Marine environment time sequence prediction method and system based on recurrent neural network

Abstract

The invention provides a marine environment time sequence prediction method and system based on a recurrent neural network, and relates to the technical field of time sequence prediction, and the method comprises the steps: obtaining marine environment historical observation data, carrying out the preprocessing of the marine environment historical observation data, and removing abnormal values and missing values from the marine environment historical observation data; according to the preprocessed marine environment historical observation data, a recurrent neural network model is constructed, and the recurrent neural network model comprises an input layer, two stacked long short-term memory (LSTM) network layers and an output layer; acquiring marine environment observation data at the current moment, and inputting the marine environment observation data into the trained recurrent neural network model; predicting marine environment data of a plurality of time steps in the future through a recurrent neural network model, and obtaining a marine environment prediction result in a preset time; and according to the marine environment prediction result, identifying a hot spot region of marine environment change, and generating a visual result representing a dynamic change rule of the marine environment in the prediction time period.

Description

Technical Field

[0001] The present invention relates to time series prediction technology, and in particular to a method and system for predicting ocean environment time series based on recursive neural network. Background Art

[0002] Marine environment time series prediction has important application value in marine disaster warning, marine resource management, marine engineering planning and other fields. Traditional marine environment time series prediction methods mainly include statistical methods and numerical model methods. Statistical methods such as autoregressive moving average model (ARMA) and seasonal autoregressive difference moving average model (SARIMA) make predictions by establishing statistical relationships between variables, but it is difficult to characterize the nonlinear dynamic characteristics of the marine environment. Numerical model methods such as ocean circulation models and wave models simulate the evolution of the marine environment by solving fluid dynamics equations, but they are limited by the uncertainty of model parameters and the limitation of computing resources, making it difficult to achieve high-precision long-term predictions. Summary of the invention

[0003] The embodiments of the present invention provide a method and system for predicting ocean environment time series based on a recursive neural network, which can at least solve some problems in the prior art.

[0004] According to a first aspect of the embodiments of the present invention,

[0005] A method for predicting ocean environment time series based on recursive neural network is provided, comprising:

[0006] Acquire historical observation data of the marine environment, pre-process the historical observation data of the marine environment, and remove abnormal values ​​and missing values ​​therein, wherein the historical observation data of the marine environment includes at least two of sea surface temperature, sea surface salinity, sea surface density, sea surface height, ocean current speed, ocean current direction, and wind speed; wherein the historical observation data of the marine environment is time series data, including observation data of multiple time steps;

[0007] According to the preprocessed historical observation data of the marine environment, a recursive neural network model is constructed, and the recursive neural network model includes an input layer, two stacked long short-term memory network LSTM layers and an output layer; wherein the input layer is used to receive the observation data at the current moment, and sequentially transmit it to the two LSTM layers, each LSTM layer includes a plurality of LSTM units, and each LSTM unit is implemented based on a gating mechanism, and the output results of the two LSTM layers are spliced ​​and transmitted to the output layer, and the output layer adopts a fully connected layer structure, realizes nonlinear transformation through an activation function, and outputs the predicted data of the next moment;

[0008] The ocean environment observation data at the current moment is obtained, and after preprocessing operations on the ocean environment observation data at the current moment, it is input into the trained recursive neural network model; through the recursive neural network model, the ocean environment data of multiple time steps in the future is predicted to obtain the ocean environment prediction results within a preset time; according to the ocean environment prediction results, the hot spots of ocean environment changes are identified, and the visualization results representing the dynamic change law of the ocean environment within the prediction time period are generated.

[0009] In an optional embodiment,

[0010] Preprocessing the marine environment historical observation data includes:

[0011] Performing data cleaning on each observation indicator in the historical observation data of the marine environment, identifying outliers that deviate from the mean within a range of plus or minus three times the standard deviation, and replacing the outliers with the average values ​​before and after the outliers, to obtain observation indicator data with outliers removed;

[0012] Interpolating the missing values ​​in the observed indicators after removing the outliers, estimating the missing values ​​by calculating the valid observed values ​​before and after the missing values ​​using the cubic spline interpolation method, and obtaining the observed indicator data without missing values;

[0013] Normalizing the observed indicators without missing values, calculating the 0.05 quantile and 0.95 quantile of each observed indicator as the minimum and maximum values ​​of the observed indicator, mapping each observed indicator to the interval of 0 to 1, and obtaining normalized observed data;

[0014] Based on the time scale analysis and cross-validation method, the optimal time window size and prediction time span are determined, and the normalized observation data are reconstructed with the optimal time window size as the window by using the sliding time window method. Each training sample is composed of observation data at multiple consecutive moments as input, and observation data with the optimal prediction time span as the span is output to obtain a reconstructed observation data set, which is the preprocessed historical observation data of the marine environment.

[0015] In an optional embodiment,

[0016] The optimal time window size and prediction time span are determined based on time scale analysis and cross-validation methods, including:

[0017] Perform time scale characteristic analysis on the time series of each marine environmental element, identify the periodic components of the time series using autocorrelation analysis, determine the optimal lag order of the time series using partial autocorrelation analysis, obtain the energy distribution of the time series on different frequency components using wavelet analysis and singular spectrum analysis, and determine the candidate value of the time window size of each marine environmental element based on the comprehensive analysis results; determine the candidate value of the prediction time span according to the actual prediction needs;

[0018] A cross-validation method is used to evaluate the performance of the prediction model under different combinations of candidate values ​​of time window sizes and prediction time spans, and the cross-validation method is any one of K-fold cross-validation, leave-one-out cross-validation and time series cross-validation; the K-fold cross-validation randomly divides the observed data into K subsets, selects one subset as the validation set each time, and the remaining subsets are used as training sets, trains the prediction model and evaluates the model performance on the validation set, and after repeating K times, takes the average value of the performance evaluation index as the performance score under the current candidate value combination; the leave-one-out cross-validation uses only one sample as the validation set each time, and the remaining samples as the training set, repeats N times until each sample is used as a validation set once, and takes the average value of the performance evaluation index as the performance score under the current candidate value combination, where N is the total number of samples; the time series cross-validation adopts a forward rolling time series segmentation method, divides the observed data into a training set and a validation set in turn, rolls forward one time step each time, and repeats until the entire time range is covered, and takes the average value of the performance evaluation index as the performance score under the current candidate value combination;

[0019] A strategy combining grid search and random search is adopted to search the candidate value combination space, and the cross-validation is performed on each candidate value combination to obtain its performance score; the candidate value combination with the best performance score is selected to determine the optimal time window size and prediction time span.

[0020] In an optional embodiment,

[0021] Based on the preprocessed historical observation data of the marine environment, a recursive neural network model is constructed, including:

[0022] The input layer of the recursive neural network model receives the preprocessed historical observation data of the ocean environment at the current moment, and transmits the received historical observation data of the ocean environment to the two stacked LSTM layers;

[0023] Each of the two stacked LSTM layers includes a plurality of LSTM units, and the plurality of LSTM units realize adaptive learning and memory of long-term dependencies in the historical observation data of the marine environment through a gating mechanism. Each of the LSTM units includes an input gate, a forget gate, an output gate and a cell state. The input gate controls the proportion of the input information at the current moment updated to the cell state according to the input at the current moment and the hidden state at the previous moment. The forget gate controls the proportion of the cell state at the previous moment that needs to be forgotten according to the input at the current moment and the hidden state at the previous moment. The output gate controls the proportion of the cell state at the current moment output to the hidden state according to the input at the current moment and the hidden state at the previous moment. The cell state is used to store long-term memory information and control the accumulation and forgetting of memory information.

