An effective wave height prediction method based on feature engineering and elder brother guidance mechanism

Abstract

The application discloses an effective wave height prediction method based on feature engineering and elder brother guidance mechanism. The method comprises the following steps: acquiring a sea wave feature dataset containing an original wave height time sequence signal; performing modal decomposition on the original wave height time sequence signal to obtain a plurality of intrinsic mode functions and a plurality of residual terms; adding the plurality of intrinsic mode functions and the plurality of residual terms to the sea wave feature dataset to construct an enhanced feature dataset; performing feature screening on the features in the enhanced feature dataset to obtain a screened enhanced feature dataset; constructing an effective wave height prediction network model based on a gated recurrent unit and an elder brother guidance mechanism; and inputting the screened enhanced feature dataset into the effective wave height prediction network model for prediction to obtain an effective wave height prediction result. The application can improve the prediction accuracy of the effective wave height.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of sea wave prediction, in particular to an effective wave height prediction method based on feature engineering and elder brother guidance mechanism. BACKGROUND

[0002] Sea wave is an important ocean phenomenon driven by wind field, and is an important part of physical oceanography. Its properties can be characterized by multiple ocean parameters: effective wave height, average wave height, average wave period, etc. Among them, the effective wave height (SWH) is the most important parameter in the sea wave, and the accurate prediction of SWH is also the main research direction of ocean prediction.

[0003] Some people combine XgBoost and LSTM model to avoid model overfitting and data missing problems; the combination of signal decomposition method and deep learning network model is proved to effectively improve the correlation and reduce the error in more and more experiments. Some people combine wavelet decomposition algorithm EMD and LSTM model, and prove that it has good superiority in predicting nonlinear and non-stationary waves. A random decomposition wave method of DSD is combined with LSTM, and better results than wavelet decomposition are achieved in multivariate input prediction. However, the existing deep learning method still pays little attention to the problems of low computational efficiency and long sequence information decay caused by long-term prediction. In recent years, data-based deep learning models such as LSTM and RNN have been widely used in single-point sea wave prediction, but these methods have been proved to have high lag and non-stationary, and poor generalization ability.

[0004] In summary, the existing related methods have relatively low prediction accuracy of effective wave height. SUMMARY

[0005] The application aims to provide an effective wave height prediction method based on feature engineering and elder brother guidance mechanism, which can improve the prediction accuracy of effective wave height.

[0006] The application embodiment provides an effective wave height prediction method based on feature engineering and elder brother guidance mechanism, which comprises the following steps:

[0007] Obtain a sea wave feature data set containing an original wave height time series signal;

[0008] Perform modal decomposition on the original wave height time series signal to obtain a plurality of intrinsic mode functions and a plurality of residual terms;

[0009] Add the plurality of intrinsic mode functions and the plurality of residual terms to the sea wave feature data set to construct an enhanced feature data set;

[0010] Perform feature screening on the features in the enhanced feature data set to obtain a screened enhanced feature data set;

[0011] An effective wave height prediction network model is constructed based on a gating recurrent unit and a big brother guidance mechanism.

[0012] The filtered enhanced feature dataset is input into the effective wave height prediction network model for prediction to obtain an effective wave height prediction result.

[0013] In some embodiments, the effective wave height prediction network model is constructed based on a gating recurrent unit and a big brother guidance mechanism, including:

[0014] A plurality of cascaded gating recurrent units are used to construct an encoder and a decoder.

[0015] The effective wave height prediction network model is constructed according to the encoder and the decoder, wherein the encoder is used as a big brother of the decoder to guide the decoder to perform prediction.

[0016] In some embodiments, the filtered enhanced feature dataset is input into the effective wave height prediction network model for prediction to obtain an effective wave height prediction result, including:

[0017] The filtered enhanced feature dataset is processed by a plurality of gating recurrent units in the encoder to obtain a last hidden state of the encoder.

[0018] The last hidden state of the encoder is used as an initial hidden state of the decoder, and the last hidden state of the encoder is used as a semantic encoding learned by the encoder.

[0019] The semantic encoding is normalized to obtain a weighting parameter.

[0020] The weighting parameter is subjected to Hadamard product with a last time sequence feature in the filtered enhanced feature dataset, and after linear layer operation, a first input of the decoder is obtained.

[0021] The first input of the decoder and the initial hidden state of the decoder are input into the decoder for prediction to obtain an effective wave height prediction result.

[0022] In some embodiments, the semantic encoding is normalized to obtain a weighting parameter, including:

[0023] ;

[0024] ;

[0025] wherein, the semantic encoding is denoted as, the last hidden state of the encoder is denoted as, represents the weighting parameter, represents a normalization layer.

[0026] In some embodiments, processing the filtered enhanced feature dataset using a plurality of gated recurrent units in the encoder includes:

[0027] ;

[0028] ;

[0029] ;

[0030] ;

[0031] in, Indicates the The encoder updates the gate at time , represents the activation function, represents the weight matrix of the update gate, Indicates the The encoder hidden state at time t, represents the encoder input features, represents the bias vector of the update gate, Indicates the encoder reset gate, represents the weight matrix of the reset gate, represents the bias vector for resetting the gate, represents the encoder candidate hidden state, represents the hyperbolic tangent function, The weight matrix representing the candidate hidden states of the encoder, The bias vector representing the candidate hidden state of the encoder, represents the Hadamard product.

