5 - Buoy 1 1157 7099 7193 Data Volume 4 - Buoy 1 934 9535 9581 Data Volume 6 - Buoy 2 6054 25313 26027 Data Volume 5 - Buoy 2 12518 19445 23126 Data Volume 4 - Buoy 2 13335 20547 24495 Data Volume 5 - Buoy 5 284 5151 5159 Data Volume 4 - Buoy 5 927 7065 7125 Data Volume 4 - Buoy 7 2076 4468 4927

[0049] In the table, σx (m) is the standard deviation of the model's prediction error in the longitude direction (unit: meters), σy (m) is the standard deviation of the model's prediction error in the latitude direction (unit: meters), and σp (m) is the standard deviation of the model's total spatial position prediction error, which is usually the square root of the sum of the squares of σx and σy.

[0050] This embodiment performs error analysis on two sets of data: the estimated coordinates and the original coordinates of Arctic sea ice. Using the measured data as the true value, the results of spatial and temporal estimations are subtracted from the measured results to obtain the deviation between the fitted coordinate data and the measured data. The results are plotted, and the error data is calculated. Based on this, the accuracy of the model predictions obtained with different amounts of training data is analyzed. The experimental error results are as follows: Figure 3-6 As shown in the figure, the overall prediction effect is relatively good.

Claims

1. A deep learning-based method for predicting Arctic buoy tracks, characterized in that, Specifically, the following steps are included: S1. Data Collection: Collect Arctic sea ice drift data and filter the data, selecting the drift time, drift location, temperature and air pressure as known data and training and testing data for the neural network model. S2. Data preprocessing: The Arctic sea ice drift data is preprocessed by missing data filling, outlier removal, standardization and feature selection to generate a standardized multi-source feature set; S3. Buoy Track Prediction and Analysis Model Construction: A buoy track prediction and analysis model is constructed based on the BuoyModel neural network model. The buoy track prediction and analysis model includes one input layer, two hidden layers, and one output layer. The input layer takes in the feature vectors from the multi-source feature set generated in step S2. The feature vectors include features such as historical trajectory of latitude and longitude, historical drift speed, temperature, air pressure, wind speed, sea ice density, etc., as well as the periodic encoding of time variables. These variables are windowed and concatenated to form a fixed-length one-dimensional feature vector, which serves as the input data for the input layer. The first hidden layer consists of a fully connected layer that maps the input feature vector to the hidden representation and introduces a nonlinear transformation using the ReLU activation function. The second hidden layer further processes the hidden representation through another fully connected layer, also using the ReLU activation function to enhance the model's expressive power. The output layer converts the hidden representation into the target output. S4. Forward Propagation: Forward propagation consists of three fully connected layers and a ReLU activation function. For an input vector x, the model output is: ,in , , , and These are the weight matrices for the first, second, and third fully connected layers, respectively. , and These are the bias vectors for the first, second, and third layers, respectively. ReLU(x) = max(0,x) is the ReLU activation function; where, during the forward propagation, the input vector x represents the set of features extracted and standardized from the original buoy observation data, including latitude and longitude, temperature, air pressure, speed, and direction. After standardization, x∈R7 is fed into the neural network model BuoyModel, which passes through three fully connected layers and the ReLU activation function in sequence, and finally outputs the predicted value y^, which is the latitude and longitude coordinates of the buoy at the future time. The weights Wi and biases bi are parameters automatically learned during the model training process. The training process uses the MSE loss function and the Adam optimizer to achieve error backpropagation and parameter updates, thereby continuously improving the model's prediction accuracy.

2. The method for predicting Arctic buoy tracks based on deep learning according to claim 1, characterized in that, During training, the mean squared error (MSE) is used to evaluate the quality of the prediction results. The MSE is calculated using the mean squared error as the loss function: Where N is the number of samples, This is the predicted value of the buoy track. This represents the actual value of the buoy's tracking trajectory.

3. The method for predicting Arctic buoy tracks based on deep learning according to claim 2, characterized in that, When updating parameters, the gradient of the loss function with respect to each parameter is first calculated using the chain rule, and then the parameters are updated using the Adam optimizer. The update formula for the Adam optimizer is shown below: ， ， ， ， ， in and It is an estimate of momentum and acceleration from Arctic ocean ice drift data. and It is the rate of decay of the momentum term. This is the current gradient. It is the learning rate, used to prevent division by zero for decimals.

4. The method for predicting Arctic buoy tracks based on deep learning according to claim 3, characterized in that, By calculating feature importance analysis and optimizing the model input structure, the MSE on the original dataset is first calculated. Then, each feature obtained in step S2 is permuted, and the MSE of the permuted feature is calculated. The importance of feature i is defined as: ,in It is the MSE on the original dataset.

Images