[0024] The output result of the last LSTM layer in the two stacked LSTM layers is passed to the output layer. The output layer maps the output of the last LSTM layer to an output space consistent with the prediction target dimension through a fully connected layer, and realizes nonlinear transformation through an activation function to generate the ocean environment prediction data for the next moment.

[0025] In an optional embodiment,

[0026] The multiple LSTM units realize adaptive learning and memory of long-term dependencies in the historical observation data of the marine environment through a gating mechanism, including:

[0027] The input gate receives the input at the current moment and the hidden state at the previous moment, combines the attention weight obtained by weighted summation of the historical hidden state sequence, and uses the sigmoid activation function to generate a gating signal that controls the proportion of the current input information updated to the cell state;

[0028] The forget gate receives the input at the current moment and the hidden state at the previous moment, introduces sine and cosine functions in the gate calculation to encode the time periodic information, and uses the sigmoid activation function to generate a gate signal that controls the proportion of the cell state at the previous moment that needs to be forgotten;

[0029] The output gate receives the input at the current moment and the hidden state at the previous moment, combines the candidate hidden state calculated by the current input and the hidden state at the previous moment, introduces the update gate mechanism in the gated recursive unit, and uses the sigmoid activation function to generate a gating signal that controls the ratio of the current cell state output to the hidden state;

[0030] Updating the cell state, by integrating the candidate cell state at the current moment, the gating signal of the forget gate and the gating signal of the input gate, the cell state at the previous moment is updated to achieve the selective retention and update of the long-term memory information; the candidate cell state is obtained by using the input at the current moment and the hidden state at the previous moment and applying the tanh activation function to calculate, as a potential cell state update;

[0031] Based on the structure of the LSTM unit, a multi-layer long short-term memory network model is constructed.

[0032] In an optional embodiment,

[0033] Based on the marine environment prediction results, identify the hot spots of marine environment changes and generate visualization results that show the dynamic changes of the marine environment during the prediction period, including:

[0034] The time series at each grid point in the marine environment prediction results is regarded as an independent signal, and the time series at each grid point is decomposed into time and frequency using wavelet transform, and the dominant cycle and key time points of the changes of marine environmental elements at each grid point are extracted to obtain the time-frequency spectrum coefficients at different scales.

[0035] The wavelet spectrum coefficients of all grid points form a high-dimensional spectrum matrix, and the dynamic time warping (DTW) distance is used to measure the similarity between grid points in the time dimension to obtain a spatiotemporal similarity matrix; the spatiotemporal similarity matrix is ​​spectrally decomposed, and the optimal number of clusters is adaptively determined by analyzing the distribution of eigenvalues, and the eigenvectors corresponding to the first k largest eigenvalues ​​are used as low-dimensional embedding coordinates of the grid points; the low-dimensional embedding coordinates of the grid points are input into the K-means clustering algorithm to obtain the clustering results of the hot spots of marine environmental changes, and each type of cluster corresponds to a hot spot area;

[0036] The clustering results are remapped back to the geographic space, and the average of the geographic coordinates of the internal grid points of each cluster is used as the representative point of the cluster in the geographic space. The standard deviation of the geographic coordinates of the internal grid points of the cluster is used as the radius to draw a spatial distribution map of the hotspot area. The color or filling mode of the hotspot area is set according to the average trend or extreme degree of the change of the marine environment in each hotspot area, and a visualization result representing the dynamic change law of the marine environment within the prediction time period is generated.

[0037] In an optional embodiment,

[0038] The wavelet spectrum coefficients of all grid points are used to form a high-dimensional spectrum matrix, and the dynamic time warping DTW distance is used to measure the similarity between grid points in the time dimension to obtain the spatiotemporal similarity matrix, including:

[0039] The high-dimensional spectrum matrix is ​​feature selected by using the mutual information criterion, and the subset of spectrum coefficients that contributes most to the similarity measurement is screened out to optimize the information representation of the spectrum matrix; for any two grid points, the optimized wavelet spectrum coefficient sequence is extracted, and the local distance matrix is ​​constructed by calculating the Mahalanobis distance between different time-frequency components, taking into account the correlation between the time-frequency components;

[0040] On the local distance matrix, a bandwidth-constrained dynamic programming algorithm is used to search for the optimal nonlinear alignment path between two wavelet spectrum coefficient sequences, and a radial basis kernel function is introduced to perform nonlinear transformation on the original cumulative distance to enhance the robustness of the dynamic time warping distance; wherein, according to the physical characteristics of the marine environmental elements, the bandwidth constraint parameters in the optimal alignment path search process are adaptively determined;

[0041] The dynamic time warping distance matrix between all grid point pairs at each time scale is calculated; a spatiotemporal data quality evaluation function is constructed, and the fusion weights of each time scale are determined according to the credibility of the distance matrix at different time scales to generate a spatiotemporal similarity matrix; the regularization term is used to constrain the fusion weights and the spatiotemporal similarity matrix is ​​locally adjusted according to prior knowledge.

[0042] In the second aspect of the embodiments of the present application,

[0043] Provides a marine environment time series prediction system based on recursive neural networks, including:

[0044] The first unit is used to obtain historical observation data of the marine environment, pre-process the historical observation data of the marine environment, and remove abnormal values ​​and missing values ​​therein, wherein the historical observation data of the marine environment includes at least two of sea surface temperature, sea surface salinity, sea surface density, sea surface height, ocean current speed, ocean current direction and wind speed; wherein the historical observation data of the marine environment is time series data, including observation data of multiple time steps;

[0045] The second unit is used to construct a recursive neural network model based on the preprocessed historical observation data of the marine environment, and the recursive neural network model includes an input layer, two stacked long short-term memory network LSTM layers and an output layer; wherein the input layer is used to receive the observation data at the current moment, and pass it to the two LSTM layers in sequence, each LSTM layer includes a plurality of LSTM units, and each LSTM unit is implemented based on a gating mechanism, and the output results of the two LSTM layers are spliced ​​and passed to the output layer, and the output layer adopts a fully connected layer structure, realizes nonlinear transformation through an activation function, and outputs the predicted data of the next moment;

[0046] The third unit is used to obtain the ocean environment observation data at the current moment, pre-process the ocean environment observation data at the current moment, and then input it into the trained recursive neural network model; through the recursive neural network model, predict the ocean environment data of multiple time steps in the future, and obtain the ocean environment prediction results within a preset time; based on the ocean environment prediction results, identify the hot spots of ocean environment changes, and generate visualization results that represent the dynamic change laws of the ocean environment within the prediction time period.