[0032] In some embodiments, performing a Hadamard product on the weighted parameter and the last temporal feature in the filtered enhanced feature dataset and performing a linear layer operation to obtain a first input of the decoder includes:

[0033] ;

[0034] in, represents the first input of the decoder, represents the weight matrix, represents the weighting parameter, represents the Hadamard product, Represents the last time series feature in the filtered enhanced feature dataset, represents the bias vector, represents a linear layer operation.

[0035] In some embodiments, the step of inputting the first input of the decoder and the initial hidden state of the decoder into the decoder to obtain the predicted significant wave height prediction result comprises:

[0036] ;

[0037] ;

[0038] ;

[0039] ;

[0040] ;

[0041] ;

[0042] wherein, represents an input of the decoder at the t-th time point, represents a weight matrix, represents a weight parameter, represents a Hadamard product, represents the last time sequence feature in the filtered enhanced feature dataset, represents a bias vector, represents an input of the decoder at the t-th time point, represents a decoder update gate at the t-th time point, represents an activation function, represents a weight matrix of the decoder update gate, represents a hidden state of the decoder at the t-th time point, represents a bias vector of the update gate, represents a decoder reset gate, represents a weight matrix of the decoder reset gate, represents a bias vector of the reset gate, represents a decoder candidate hidden state, represents a hyperbolic tangent function, represents a weight matrix of the decoder candidate hidden state, represents a bias vector of the decoder candidate hidden state, represents a predicted significant wave height prediction result at the t-th time point.

[0043] ​​​​​In some embodiments, the modal decomposition of the original wave height time series signal comprises:

[0044] The modal decomposition of the original wave height time series signal is performed by using adaptive noise complete ensemble empirical mode decomposition to obtain a plurality of intrinsic mode functions and a plurality of residual terms.

[0045] In some embodiments, the feature screening of the features in the enhanced feature dataset comprises:

[0046] The features in the enhanced feature dataset are sorted by using an XGBoost regression model to obtain a sorted enhanced feature dataset;

[0047] The features in the sorted enhanced feature dataset are sequentially input into the XGBoost regression model for prediction to obtain a current prediction value, and a target mean absolute percentage error between the current prediction value and a true value is calculated;

[0048] If the target mean absolute percentage error is smaller than a historical mean absolute percentage error, the features input into the XGBoost regression model are retained until a plurality of features are sequentially input into the XGBoost regression model, the target mean absolute percentage error calculated is equal to or greater than the historical mean absolute percentage error, the iterative prediction is ended, and a screened enhanced feature dataset is obtained.

[0049] In some embodiments, the feature screening of the features in the enhanced feature dataset comprises:

[0050] The features in the enhanced feature dataset are sorted by using an XGBoost regression model to obtain a sorted enhanced feature dataset;

[0051] The features in the sorted enhanced feature dataset are sequentially input into the XGBoost regression model for prediction to obtain a current prediction value, and a target mean absolute percentage error between the current prediction value and a true value is calculated;

[0052] If the target mean absolute percentage error is greater than or equal to a historical mean absolute percentage error, the features input into the XGBoost regression model are removed until a plurality of features are sequentially input into the XGBoost regression model, the target mean absolute percentage error calculated is greater than or equal to the historical mean absolute percentage error, the iterative prediction is ended, and a screened enhanced feature dataset is obtained.

[0053] Compared with the prior art, the present application has the following beneficial effects:

[0054] The method comprises the following steps: acquiring a sea wave feature dataset containing an original wave height time sequence signal; performing modal decomposition on the original wave height time sequence signal to obtain a plurality of intrinsic mode functions and a plurality of residual terms; adding the plurality of intrinsic mode functions and the plurality of residual terms to the sea wave feature dataset to construct an enhanced feature dataset; performing feature screening on the features in the enhanced feature dataset to obtain a screened enhanced feature dataset; constructing an effective wave height prediction network model based on a gated recurrent unit and using an elder guiding mechanism; and inputting the screened enhanced feature dataset into the effective wave height prediction network model for prediction to obtain an effective wave height prediction result. In this way, through modal decomposition and feature screening, the model can better learn the nonlinear relationship between beneficial features and prediction results, and then an effective wave height prediction network model is constructed through the elder guiding mechanism, which alleviates the error accumulation caused by model autoregression and enhances the model's ability to resist prediction lag, thereby improving the prediction accuracy of the effective wave height. BRIEF DESCRIPTION OF DRAWINGS

[0055] The above and / or additional aspects and advantages of the present application will become apparent and more readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings, in which:

[0056] Figure 1 is a flowchart of an embodiment of the effective wave height prediction method based on feature engineering and elder guiding mechanism provided by the present application;

[0057] Figure 2 is a structural diagram of a sequence-to-sequence neural network model based on an elder guiding mechanism in the best embodiment of the effective wave height prediction method based on feature engineering and elder guiding mechanism provided by the present application;

[0058] Figure 3 is a diagram showing the change of MAE and MAPE with the number of features K in the best embodiment of the effective wave height prediction method based on feature engineering and elder guiding mechanism provided by the present application;

[0059] Figure 4 is a diagram showing the SHAP value swarm plot in the best embodiment of the effective wave height prediction method based on feature engineering and elder guiding mechanism provided by the present application. DETAILED DESCRIPTION

[0060] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, in which the same or similar reference numerals represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present application, and cannot be understood as a limitation of the present application.