[0047] According to a third aspect of the embodiments of the present invention,

[0048] An electronic device is provided, comprising:

[0049] processor;

[0050] a memory for storing processor-executable instructions;

[0051] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0052] According to a fourth aspect of the embodiments of the present invention,

[0053] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0054] This application adopts a recursive neural network model, especially using multiple stacked long short-term memory network (LSTM) layers, which can effectively capture the long-term dependencies and nonlinear dynamic characteristics in the marine environment time series data, and improve the accuracy of marine environment prediction. The gating mechanism is introduced into the recursive neural network model, and the information flow is adaptively controlled through the input gate, forget gate and output gate, which enhances the model's ability to extract and memorize key spatiotemporal features and overcomes the problems of information forgetting and gradient disappearance in traditional recursive neural networks.

[0055] By using observation data of multiple time steps as input, the ocean environment data of multiple time steps in the future is predicted, which realizes the long-term ocean environment prediction and provides an important basis for the analysis of the medium- and long-term evolution of the ocean environment. By identifying the hot spots of ocean environment changes, the visualization results representing the dynamic changes of the ocean environment in the prediction time period are generated, which intuitively presents the spatiotemporal evolution characteristics of the ocean environment, and provides intuitive and easy-to-interpret decision support for ocean environment early warning, ocean resource management, etc.

[0056] The marine environment time series prediction method based on recursive neural network provided by the present invention makes comprehensive use of artificial intelligence, data mining, visualization and other technologies, realizes the intelligence and refinement of marine environment prediction, improves the efficiency and accuracy of marine environment prediction, and provides important technical support for marine environmental protection and development. The prediction method of the present invention is universal and scalable, and is not only suitable for the prediction of various marine environmental elements such as sea surface temperature and sea surface salinity, but also can flexibly adjust the model structure and parameters according to actual needs, and adapt to marine environment prediction tasks in different regions and different time and space scales. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 Schematic diagram of the flow of a method for predicting ocean environment time series based on a recursive neural network according to an embodiment of the present invention;

[0058] Figure 2 It is a schematic diagram of the structure of a marine environment time series prediction system based on a recursive neural network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0061] Figure 1 FIG. 1 is a flow chart of a method for predicting ocean environment time series based on a recursive neural network according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0062] S101. Acquire historical marine environment observation data, pre-process the historical marine environment observation data, and remove abnormal values ​​and missing values ​​therein, wherein the historical marine environment observation data includes at least two of sea surface temperature, sea surface salinity, sea surface density, sea surface height, ocean current speed, ocean current direction, and wind speed; wherein the historical marine environment observation data is time series data, including observation data of multiple time steps;

[0063] In an optional embodiment,

[0064] Preprocessing the marine environment historical observation data includes:

[0065] Performing data cleaning on each observation indicator in the historical observation data of the marine environment, identifying outliers that deviate from the mean within a range of plus or minus three times the standard deviation, and replacing the outliers with the average values ​​before and after the outliers, to obtain observation indicator data with outliers removed;

[0066] Interpolating the missing values ​​in the observed indicators after removing the outliers, estimating the missing values ​​by calculating the valid observed values ​​before and after the missing values ​​using the cubic spline interpolation method, and obtaining the observed indicator data without missing values;

[0067] Normalizing the observed indicators without missing values, calculating the 0.05 quantile and 0.95 quantile of each observed indicator as the minimum and maximum values ​​of the observed indicator, mapping each observed indicator to the interval of 0 to 1, and obtaining normalized observed data;

[0068] Based on the time scale analysis and cross-validation method, the optimal time window size and prediction time span are determined, and the normalized observation data are reconstructed with the optimal time window size as the window by using the sliding time window method. Each training sample is composed of observation data at multiple consecutive moments as input, and observation data with the optimal prediction time span as the span is output to obtain a reconstructed observation data set, which is the preprocessed historical observation data of the marine environment.

[0069] Exemplarily, for each observation indicator in the historical observation data of the marine environment, its mean and standard deviation are calculated; based on the mean and standard deviation, outliers that deviate from the mean within a range of plus or minus three times the standard deviation are identified; for each outlier, the observation values ​​before and after it are taken to calculate the average value; the outlier is replaced with the calculated average value to obtain the observation indicator data with the outlier removed.

[0070] For each observed indicator data after removing outliers, identify the missing values; for each missing value, take the valid observed values ​​before and after it; use the cubic spline interpolation method to estimate the missing value based on the valid observed values ​​before and after the missing value; fill the estimated value into the position of the missing value to obtain the observed indicator data without missing values.

[0071] For each observation indicator data without missing values, calculate its 0.05 quantile and 0.95 quantile; take the 0.05 quantile as the minimum value of the observation indicator and the 0.95 quantile as the maximum value; use the minimum-maximum normalization formula to map the data of each observation indicator to the interval of 0 to 1; and obtain the normalized observation data.

[0072] Set a set of candidate time window sizes and prediction time spans; use the cross-validation method to divide the normalized observation data into a training set and a validation set; for each candidate time window size and prediction time span, perform model training on the training set, and evaluate the model performance on the validation set; select the time window size and prediction time span with the best performance on the validation set as the optimal parameters.

[0073] Taking the optimal time window size as the window, slide on the normalized observation data; for each sliding window, take the observation data of multiple consecutive moments within it as the input of a training sample; taking the optimal prediction time span as the span, take the observation data after the sliding window as the output of the training sample; traverse the entire normalized observation data to obtain the reconstructed observation data set.

[0074] The reconstructed observation dataset is used as the preprocessed historical observation data of the marine environment for subsequent marine environment prediction tasks.

[0075] Through the above technical solutions, the historical observation data of the marine environment can be effectively preprocessed. Data cleaning and missing value interpolation can improve data quality, remove outliers and estimate missing values; normalization processing can eliminate the influence of different observation indicator dimensions, which is convenient for subsequent modeling; based on time scale analysis and cross-validation, the optimal time window size and prediction time span can be determined to optimize the extraction of time series features; the data reconstruction method of sliding windows can retain the time series features while constructing training samples suitable for supervised learning. The preprocessed historical observation data of the marine environment can significantly improve the performance of subsequent prediction models.

[0076] In an optional embodiment,

[0077] The optimal time window size and prediction time span are determined based on time scale analysis and cross-validation methods, including:

[0078] Perform time scale characteristic analysis on the time series of each marine environmental element, identify the periodic components of the time series using autocorrelation analysis, determine the optimal lag order of the time series using partial autocorrelation analysis, obtain the energy distribution of the time series on different frequency components using wavelet analysis and singular spectrum analysis, and determine the candidate value of the time window size of each marine environmental element based on the comprehensive analysis results; determine the candidate value of the prediction time span according to the actual prediction needs;

[0079] A cross-validation method is used to evaluate the performance of the prediction model under different combinations of candidate values ​​of time window sizes and prediction time spans, and the cross-validation method is any one of K-fold cross-validation, leave-one-out cross-validation and time series cross-validation; the K-fold cross-validation randomly divides the observed data into K subsets, selects one subset as the validation set each time, and the remaining subsets are used as training sets, trains the prediction model and evaluates the model performance on the validation set, and after repeating K times, takes the average value of the performance evaluation index as the performance score under the current candidate value combination; the leave-one-out cross-validation uses only one sample as the validation set each time, and the remaining samples as the training set, repeats N times until each sample is used as a validation set once, and takes the average value of the performance evaluation index as the performance score under the current candidate value combination, where N is the total number of samples; the time series cross-validation adopts a forward rolling time series segmentation method, divides the observed data into a training set and a validation set in turn, rolls forward one time step each time, and repeats until the entire time range is covered, and takes the average value of the performance evaluation index as the performance score under the current candidate value combination;

[0080] A strategy combining grid search and random search is adopted to search the candidate value combination space, and the cross-validation is performed on each candidate value combination to obtain its performance score; the candidate value combination with the best performance score is selected to determine the optimal time window size and prediction time span.