[0061] In the description of the present application, if the first, second, etc. are described, it is only for the purpose of distinguishing technical features, and cannot be understood as indicating or implying the relative importance of the indicated technical features or implying the number of the indicated technical features or the order of the indicated technical features.

[0062] In the description of the present application, it should be understood that the orientation description, such as the orientation or position relationship indicated by up, down, etc. is based on the orientation or position relationship shown in the drawings, only for the purpose of facilitating the description of the present application and simplifying the description, and therefore cannot be understood as indicating or implying that the indicated device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present application.

[0063] In the description of the present application, it should be noted that, unless otherwise explicitly limited, the words such as setting, installing, connecting, etc. should be broadly understood, and the person skilled in the art can reasonably determine the specific meaning of the above words in the present application in combination with the specific content of the technical solution.

[0064] Since the existing deep learning method still pays little attention to the problem of low computational efficiency and long sequence information decay caused by long-term prediction. In recent years, data-based deep learning models such as LSTM and RNN have been widely used in single-point sea wave prediction, but these methods have been proved to have high lag and non-stationary, and poor generalization ability.

[0065] In order to solve the problem that the prediction accuracy of the effective wave height of the existing related method is relatively low, the present application provides an effective wave height prediction method based on feature engineering and elder brother guidance mechanism.

[0066] Referring to Figure 1 , the effective wave height prediction method based on feature engineering and elder brother guidance mechanism provided by the embodiments of the present application. The effective wave height prediction method based on feature engineering and elder brother guidance mechanism is applied to an electronic device, which can be a server or a mobile terminal, etc. As Figure 1 shown, the effective wave height prediction method based on feature engineering and elder brother guidance mechanism can include the following steps:

[0067] Step S100, obtaining a sea wave feature data set containing an original wave height time series signal;

[0068] Step S200, modal decomposition is performed on the original wave height time series signal to obtain a plurality of intrinsic mode functions and a plurality of residual terms;

[0069] Step S300, adding the plurality of intrinsic mode functions and the plurality of residual terms to the sea wave feature data set to construct an enhanced feature data set;

[0070] Step S400, feature screening is performed on the features in the enhanced feature dataset to obtain a screened enhanced feature dataset.

[0071] Step S500, an effective wave height prediction network model is constructed based on the gated recurrent unit and using the elder guidance mechanism.

[0072] Step S600, the screened enhanced feature dataset is input into the effective wave height prediction network model for prediction to obtain an effective wave height prediction result.

[0073] In the embodiment, the sea wave feature dataset containing the original wave height time series signal is obtained; the original wave height time series signal is subjected to modal decomposition to obtain a plurality of intrinsic mode functions and a plurality of residual terms; the plurality of intrinsic mode functions and the plurality of residual terms are added to the sea wave feature dataset to construct an enhanced feature dataset; feature screening is performed on the features in the enhanced feature dataset to obtain a screened enhanced feature dataset; an effective wave height prediction network model is constructed based on the gated recurrent unit and using the elder guidance mechanism; and the screened enhanced feature dataset is input into the effective wave height prediction network model for prediction to obtain an effective wave height prediction result. In this way, through modal decomposition and feature screening, the model can better learn the nonlinear relationship between the beneficial features and the prediction result, and then the effective wave height prediction network model is constructed through the elder guidance mechanism, which alleviates the error accumulation caused by the model autoregression and enhances the model's ability to resist prediction lag, thereby improving the prediction accuracy of the effective wave height.

[0074] The above-mentioned sea wave feature dataset containing the original wave height time series signal can be obtained by obtaining single-point sea wave prediction data, which can include original wave height time series signal and other data, and the embodiment is not limited specifically.

[0075] The above-mentioned modal decomposition of the original wave height time series signal can be adaptive noise complete ensemble empirical mode decomposition (CEEMDAN) or other modal decomposition techniques known to those skilled in the art, and the embodiment is not limited specifically.

[0076] In some embodiments, the effective wave height prediction network model is constructed based on the gated recurrent unit and using the elder guidance mechanism, including:

[0077] A plurality of cascaded gated recurrent units are used to construct an encoder and a decoder.

[0078] The effective wave height prediction network model is constructed according to the encoder and the decoder, wherein the encoder is used as the elder of the decoder to guide the decoder to perform prediction.

[0079] In the embodiment, the encoder and the decoder are constructed by adopting a plurality of cascaded gated recurrent units; and the effective wave height prediction network model is constructed according to the encoder and the decoder, wherein the encoder is used as an elder brother of the decoder to guide the decoder to perform prediction. In this way, the encoder is used as the elder brother of the decoder, the encoder guides the decoder to perform prediction, the encoder and the decoder perform parallel prediction through the elder brother guiding mechanism, error accumulation is reduced, and the prediction accuracy of the effective wave height is improved.

[0080] In some embodiments, the screened enhanced feature dataset is input into the effective wave height prediction network model for prediction to obtain an effective wave height prediction result, including:

[0081] The screened enhanced feature dataset is processed by the plurality of gated recurrent units in the encoder to obtain a last hidden state of the encoder;

[0082] The last hidden state of the encoder is used as an initial hidden state of the decoder, and the last hidden state of the encoder is used as semantic encoding learned by the encoder;

[0083] The semantic encoding is normalized to obtain a weighting parameter;

[0084] The weighting parameter and a last time sequence feature in the screened enhanced feature dataset are subjected to Hadamard product, and after linear layer operation, a first input of the decoder is obtained;

[0085] The first input of the decoder and the initial hidden state of the decoder are input into the decoder for prediction to obtain the effective wave height prediction result.

[0086] In the embodiment, the screened enhanced feature dataset is processed by the plurality of gated recurrent units in the encoder to obtain a last hidden state of the encoder; the last hidden state of the encoder is used as an initial hidden state of the decoder, and the last hidden state of the encoder is used as semantic encoding learned by the encoder; the semantic encoding is normalized to obtain a weighting parameter; the weighting parameter and a last time sequence feature in the screened enhanced feature dataset are subjected to Hadamard product, and after linear layer operation, a first input of the decoder is obtained; the first input of the decoder and the initial hidden state of the decoder are input into the decoder for prediction to obtain the effective wave height prediction result. In this way, in the decoding stage of the decoder, the encoder receives the input data at the end of the encoder again, and the prediction output (i.e., the last hidden state of the encoder is used as the semantic encoding learned by the encoder) is given to the decoder as input, so as to alleviate the error accumulation caused by the autoregressive prediction, and better resist the lag of prediction.

[0087] In some embodiments, the semantic encoding is normalized to obtain a weighting parameter, including:

[0088] ;

[0089] ;

[0090] wherein, denotes semantic encoding, denotes the last hidden state of the encoder, denotes a weighting parameter, denotes a normalization layer.

[0091] In some embodiments, the screened enhanced feature dataset is processed by a plurality of gated recurrent units in the encoder, comprising:

[0092] ;

[0093] ;

[0094] ;

[0095] ;

[0096] wherein, denotes an encoder update gate at the time step, denotes an activation function, denotes a weight matrix of the update gate, denotes an encoder hidden state at the time step, denotes an encoder input feature, denotes a bias vector of the update gate, denotes an encoder reset gate, denotes a weight matrix of the reset gate, denotes a bias vector of the reset gate, denotes an encoder candidate hidden state, denotes a hyperbolic tangent function, denotes a weight matrix of the encoder candidate hidden state, denotes a bias vector of the encoder candidate hidden state, denotes a Hadamard product.

[0097] In some embodiments, the weighting parameter is Hadamard multiplied with the last time step feature in the screened enhanced feature dataset, and after a linear layer operation, the first input of the decoder is obtained, comprising:

[0098] ;

[0099] wherein, represents the first input of the decoder, represents the weight matrix, represents the weighting parameter, represents the Hadamard product, Represents the last time series feature in the filtered enhanced feature dataset, represents the bias vector, Represents a linear layer operation.

[0100] In some embodiments, the first input of the decoder and the initial hidden state of the decoder are input into the decoder for prediction to obtain a significant wave height prediction result, including:

[0101] ;

[0102] ;

[0103] ;

[0104] ;

[0105] ;

[0106] ;

[0107] in, Indicates the The input of the time decoder, represents the weight matrix, represents the weighting parameter, represents the Hadamard product, Represents the last time series feature in the filtered enhanced feature dataset, represents the bias vector, Indicates the The input of the time decoder, Indicates the The decoder updates the gate at time , represents the activation function, represents the weight matrix of the update gate, Indicates the The decoder hidden state at time t, represents the bias vector of the update gate, Indicates the decoder reset gate, represents the weight matrix of the reset gate, represents the bias vector for resetting the gate, represents the decoder candidate hidden state, represents the hyperbolic tangent function, The weight matrix representing the candidate hidden states of the decoder, a bias vector representing a decoder candidate hidden state, a prediction result of the significant wave height at the

[0108] In some embodiments, the original wave height time series signal is subjected to modal decomposition to obtain a plurality of intrinsic mode functions and a plurality of residual terms, including:

[0109] The original wave height time series signal is subjected to modal decomposition by adaptive noise complete ensemble empirical mode decomposition to obtain a plurality of intrinsic mode functions and a plurality of residual terms.

[0110] In this embodiment, the original wave height time series signal is subjected to modal decomposition by adaptive noise complete ensemble empirical mode decomposition to obtain intrinsic mode functions (i.e., IMFs) that separate signals from high frequency to low frequency in turn, i.e., to obtain detail components. Through noise-assisted decomposition and multi-scale integrated operation, the frequency domain completeness and low reconstruction error characteristics of each IMF are ensured. Adding these detail components to the original feature data set can enhance the feature data set and lay a good data foundation for later accurate prediction of the significant wave height.

[0111] In some embodiments, the features in the enhanced feature data set are subjected to feature screening to obtain a screened enhanced feature data set, including:

[0112] The features in the enhanced feature data set are sorted by an XGBoost regression model to obtain a sorted enhanced feature data set.

[0113] The features in the sorted enhanced feature data set are sequentially input into the XGBoost regression model for prediction to obtain a current prediction value, and a target mean absolute percentage error between the current prediction value and the true value is calculated.