[0081] Exemplarily, an autocorrelation analysis is performed on the time series of each marine environmental element to identify the periodic components therein; a partial autocorrelation analysis is performed on the time series of each marine environmental element to determine its optimal lag order; a wavelet analysis and a singular spectrum analysis are performed on the time series of each marine environmental element to obtain its energy distribution on different frequency components; the results of autocorrelation analysis, partial autocorrelation analysis, wavelet analysis and singular spectrum analysis are combined to determine the candidate value of the time window size for each marine environmental element.

[0082] According to the actual needs of marine environment prediction, such as short-term, medium-term or long-term prediction, the candidate values ​​of the prediction time span are determined.

[0083] Select an appropriate cross-validation method, such as K-fold cross-validation, leave-one-out cross-validation, or time series cross-validation; for each candidate value combination of time window size and forecast time span, perform cross-validation:

[0084] If K-fold cross validation is used:

[0085] The observed data is randomly divided into K subsets; each time a subset is selected as the validation set, and the remaining subsets are used as training sets; the prediction model is trained on the training set, and the model performance is evaluated on the validation set; repeat K times, and take the average value of the performance evaluation index as the performance score under the current candidate value combination.

[0086] If leave-one-out cross validation is used:

[0087] Only one sample is used as a validation set each time, and the remaining samples are used as training sets; repeat N times until each sample is used as a validation set once, where N is the total number of samples; take the average value of the performance evaluation index as the performance score under the current candidate value combination.

[0088] If time series cross validation is used:

[0089] The forward rolling time series segmentation method is adopted to divide the observed data into training set and validation set in turn; each time it rolls forward one time step, and repeats until the entire time range is covered; the average value of the performance evaluation index is taken as the performance score under the current candidate value combination.

[0090] Search for the best parameter combination:

[0091] A strategy combining grid search and random search is adopted to search the candidate value combination space of time window size and prediction time span; cross-validation is performed on each candidate value combination to obtain its performance score; the candidate value combination with the best performance score is selected to determine the optimal time window size and prediction time span.

[0092] Through the above technical solutions, the optimal time window size and prediction time span for the marine environmental element prediction task can be effectively determined. Time scale characteristic analysis can reveal the intrinsic characteristics of the time series from different angles and provide a basis for the selection of the time window size; cross-validation evaluation can comprehensively examine the performance of the prediction model under different parameter combinations to avoid overfitting; the strategy of combining grid search and random search can efficiently search for the optimal solution in the parameter space. This solution comprehensively considers the time series characteristics, model performance and computational efficiency, and can provide a reliable time window size and prediction time span selection for the marine environmental element prediction task.

[0093] S102. According to the pre-processed marine environment historical observation data, a recursive neural network model is constructed, wherein the recursive neural network model includes an input layer, two stacked long short-term memory network LSTM layers and an output layer; wherein the input layer is used to receive the observation data at the current moment, and sequentially transmit it to the two LSTM layers, each LSTM layer includes a plurality of LSTM units, each LSTM unit is implemented based on a gating mechanism, and the output results of the two LSTM layers are spliced ​​and transmitted to the output layer, wherein the output layer adopts a fully connected layer structure, realizes nonlinear transformation through an activation function, and outputs the prediction data of the next moment;

[0094] In an optional embodiment,

[0095] Based on the preprocessed historical observation data of the marine environment, a recursive neural network model is constructed, including:

[0096] The input layer of the recursive neural network model receives the preprocessed historical observation data of the ocean environment at the current moment, and transmits the received historical observation data of the ocean environment to the two stacked LSTM layers;

[0097] Each of the two stacked LSTM layers includes a plurality of LSTM units, and the plurality of LSTM units realize adaptive learning and memory of long-term dependencies in the historical observation data of the marine environment through a gating mechanism. Each of the LSTM units includes an input gate, a forget gate, an output gate and a cell state. The input gate controls the proportion of the input information at the current moment updated to the cell state according to the input at the current moment and the hidden state at the previous moment. The forget gate controls the proportion of the cell state at the previous moment that needs to be forgotten according to the input at the current moment and the hidden state at the previous moment. The output gate controls the proportion of the cell state at the current moment output to the hidden state according to the input at the current moment and the hidden state at the previous moment. The cell state is used to store long-term memory information and control the accumulation and forgetting of memory information.

[0098] The output result of the last LSTM layer in the two stacked LSTM layers is passed to the output layer. The output layer maps the output of the last LSTM layer to an output space consistent with the prediction target dimension through a fully connected layer, and realizes nonlinear transformation through an activation function to generate the ocean environment prediction data for the next moment.

[0099] Exemplarily, the input layer of the recursive neural network model receives preprocessed historical observation data of the ocean environment at the current moment; and passes the received historical observation data of the ocean environment to two subsequent stacked LSTM layers.

[0100] The recursive neural network model consists of two stacked LSTM layers; each LSTM layer consists of multiple LSTM units, which realize adaptive learning and memory of long-term dependencies in historical observation data of the marine environment through a gating mechanism;

[0101] Each LSTM unit consists of an input gate, a forget gate, an output gate, and a cell state:

[0102] Input Gate:

[0103] According to the input at the current moment and the hidden state at the previous moment, the activation value of the input gate is calculated; the proportion of the input information at the current moment updated to the cell state is controlled.

[0104] Forget Gate:

[0105] According to the input at the current moment and the hidden state at the previous moment, the activation value of the forget gate is calculated; the proportion of the cell state at the previous moment that needs to be forgotten is controlled.

[0106] Output Gate:

[0107] According to the input at the current moment and the hidden state at the previous moment, the activation value of the output gate is calculated; the ratio of the cell state output at the current moment to the hidden state is controlled.

[0108] Cell Status:

[0109] It is used to store the long-term memory information of the LSTM unit; it is updated through the input gate, forget gate and the cell state of the previous moment to control the accumulation and forgetting of memory information. The stacked LSTM layers are connected in sequence, and the output of the previous LSTM layer is used as the input of the next LSTM layer.

[0110] Output layer:

[0111] The output result of the LSTM layer at the last layer is passed to the output layer; the output layer maps the output of the last LSTM layer to an output space consistent with the prediction target dimension through a fully connected layer; the output layer performs a nonlinear transformation on the output of the fully connected layer through an activation function (such as Sigmoid, tanh or ReLU); the output layer generates the ocean environment prediction data for the next moment.