[0114] If the target mean absolute percentage error is smaller than the historical mean absolute percentage error, the features input into the XGBoost regression model are retained, until a plurality of features are continuously input into the XGBoost regression model, the target mean absolute percentage error calculated is equal to or greater than the historical mean absolute percentage error, the iterative prediction is ended, and a screened enhanced feature data set is obtained.

[0115] ​In the embodiment, the features in the enhanced feature dataset are sorted by adopting the XGBoost regression model to obtain a sorted enhanced feature dataset; the features in the sorted enhanced feature dataset are sequentially input into the XGBoost regression model in order to obtain a current prediction value, and a target mean absolute percentage error between the current prediction value and a true value is calculated; if the target mean absolute percentage error is smaller than a historical mean absolute percentage error, the features input into the XGBoost regression model are retained, until a plurality of features are sequentially input into the XGBoost regression model, the target mean absolute percentage error calculated is equal to or greater than the historical mean absolute percentage error, the iterative prediction is ended, and a screened enhanced feature dataset is obtained. In this way, the XGBoost regression model is used to select the modal beneficial to the prediction into the feature subset (i.e., the screened enhanced feature dataset is obtained), which can make the effective wave height prediction network model better learn the nonlinear relationship between the beneficial features and the prediction result, thereby improving the prediction accuracy of the effective wave height.

[0116] In some embodiments, the features in the enhanced feature dataset are screened to obtain a screened enhanced feature dataset, including:

[0117] The features in the enhanced feature dataset are sorted by adopting the XGBoost regression model to obtain a sorted enhanced feature dataset;

[0118] The features in the sorted enhanced feature dataset are sequentially input into the XGBoost regression model in order to obtain a current prediction value, and a target mean absolute percentage error between the current prediction value and a true value is calculated;

[0119] If the target mean absolute percentage error is greater than or equal to a historical mean absolute percentage error, the features input into the XGBoost regression model are removed, until a plurality of features are sequentially input into the XGBoost regression model, the target mean absolute percentage error calculated is greater than or equal to the historical mean absolute percentage error, the iterative prediction is ended, and a screened enhanced feature dataset is obtained.

[0120] In the embodiment, the features in the enhanced feature dataset are sorted by adopting an XGBoost regression model, to obtain a sorted enhanced feature dataset; the features in the sorted enhanced feature dataset are sequentially input into the XGBoost regression model for prediction, to obtain a current prediction value, and a target mean absolute percentage error between the current prediction value and a true value is calculated; if the target mean absolute percentage error is greater than or equal to a historical mean absolute percentage error, the features input into the XGBoost regression model are removed, until after a plurality of features are sequentially input into the XGBoost regression model, the target mean absolute percentage error calculated is greater than or equal to the historical mean absolute percentage error, the iterative prediction is ended, and a screened enhanced feature dataset is obtained. In this way, by excluding the features in the enhanced feature dataset that affect the learning of the effective wave height prediction network model, the effective wave height prediction network model can better learn the nonlinear relationship between the beneficial features and the prediction results, thereby improving the prediction accuracy of the effective wave height.

[0121] For the convenience of those skilled in the art, a set of best embodiments is provided below:

[0122] Sea waves are a complex wave phenomenon generated by wind in the ocean, and the significant wave height (SWH) is an important parameter for describing waves and is also the key prediction target of single-point sea wave prediction. The properties of sea waves can be characterized by multiple ocean parameters: significant wave height, average wave height, and average wave period, etc. Among them, the significant wave height (SWH) is the most important parameter of sea waves, and the accurate prediction of SWH is also the main research direction of ocean prediction.

[0123] Timely and accurate sea wave prediction not only plays an important role in marine activities such as fishing, exploration, power generation, and route planning, but also has many specific applications in the field of engineering. However, in recent years, data-based deep learning models such as long short-term memory networks (LSTM) and neural networks (RNN) have been widely used in single-point sea wave prediction, but these methods have been proven to have high lag and non-stationarity, and poor generalization ability. This is because the ocean environment itself has randomness and uncertainty, there are a large number of random wind fields on the sea surface, and the sea waves themselves have non-stationarity and high noise characteristics, which all have an impact on the accuracy of sea wave prediction.

[0124] To address these limitations, this embodiment proposes a sequence-to-sequence model (FE-BG-Seq2Seq) (i.e., a significant wave height prediction network model) based on feature selection and a brother guidance mechanism. This model first decomposes the original wave height signal using the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) algorithm, incorporating detailed components of each order of approximation into wave features. An XGBoost-based feature selection process then selects features that enhance model prediction. Finally, the sequence-to-sequence model, based on the brother guidance mechanism, predicts future wave heights. This embodiment's sequence-to-sequence model, based on the brother guidance mechanism, mitigates the error accumulation caused by autoregression in the sequence-to-sequence model and enhances the model's ability to mitigate prediction lag. Experiments comparing this embodiment's sequence-to-sequence model with the currently effective CEEMDAN-BiDLS™ model demonstrate that this embodiment improves the predictive power and accuracy of long-term series.

[0125] The technical solution of this embodiment specifically includes the following contents:

[0126] 1. Restate the problem.

[0127] The input data for single-point ocean wave forecast (i.e., ocean wave characteristic dataset) can be expressed as , the output data can be expressed as ,in is the number of wave features, is the predicted significant wave height, is the input time step, is the output time step. The sequence-to-sequence model of this embodiment is Time step and The wave features of the dimension are trained for multiple rounds to obtain a trained sequence-to-sequence model, and then the trained sequence-to-sequence model is used to predict The significant wave height data of the ocean waves for the time step.