[0112] Use preprocessed historical marine environment observation data to construct training samples and labels; use appropriate loss functions (such as mean square error or cross entropy) and optimization algorithms (such as Adam or RMSprop) to train the recursive neural network model; optimize model parameters through back propagation algorithm and gradient descent to minimize prediction error.

[0113] Use methods such as holdout or cross-validation to evaluate the performance of the trained model; adjust and optimize the model architecture, hyperparameters, etc. based on the evaluation results to improve the model's predictive performance.

[0114] Through the above technical solutions, a marine environment prediction model based on recursive neural network can be constructed. The model uses stacked LSTM layers to extract features and learn long-term dependencies of historical marine environmental observation data, adaptively memorizes and forgets information through a gating mechanism, and generates prediction results for future moments in the output layer. The training and tuning process of the model can further improve the accuracy and generalization ability of the prediction. This solution fully considers the time series characteristics of marine environmental data and takes advantage of the advantages of recursive neural networks in processing sequence data, providing an effective solution for marine environmental prediction tasks.

[0115] In an optional embodiment,

[0116] The multiple LSTM units realize adaptive learning and memory of long-term dependencies in the historical observation data of the marine environment through a gating mechanism, including:

[0117] The input gate receives the input at the current moment and the hidden state at the previous moment, combines the attention weight obtained by weighted summation of the historical hidden state sequence, and uses the sigmoid activation function to generate a gating signal that controls the proportion of the current input information updated to the cell state;

[0118] The forget gate receives the input at the current moment and the hidden state at the previous moment, introduces sine and cosine functions in the gate calculation to encode the time periodic information, and uses the sigmoid activation function to generate a gate signal that controls the proportion of the cell state at the previous moment that needs to be forgotten;

[0119] The output gate receives the input at the current moment and the hidden state at the previous moment, combines the candidate hidden state calculated by the current input and the hidden state at the previous moment, introduces the update gate mechanism in the gated recursive unit, and uses the sigmoid activation function to generate a gating signal that controls the ratio of the current cell state output to the hidden state;

[0120] Updating the cell state, by integrating the candidate cell state at the current moment, the gating signal of the forget gate and the gating signal of the input gate, the cell state at the previous moment is updated to achieve the selective retention and update of the long-term memory information; the candidate cell state is obtained by using the input at the current moment and the hidden state at the previous moment and applying the tanh activation function to calculate, as a potential cell state update;

[0121] Based on the structure of the LSTM unit, a multi-layer long short-term memory network model is constructed.

[0122] For example,

[0123] Input Gate:

[0124] Receive the current input and the previous hidden state; perform weighted summation on the historical hidden state sequence to obtain the attention weight: use the attention mechanism to calculate the correlation between the current and historical hidden states; convert the correlation into the attention weight through the softmax function. Combine the current input, the previous hidden state and the attention weight, and use the sigmoid activation function to generate the gating signal of the input gate; the gating signal of the input gate controls the proportion of the current input information updated to the cell state.

[0125] Forget Gate:

[0126] Receive the input at the current moment and the hidden state at the previous moment; introduce sine and cosine functions in the gating calculation to encode the time periodicity information: pass the time step information of the current moment into the sine and cosine functions; combine the encoded time periodicity information with the input at the current moment and the hidden state at the previous moment. Use the sigmoid activation function to generate the gating signal of the forget gate; the gating signal of the forget gate controls the proportion of the cell state at the previous moment that needs to be forgotten.

[0127] Output Gate:

[0128] Receive the current input and the previous hidden state; calculate the candidate hidden state: use the current input and the previous hidden state; apply the tanh activation function to calculate. Use the current input and the previous hidden state to calculate the gating signal of the update gate; use the sigmoid activation function to generate the gating signal of the update gate.

[0129] The candidate hidden state and the gating signal of the update gate are combined, and the gating signal of the output gate is generated using the sigmoid activation function; the gating signal of the output gate controls the ratio of the current cell state output to the hidden state.

[0130] Cell status update:

[0131] Calculate the candidate cell state: use the input at the current moment and the hidden state at the previous moment; apply the tanh activation function for calculation.

[0132] The cell state at the previous moment is updated by combining the candidate cell state at the current moment, the gating signal of the forget gate and the gating signal of the input gate; the influence of the candidate cell state, forgetting information and new input information on the cell state is controlled by the gating signal to achieve selective retention and updating of long-term memory information.

[0133] Based on the structure of the above LSTM unit, a multi-layer long short-term memory network model is constructed; when stacking multiple layers of LSTM units, the output of the previous layer of LSTM units is used as the input of the next layer of LSTM units; by stacking multiple layers of LSTM units, adaptive learning and memory of long-term dependencies in historical observation data of the marine environment are achieved.

[0134] Through the above technical solutions, the gating mechanism in the LSTM unit can realize the adaptive learning and memory of the long-term dependencies in the historical observation data of the marine environment. The input gate introduces the attention mechanism to adaptively screen important input information; the forget gate introduces the time periodic information encoding to adaptively forget the historical memory; the output gate introduces the update gate mechanism to adaptively output key information; the cell state integrates the candidate state, forgotten information and new input to achieve the selective retention and update of long-term memory. By constructing a multi-layer long short-term memory network model, the gating mechanism of the LSTM unit can be fully utilized to effectively capture and learn the long-term dependencies in the historical observation data of the marine environment and improve the prediction performance of the model.

[0135] S103. Obtain the ocean environment observation data at the current moment, perform preprocessing operations on the ocean environment observation data at the current moment, and then input it into the trained recursive neural network model; predict the ocean environment data of multiple time steps in the future through the recursive neural network model to obtain the ocean environment prediction results within a preset time; based on the ocean environment prediction results, identify the hot spots of ocean environment changes, and generate visualization results that represent the dynamic change laws of the ocean environment within the prediction time period.

[0136] In an optional embodiment,

[0137] Based on the marine environment prediction results, identify the hot spots of marine environment changes and generate visualization results that show the dynamic changes of the marine environment during the prediction period, including:

[0138] The time series at each grid point in the marine environment prediction results is regarded as an independent signal, and the time series at each grid point is decomposed into time and frequency using wavelet transform, and the dominant cycle and key time points of the changes of marine environmental elements at each grid point are extracted to obtain the time-frequency spectrum coefficients at different scales.

[0139] The wavelet spectrum coefficients of all grid points form a high-dimensional spectrum matrix, and the dynamic time warping (DTW) distance is used to measure the similarity between grid points in the time dimension to obtain a spatiotemporal similarity matrix; the spatiotemporal similarity matrix is ​​spectrally decomposed, and the optimal number of clusters is adaptively determined by analyzing the distribution of eigenvalues, and the eigenvectors corresponding to the first k largest eigenvalues ​​are used as low-dimensional embedding coordinates of the grid points; the low-dimensional embedding coordinates of the grid points are input into the K-means clustering algorithm to obtain the clustering results of the hot spots of marine environmental changes, and each type of cluster corresponds to a hot spot area;

[0140] The clustering results are remapped back to the geographic space, and the average of the geographic coordinates of the internal grid points of each cluster is used as the representative point of the cluster in the geographic space. The standard deviation of the geographic coordinates of the internal grid points of the cluster is used as the radius to draw a spatial distribution map of the hotspot area. The color or filling mode of the hotspot area is set according to the average trend or extreme degree of the change of the marine environment in each hotspot area, and a visualization result representing the dynamic change law of the marine environment within the prediction time period is generated.