[0128] 2. Introduction to CEEMDAN decomposition.

[0129] Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) is an adaptive noise-assisted improvement to empirical mode decomposition. CEEMDAN adaptively injects scale-modulated white noise during the decomposition process and utilizes an ensemble averaging strategy to effectively overcome the modal aliasing and residual noise interference inherent in traditional methods. This method iteratively separates the signal's intrinsic mode functions (IMFs) from high to low frequencies. Through noise-assisted decomposition and multi-scale ensemble operations, it ensures frequency-domain completeness and low reconstruction error for each IMF.

[0130] 3. Introduction to Seq2Seq.

[0131] Seq2Seq is a sequence-to-sequence learning method based on gated recurrent unit (GRU) to solve the problem of fixed-length output, which is widely used in natural language processing field, such as translation task and dialogue generation. By using multi-layer GRU structure or other encoder structure, the input sequence is mapped to a fixed-dimensional vector to capture the semantic information in the specified dimension, and then another GRU or other decoder structure is used to decode the input sequence and generate the specified length of output sequence by self-recurrence.

[0132] 4. Introduction of XGBoost.

[0133] XGBoost is a high-efficiency machine learning algorithm based on gradient boosting framework, which optimizes the objective function by iteratively adding weak learners (usually CART regression trees).

[0134] The objective function of XGBoost in the first iteration can be defined as:

[0135] (1);

[0136] where is the feature vector, represents the label, represents the model complexity, and the regular term can be further written as:

[0137] (2);

[0138] where is the number of leaf nodes, is the leaf node weight, represents the complexity control coefficient, which is the minimum gain threshold for leaf node splitting, represents the L2 regularization coefficient of leaf weight (output value).

[0139] When the split node is the left and right child nodes , the gain of the objective function in the first iteration can be represented as:

[0140] (3);

[0141] where is the gain of the decision tree, are the first and second derivatives of the left and right child nodes, is the L2 regularization factor, is the complexity control coefficient.

[0142] 5. Introduction of SHAP. ​​

[0143] SHAP is a model interpretation method based on game theory that quantifies the contribution of each feature to model predictions. In deep learning, explaining the decision logic is challenging due to the complexity and high nonlinearity of the model. SHAP provides a mathematically rigorous and intuitive way to understand the internal mechanisms of black-box models by randomly masking features and observing the changes in predictions. The following is the calculation formula for SHAP:

[0144] (4);

[0145] Among them, S represents each feature, and F represents the set of sample points of all features. It represents the model output after filling non-S features with background data. The Shapley diagram can be used to visually depict the contribution weight of each feature to the prediction result, including the positive and negative impact and the degree of impact, including the distribution of sample points. Represents the factorial operator.

[0146] 6. Feature gain.

[0147] To capture the non-stationary and multi-scale characteristics of significant wave height (SWH), this embodiment uses CEEMDAN decomposition to perform adaptive modal decomposition on the original wave height time series signal. After decomposing it into several intrinsic mode functions and residual terms, these high-frequency and low-frequency components are incorporated into the original feature set (i.e., the wave feature dataset) to construct an enhanced feature dataset.

[0148] 7. Feature selection.

[0149] The feature analysis of data is very important, and the selection of the number of features will affect the accuracy and speed of the model. Therefore, it is necessary to screen out features that can improve prediction accuracy and eliminate redundant features in advance. First, the enhanced feature data set is input, and the relative importance of each feature is calculated and sorted using the XGBoost regression model. However, the gain of the feature to the model still needs to be tested, so this embodiment designs an algorithm for the feature selection process.

[0150] Specifically, by selecting the mean absolute percentage error (MAPE) as the evaluation metric, the features in the enhanced feature dataset are sequentially input into XGBoost and the current MAPE (i.e., the target mean absolute percentage error) is calculated. If the current MAPE is better than the historical optimal MAPE, the feature is retained and the optimal MAPE is updated. Otherwise, the feature is removed and the patience counter is triggered. When k consecutive features fail to improve the MAPE, the iteration is terminated to prevent overfitting.

[0151] 8. Sequence-to-sequence neural network model based on brother guidance mechanism (i.e., effective wave height prediction network model).

[0152] Sequence-to-sequence network models are widely used in time series prediction, especially when the input and output sequences are of unequal length. The GRU neural network model can effectively transmit important information of the input sequence and suppress the propagation of invalid information through the gating mechanism. At the same time, it is simpler and more efficient than the LSTM neural network model because the GRU does not need to store cell information separately. This embodiment constructs a prediction model suitable for effective wave height (SWH). The encoder and decoder of the model of this embodiment both use GRU neural network units. The encoder of this embodiment will fully learn the information of the input sequence and pass the learning information to the decoder. Unlike conventional sequence-to-sequence networks, the brother guidance mechanism adopted in this embodiment enables the encoder and decoder to make predictions together. Figure 2 .

[0153] Specifically, given the input sequence ,in is the input sequence length and initializes the GRU hidden layer state The encoder GRU unit processes the input sequence sequentially and updates the hidden state according to formulas (5) to (8):

[0154] (5);

[0155] (6);

[0156] (7);

[0157] (8);

[0158] in, Indicates the The encoder updates the gate at time , represents the activation function, represents the weight matrix of the update gate, Indicates the The encoder hidden state at time t, represents the encoder input features, represents the bias vector of the update gate, Indicates the encoder reset gate, represents the weight matrix of the reset gate, represents the bias vector for resetting the gate, represents the encoder candidate hidden state, represents the hyperbolic tangent function, The weight matrix representing the candidate hidden states of the encoder, The bias vector representing the candidate hidden state of the encoder, represents the Hadamard product.

[0159] In the decoding stage, the future time series is predicted step by step using an autoregressive mechanism. To preserve the sequence continuity of the encoder and the decoder, the last hidden state of the encoder is taken as the initial hidden state of the decoder , which ensures the continuity of the hidden layer, i.e.

[0160] (9);

[0161] At the same time, the last hidden state of the encoder is taken as the semantic encoding obtained after the encoder learns the timing information of the input sequence , and it is normalized through a Softmax layer to obtain the weighted parameters with global information of the input sequence :

[0162] (10);

[0163] (11);

[0164] The weighted parameters are multiplied with the last value in the input sequence through Hadamard product operation, and through a linear layer, the first input of the decoder is obtained, i.e.

[0165] (12);

[0166] After that, the decoder updates the encoder hidden state according to the following formula and gradually obtains the output significant wave height , where , is the length of the sequence to be predicted (i.e. the output time step):

[0167] (13);

[0168] (14);

[0169] (15);

[0170] (16);

[0171] (17);

[0172] (18);

[0173] wherein denotes the input of the decoder at the moment, denotes the weight matrix, represents the weighting parameter, represents the Hadamard product, Represents the last time series feature in the filtered enhanced feature dataset, represents the bias vector, Indicates the The input of the time decoder, Indicates the The decoder updates the gate at time , Indicates the The decoder hidden state at time t, Indicates the decoder reset gate, represents the decoder candidate hidden state, The weight matrix representing the candidate hidden states of the decoder, The bias vector representing the candidate hidden state of the decoder, Indicates the The effective wave height prediction result at the moment.

[0174] In the model of this embodiment, the encoder is considered the decoder's elder brother, providing "guidance" to the decoder at each prediction step. This is because the encoder has fully learned the global information of the input sequence, and guiding the decoder will improve the performance of the final model. Compared with traditional autoregressive models, the decoder based on the elder brother guidance mechanism will fully utilize the global information of the input sequence provided by the encoder for each autoregressive input, which will alleviate the accumulation of prediction errors caused by autoregression. At the same time, in this embodiment, the use of the elder brother guidance mechanism ensures that the input data received by the single GRU does not only contain the information transmitted by the hidden layer, which can enrich the GRU's input knowledge and thus improve the predictive ability of the model of this embodiment.

[0175] In order to better illustrate the effect of the technical solution of this embodiment, this embodiment conducted the following experiments:

[0176] This embodiment uses four indicators: mean absolute error (MAE), mean absolute percentage error (MAPE), root mean square error (RMSE) and correlation coefficient ( ) to measure the model training effect. MAE and RMSE measure the absolute error between the predicted value and the true value, while MAPE measures the relative error. Relatively speaking, MAE and MAPE are not easily affected by extreme values, while RMSE is easy to amplify the error after squared and is more sensitive to outlier data. The closer the value is to 1, the more reliable the prediction is. The following are the calculation formulas for the four indicators:

[0177] (19);

[0178] (20);

[0179] (21);

[0180] (22);

[0181] where, is the true value, is the predicted value, is the mean of the true values, is the mean of the predicted values.

[0182] (1) Data preprocessing.

[0183] Ocean buoy is one of the most effective methods to obtain sea wave data in a small range of sea area. The study of single-point sea wave prediction usually uses buoy data. By deploying a large number of ocean buoys around the world, a large amount of ocean data can be provided stably for a long time. In this embodiment, the data of four stations 41008, 41010, 46080 and 46013 are selected as examples, and the time resolution reaches 1 hour. In this embodiment, the sliding method is used to cut the long time series data into the required samples, and the training is carried out according to the proportion of 0.8, 0.1 and 0.1 of the training set, the validation set and the test set, and the training is carried out in batches of samples during the process.

[0184] Because the buoy works in an unstable marine environment, data loss often occurs, which affects the training accuracy of the model. Therefore, a large number of continuous missing values are deleted, and a single missing value is processed using a linear interpolation method. This filling method is mainly based on the observation of past and future data:

[0185] (23);

[0186] where, is the filled value, and are the first non-missing values before and after the missing value, , and represent the time positions of the missing values, and . This method effectively makes up for the defects of the data in space and time.

[0187] By normalizing the data, the problems caused by dimension and numerical difference are eliminated, the gradient descent process of the algorithm is optimized, the convergence speed is improved, and the model robustness is enhanced. The normalization method is as follows:

[0188] (24);

[0189] Here is the normalized feature data, is the maximum value in the feature, is the minimum value in the feature, by this processing the feature is mapped to the 0-1 interval, improve the convergence speed.

[0190] (2) Feature selection evaluation.