[0141] Exemplarily, the time series at each grid point in the marine environment prediction results is regarded as an independent signal; the time series of each grid point is subjected to wavelet transform: the appropriate wavelet basis function and decomposition layer number are selected; the time series is subjected to wavelet decomposition to obtain time-frequency spectrum coefficients at different scales.

[0142] Analyze the time-frequency spectrum coefficients at different scales, identify the dominant period of energy concentration, and determine the key time points when energy mutations or extreme values ​​occur within the dominant period.

[0143] The wavelet spectrum coefficients of all grid points form a high-dimensional spectrum matrix; the dynamic time warping (DTW) distance is used to measure the similarity between grid points in the time dimension: the DTW distance between each pair of grid points is calculated to obtain the spatiotemporal similarity matrix; the DTW distance takes into account the displacement and expansion of the time series and can effectively measure nonlinear similarity.

[0144] Perform spectral decomposition on the spatiotemporal similarity matrix: calculate the eigenvalues ​​and eigenvectors of the spatiotemporal similarity matrix; analyze the distribution of eigenvalues ​​and adaptively determine the optimal number of clusters k. Use the eigenvectors corresponding to the first k largest eigenvalues ​​as the low-dimensional embedding coordinates of the grid points; input the low-dimensional embedding coordinates of the grid points into the K-means clustering algorithm:

[0145] Randomly select k initial cluster centers; assign each grid point to the nearest cluster according to the Euclidean distance between the grid point and the cluster center; update the center of each cluster to the mean of the coordinates of all grid points in the cluster; repeat the iteration until the cluster center no longer changes significantly. The clustering results of the hotspot areas of marine environmental change are obtained, and each type of cluster corresponds to a hotspot area.

[0146] Calculate the average value of the geographic coordinates of the grid points within each cluster as the representative point of the cluster in the geographic space; calculate the standard deviation of the geographic coordinates of the grid points within each cluster as the radius of the cluster in the geographic space.

[0147] With the representative point of each cluster as the center and the radius as the range, draw a circle or other shape to represent the hotspot area; set the color or fill mode of the hotspot area according to the average trend or extreme degree of the change of the marine environment in each hotspot area. On the basis of the spatial distribution map, add time dimension information, such as timeline or animation effects; through the changes of visual elements such as color, size, shape, etc., show the evolution process of the hotspot area in different time periods.

[0148] Through the above technical solutions, the identification and visualization of hotspots of marine environmental changes can be realized. Time-frequency decomposition and spatiotemporal similarity measurement can capture the temporal dynamic characteristics and spatial correlation of marine environmental changes; cluster analysis can adaptively discover hotspots of marine environmental changes; and visualization results can intuitively display the dynamic changes of the marine environment within the forecast period, providing intuitive decision support for marine environmental prediction and analysis. This solution comprehensively utilizes time-frequency analysis, similarity measurement, clustering algorithm and visualization technology, which can effectively mine and present the spatiotemporal patterns of marine environmental changes and improve the interpretability and practicality of marine environmental prediction results.

[0149] In an optional embodiment,

[0150] The wavelet spectrum coefficients of all grid points are used to form a high-dimensional spectrum matrix, and the dynamic time warping DTW distance is used to measure the similarity between grid points in the time dimension to obtain the spatiotemporal similarity matrix, including:

[0151] The high-dimensional spectrum matrix is ​​feature selected by using the mutual information criterion, and the subset of spectrum coefficients that contributes most to the similarity measurement is screened out to optimize the information representation of the spectrum matrix; for any two grid points, the optimized wavelet spectrum coefficient sequence is extracted, and the local distance matrix is ​​constructed by calculating the Mahalanobis distance between different time-frequency components, taking into account the correlation between the time-frequency components;

[0152] On the local distance matrix, a bandwidth-constrained dynamic programming algorithm is used to search for the optimal nonlinear alignment path between two wavelet spectrum coefficient sequences, and a radial basis kernel function is introduced to perform nonlinear transformation on the original cumulative distance to enhance the robustness of the dynamic time warping distance; wherein, according to the physical characteristics of the marine environmental elements, the bandwidth constraint parameters in the optimal alignment path search process are adaptively determined;

[0153] The dynamic time warping distance matrix between all grid point pairs at each time scale is calculated; a spatiotemporal data quality evaluation function is constructed, and the fusion weights of each time scale are determined according to the credibility of the distance matrix at different time scales to generate a spatiotemporal similarity matrix; the regularization term is used to constrain the fusion weights and the spatiotemporal similarity matrix is ​​locally adjusted according to prior knowledge.

[0154] Exemplarily, the mutual information criterion is used to perform feature selection on the high-dimensional spectrum matrix:

[0155] Calculate the mutual information between each spectral coefficient and the similarity measurement target; select the first k spectral coefficients with the highest mutual information as the optimal spectral coefficient subset. According to the selected optimal spectral coefficient subset, construct the optimized spectral matrix; the optimized spectral matrix only contains the spectral coefficients that contribute most to the similarity measurement.

[0156] For any two grid points, the optimized wavelet spectrum coefficient sequence is extracted; the Mahalanobis distance between different time-frequency components is calculated, considering the correlation between time-frequency components; the Mahalanobis distance takes into account the covariance structure of the time-frequency components and can more accurately measure the similarity between time-frequency components.

[0157] On the local distance matrix, a bandwidth-constrained dynamic programming algorithm is used to search for the optimal nonlinear alignment path between two wavelet spectrum coefficient sequences: bandwidth constraint parameters are defined to limit the search range of the alignment path; the dynamic programming algorithm is used to search for the optimal alignment path that satisfies the bandwidth constraint and obtain the original cumulative distance. The radial basis kernel function is introduced to perform a nonlinear transformation on the original cumulative distance: a suitable radial basis kernel function, such as a Gaussian kernel function, is selected; the original cumulative distance is nonlinearly transformed to enhance the robustness of the dynamic time warping distance.

[0158] According to the physical characteristics of marine environmental elements, such as time scale and spatial correlation, the initial value of the bandwidth constraint parameter is set; the performance of the dynamic time warping distance under different bandwidth constraint parameters is evaluated through cross-validation and other methods, and the optimal parameter value is selected. The dynamic time warping distance matrix between all grid point pairs at each time scale is repeatedly calculated; the evaluation function is designed to consider the credibility of the distance matrix at different time scales, such as data missing rate and noise level; by optimizing the evaluation function, the fusion weights of each scale are adaptively learned.

[0159] According to the learned fusion weights, the distance matrices at different time scales are weightedly fused; the regularization term is introduced to constrain the fusion weights to avoid overfitting; the spatiotemporal similarity matrix is ​​locally adjusted based on prior knowledge, such as considering factors such as geographical location and marine environmental characteristics.

[0160] Through the above technical solutions, a fine measurement of the spatiotemporal similarity between grid points can be achieved. Feature selection and spectral matrix optimization can extract the most representative time-frequency information; the construction of the local distance matrix takes into account the correlation between time-frequency components; the calculation of the dynamic time warping distance introduces bandwidth constraints and nonlinear transformations, which improves the robustness of the distance metric; the adaptive learning of the spatiotemporal data quality evaluation function and fusion weights realizes the effective fusion of multi-scale information; the introduction of regularization and prior knowledge further improves the reliability of the spatiotemporal similarity matrix. This solution makes full use of technologies such as time-frequency analysis, similarity measurement, and optimization algorithms, and can accurately capture the spatiotemporal dynamic characteristics of marine environmental changes, providing a high-quality spatiotemporal similarity matrix for subsequent clustering analysis and visualization.

[0161] Figure 2 FIG. 1 is a schematic diagram of the structure of a marine environment time series prediction system based on a recursive neural network according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0162] The first unit is used to obtain historical observation data of the marine environment, pre-process the historical observation data of the marine environment, and remove abnormal values ​​and missing values ​​therein, wherein the historical observation data of the marine environment includes at least two of sea surface temperature, sea surface salinity, sea surface density, sea surface height, ocean current speed, ocean current direction and wind speed; wherein the historical observation data of the marine environment is time series data, including observation data of multiple time steps;

[0163] The second unit is used to construct a recursive neural network model based on the preprocessed historical observation data of the marine environment, and the recursive neural network model includes an input layer, two stacked long short-term memory network LSTM layers and an output layer; wherein the input layer is used to receive the observation data at the current moment, and pass it to the two LSTM layers in sequence, each LSTM layer includes a plurality of LSTM units, and each LSTM unit is implemented based on a gating mechanism, and the output results of the two LSTM layers are spliced ​​and passed to the output layer, and the output layer adopts a fully connected layer structure, realizes nonlinear transformation through an activation function, and outputs the predicted data of the next moment;

[0164] The third unit is used to obtain the ocean environment observation data at the current moment, pre-process the ocean environment observation data at the current moment, and then input it into the trained recursive neural network model; through the recursive neural network model, predict the ocean environment data of multiple time steps in the future, and obtain the ocean environment prediction results within a preset time; based on the ocean environment prediction results, identify the hot spots of ocean environment changes, and generate visualization results that represent the dynamic change laws of the ocean environment within the prediction time period.

[0165] According to a third aspect of the embodiments of the present invention,

[0166] An electronic device is provided, comprising:

[0167] processor;

[0168] a memory for storing processor-executable instructions;

[0169] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0170] According to a fourth aspect of the embodiments of the present invention,

[0171] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0172] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0173] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting ocean environment time series based on recursive neural network, characterized in that: include: Acquire historical observation data of the marine environment, pre-process the historical observation data of the marine environment, and remove abnormal values ​​and missing values ​​therein, wherein the historical observation data of the marine environment includes at least two of sea surface temperature, sea surface salinity, sea surface density, sea surface height, ocean current speed, ocean current direction, and wind speed; wherein the historical observation data of the marine environment is time series data, including observation data of multiple time steps; According to the preprocessed historical observation data of the marine environment, a recursive neural network model is constructed, and the recursive neural network model includes an input layer, two stacked long short-term memory network LSTM layers and an output layer; wherein the input layer is used to receive the observation data at the current moment, and sequentially transmit it to the two LSTM layers, each LSTM layer includes a plurality of LSTM units, and each LSTM unit is implemented based on a gating mechanism, and the output results of the two LSTM layers are spliced ​​and transmitted to the output layer, and the output layer adopts a fully connected layer structure, realizes nonlinear transformation through an activation function, and outputs the predicted data of the next moment; Obtain the current ocean environment observation data, perform preprocessing operations on the current ocean environment observation data, and then input it into the trained recursive neural network model; predict the ocean environment data of multiple time steps in the future through the recursive neural network model to obtain the ocean environment prediction results within a preset time; identify the hot spots of ocean environment changes based on the ocean environment prediction results, and generate visualization results that represent the dynamic change rules of the ocean environment within the prediction time period; Based on the marine environment prediction results, identify the hot spots of marine environment changes and generate visualization results that show the dynamic changes of the marine environment during the prediction period, including: The time series at each grid point in the marine environment prediction results is regarded as an independent signal, and the time series at each grid point is decomposed into time and frequency using wavelet transform, and the dominant cycle and key time points of the changes of marine environmental elements at each grid point are extracted to obtain the time-frequency spectrum coefficients at different scales. The wavelet spectrum coefficients of all grid points form a high-dimensional spectrum matrix, and the dynamic time warping (DTW) distance is used to measure the similarity between grid points in the time dimension to obtain a spatiotemporal similarity matrix; the spatiotemporal similarity matrix is ​​spectrally decomposed, and the optimal number of clusters is adaptively determined by analyzing the distribution of eigenvalues, and the eigenvectors corresponding to the first k largest eigenvalues ​​are used as low-dimensional embedding coordinates of the grid points; the low-dimensional embedding coordinates of the grid points are input into the K-means clustering algorithm to obtain the clustering results of the hot spots of marine environmental changes, and each type of cluster corresponds to a hot spot area; The clustering results are remapped back to the geographic space, and the average of the geographic coordinates of the internal grid points of each cluster is used as the representative point of the cluster in the geographic space. The standard deviation of the geographic coordinates of the internal grid points of the cluster is used as the radius to draw a spatial distribution map of the hotspot area. The color or filling mode of the hotspot area is set according to the average trend or extreme degree of the change of the marine environment in each hotspot area, and a visualization result representing the dynamic change law of the marine environment within the prediction time period is generated.

2. The method according to claim 1, characterized in that: Preprocessing the historical marine environment observation data includes: Performing data cleaning on each observation indicator in the historical observation data of the marine environment, identifying outliers that deviate from the mean within a range of plus or minus three times the standard deviation, and replacing the outliers with the average values ​​before and after the outliers, to obtain observation indicator data with outliers removed; Interpolating the missing values ​​in the observed indicators after removing the outliers, estimating the missing values ​​by calculating the valid observed values ​​before and after the missing values ​​using the cubic spline interpolation method, and obtaining the observed indicator data without missing values; Normalizing the observed indicators without missing values, calculating the 0.05 quantile and 0.95 quantile of each observed indicator as the minimum and maximum values ​​of the observed indicator, mapping each observed indicator to the interval of 0 to 1, and obtaining normalized observed data; Based on the time scale analysis and cross-validation method, the optimal time window size and prediction time span are determined, and the normalized observation data are reconstructed with the optimal time window size as the window by using the sliding time window method. Each training sample is composed of observation data at multiple consecutive moments as input, and observation data with the optimal prediction time span as the span is output to obtain a reconstructed observation data set, which is the preprocessed historical observation data of the marine environment.

3. The method according to claim 2, characterized in that The optimal time window size and prediction time span are determined based on time scale analysis and cross-validation methods, including: Perform time scale characteristic analysis on the time series of each marine environmental element, identify the periodic components of the time series using autocorrelation analysis, determine the optimal lag order of the time series using partial autocorrelation analysis, obtain the energy distribution of the time series on different frequency components using wavelet analysis and singular spectrum analysis, and determine the candidate value of the time window size of each marine environmental element based on the comprehensive analysis results; determine the candidate value of the prediction time span according to the actual prediction needs; A cross-validation method is used to evaluate the performance of the prediction model under different combinations of candidate values ​​of time window sizes and prediction time spans, and the cross-validation method is any one of K-fold cross-validation, leave-one-out cross-validation and time series cross-validation; the K-fold cross-validation randomly divides the observed data into K subsets, selects one subset as the validation set each time, and the remaining subsets are used as training sets, trains the prediction model and evaluates the model performance on the validation set, and after repeating K times, takes the average value of the performance evaluation index as the performance score under the current candidate value combination; the leave-one-out cross-validation uses only one sample as the validation set each time, and the remaining samples as the training set, repeats N times until each sample is used as a validation set once, and takes the average value of the performance evaluation index as the performance score under the current candidate value combination, where N is the total number of samples; the time series cross-validation adopts a forward rolling time series segmentation method, divides the observed data into a training set and a validation set in turn, rolls forward one time step each time, and repeats until the entire time range is covered, and takes the average value of the performance evaluation index as the performance score under the current candidate value combination; A strategy combining grid search and random search is adopted to search the candidate value combination space, and the cross-validation is performed on each candidate value combination to obtain its performance score; the candidate value combination with the best performance score is selected to determine the optimal time window size and prediction time span.

4. The method according to claim 1, characterized in that: Based on the preprocessed historical observation data of the marine environment, a recursive neural network model is constructed, including: The input layer of the recursive neural network model receives the preprocessed historical observation data of the ocean environment at the current moment, and transmits the received historical observation data of the ocean environment to the two stacked LSTM layers; Each of the two stacked LSTM layers includes a plurality of LSTM units, and the plurality of LSTM units realize adaptive learning and memory of long-term dependencies in the historical observation data of the marine environment through a gating mechanism. Each of the LSTM units includes an input gate, a forget gate, an output gate and a cell state. The input gate controls the proportion of the input information at the current moment updated to the cell state according to the input at the current moment and the hidden state at the previous moment. The forget gate controls the proportion of the cell state at the previous moment that needs to be forgotten according to the input at the current moment and the hidden state at the previous moment. The output gate controls the proportion of the cell state at the current moment output to the hidden state according to the input at the current moment and the hidden state at the previous moment. The cell state is used to store long-term memory information and control the accumulation and forgetting of memory information. The output result of the last LSTM layer in the two stacked LSTM layers is passed to the output layer. The output layer maps the output of the last LSTM layer to an output space consistent with the prediction target dimension through a fully connected layer, and realizes nonlinear transformation through an activation function to generate the ocean environment prediction data for the next moment.

5. The method according to claim 4, characterized in that The multiple LSTM units realize adaptive learning and memory of long-term dependencies in the historical observation data of the marine environment through a gating mechanism, including: The input gate receives the input at the current moment and the hidden state at the previous moment, combines the attention weight obtained by weighted summation of the historical hidden state sequence, and uses the sigmoid activation function to generate a gating signal that controls the proportion of the current input information updated to the cell state; The forget gate receives the input at the current moment and the hidden state at the previous moment, introduces sine and cosine functions in the gate calculation to encode the time periodic information, and uses the sigmoid activation function to generate a gate signal that controls the proportion of the cell state at the previous moment that needs to be forgotten; The output gate receives the input at the current moment and the hidden state at the previous moment, combines the candidate hidden state calculated by the current input and the hidden state at the previous moment, introduces the update gate mechanism in the gated recursive unit, and uses the sigmoid activation function to generate a gating signal that controls the ratio of the current cell state output to the hidden state; Updating the cell state, by integrating the candidate cell state at the current moment, the gating signal of the forget gate and the gating signal of the input gate, the cell state at the previous moment is updated to achieve the selective retention and update of the long-term memory information; the candidate cell state is obtained by using the input at the current moment and the hidden state at the previous moment and applying the tanh activation function to calculate, as a potential cell state update; Based on the structure of the LSTM unit, a multi-layer long short-term memory network model is constructed.

6. The method according to claim 1, characterized in that The wavelet spectrum coefficients of all grid points are used to form a high-dimensional spectrum matrix, and the dynamic time warping DTW distance is used to measure the similarity between grid points in the time dimension to obtain the spatiotemporal similarity matrix, including: The high-dimensional spectrum matrix is ​​feature selected by using the mutual information criterion, and the subset of spectrum coefficients that contributes most to the similarity measurement is screened out to optimize the information representation of the spectrum matrix; for any two grid points, the optimized wavelet spectrum coefficient sequence is extracted, and the local distance matrix is ​​constructed by calculating the Mahalanobis distance between different time-frequency components, taking into account the correlation between the time-frequency components; On the local distance matrix, a bandwidth-constrained dynamic programming algorithm is used to search for the optimal nonlinear alignment path between two wavelet spectrum coefficient sequences, and a radial basis kernel function is introduced to perform nonlinear transformation on the original cumulative distance to enhance the robustness of the dynamic time warping distance; wherein, according to the physical characteristics of the marine environmental elements, the bandwidth constraint parameters in the optimal alignment path search process are adaptively determined; The dynamic time warping distance matrix between all grid point pairs at each time scale is calculated; a spatiotemporal data quality evaluation function is constructed, and the fusion weights of each time scale are determined according to the credibility of the distance matrix at different time scales to generate a spatiotemporal similarity matrix; the regularization term is used to constrain the fusion weights and the spatiotemporal similarity matrix is ​​locally adjusted according to prior knowledge.

7. A marine environment time series prediction system based on a recursive neural network, used to implement the method described in any one of claims 1 to 6, characterized in that: include: The first unit is used to obtain historical observation data of the marine environment, pre-process the historical observation data of the marine environment, and remove abnormal values ​​and missing values ​​therein, wherein the historical observation data of the marine environment includes at least two of sea surface temperature, sea surface salinity, sea surface density, sea surface height, ocean current speed, ocean current direction and wind speed; wherein the historical observation data of the marine environment is time series data, including observation data of multiple time steps; The second unit is used to construct a recursive neural network model based on the preprocessed historical observation data of the marine environment, and the recursive neural network model includes an input layer, two stacked long short-term memory network LSTM layers and an output layer; wherein the input layer is used to receive the observation data at the current moment, and pass it to the two LSTM layers in sequence, each LSTM layer includes a plurality of LSTM units, and each LSTM unit is implemented based on a gating mechanism, and the output results of the two LSTM layers are spliced ​​and passed to the output layer, and the output layer adopts a fully connected layer structure, realizes nonlinear transformation through an activation function, and outputs the predicted data of the next moment; The third unit is used to obtain the ocean environment observation data at the current moment, pre-process the ocean environment observation data at the current moment, and then input it into the trained recursive neural network model; through the recursive neural network model, predict the ocean environment data of multiple time steps in the future, and obtain the ocean environment prediction results within a preset time; based on the ocean environment prediction results, identify the hot spots of ocean environment changes, and generate visualization results that represent the dynamic change laws of the ocean environment within the prediction time period.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images