[0191] Table 1 is the performance of two models for 24 hours

[0192]

[0193] In this experiment, the embodiment adopts XGBoost to screen 19 features, and removes features that do not increase the prediction accuracy of the model. By using the XGBoost regressor to sort the feature importance, the importance of each modal component can be clearly seen. The experimental results of the embodiment show that the modal component GST has the highest importance, WSPD has the second importance, and APD follows.

[0194] Figure 3 The average absolute error (MAE) and the average absolute percentage error (MAPE) are plotted in the order of feature importance. The MAE value decreases significantly as the number of features increases at the initial time. When the number of input features reaches 11, the MAE and MAPE of the model reach the extreme point, and the error further increases when the number of features is too large, showing a tendency of overfitting. Therefore, the experiment selects these 11 features for model training.

[0195] Due to the limitation of the length, only the comparison of the model before and after adding feature selection for 24 hours is selected here. From Table 1, it can be seen that the prediction accuracy of the model after adding feature engineering is greatly improved compared to before adding. Among them, Seq2Seq is before adding, XB-Seq2Seq is after adding feature engineering. The main reason is that by separating different modalities of features and selecting modalities that have gain for model prediction, the stacking effect is clearly depicted. Each feature modality alone affects the model, reducing the ambiguity of full feature prediction.

[0196] (3) Performance evaluation.

[0197] The training effect of a deep learning model is crucial to parameters. The main parameters of the innovative FE-BG-Seq2Seq model include seqlength, hiddensize, batchsize, nhead, numlayers and learningrate. In this embodiment, a medium batchsize of 256 and a relatively small num_layer of 1 are uniformly set to balance the model accuracy and the model training speed. Through a series of parameter experiments, the parameters seqlength=128, hiddensize=72, nhead=8 and learningrate=0.002 are determined.

[0198] Table 2 is the performance comparison results of each model for 24 hours

[0199]

[0200] The comparative models selected in this experiment include Seq2Seq, XB-Seq2Seq and CEEMDAN-BiDLSTM models, which are all models in the prior art and will not be described in detail in this embodiment. The XB-Seq2Seq model combines the feature selection methods of XGBoost and Seq2Seq, and the CEEMDAN-BiDLSTM model combines the CEEMDAN decomposition algorithm and the BiDLSTM model. The model has been proven to have good performance in the nonlinear and non-stationary sea wave time series prediction in recent research.

[0201] Table 3 is the performance comparison results of each model for 48 hours

[0202]

[0203] Tables 2 and 3 list the prediction results of the four models for 24 hours and 48 hours of sea wave significant wave height, using long time series data (such as 41008 site data from 2005 to 2023). Vertical comparison shows that the prediction effect of the four models for 24 hours is better than that for 48 hours, proving that the four models have good prediction ability for short-term prediction of 24 hours. Horizontal comparison shows that the model (FE-BG-Seq2Seq) of this embodiment achieves the best effect, followed by the CEEMDAN-BiDLSTM model. Taking MAE as an example, the error of the model of this embodiment is reduced by 15%, 18%, 17% and 21% compared with the CEEMDAN-BiDLSTM model for 24 hours prediction at each site, and the error is reduced by 22%, 24%, 22% and 25% compared with the CEEMDAN-BiDLSTM model for 48 hours prediction at each site. This shows that the model of this embodiment has a larger improvement compared with the CEEMDAN-BiDLSTM model after adding the elder brother guidance mechanism.

[0204] (4) Model verification.

[0205] Feature analysis is very important to understand the internal mechanism of deep learning model. Therefore, in the experiment, the SHAP value is used to analyze the network model based on the brother guidance mechanism constructed in the embodiment, so as to quantify and visualize the influence of each sea wave feature on the prediction result, which can further verify the correctness of the feature selection process and the correctness of the model for different feature weight allocation.

[0206] Since SHAP can only identify the two dimensions of batchsize and hiddensize, the embodiment aggregates the dimension of seqlength, and takes the average on the time step. The embodiment takes the training set as the background sample and the test set as the test instance, and draws a representative SHAP contribution chart (bee chart).

[0207] Figure 4 The SHAP value bee chart drawn by the model of the embodiment, wherein each point represents a sample point, and WVHT, GST, WSPD, APD, IMFS_7, MFS_6 and IMFS_8 represent various features. The abscissa represents the SHAP value influence value of the sample point corresponding to the feature, the size of the abscissa represents the SHAP influence amount, and the positive and negative represent the influence direction. Therefore, the more densely the sample points are distributed on the right side of the coordinate axis, the greater the positive contribution to the model output value. The SHAP chart from top to bottom shows that the feature importance decreases in turn, which is consistent with the feature importance ranking in the feature selection process, and the sample points corresponding to GST, WSPD and APD are densely distributed on the positive half axis, indicating that these features have a significant positive correlation with the significant wave height, which is consistent with the results obtained in the feature selection evaluation. These further verify the correctness of the model of the embodiment.

[0208] Compared with the prior art, the method of the embodiment has the following advantages:

[0209] The past models such as ARIMA, LSTM and Seq2Seq are often based on short sequence data for short-term prediction, and have defects such as lag and non-stationarity, which are difficult to capture the global multi-dimensional relationship of data. The FE-BG-Seq2Seq model proposed in the embodiment combines feature engineering, brother guidance mechanism and sequence-to-sequence main model, which not only effectively transmits the information of the input sequence through the sequence-to-sequence network, but also makes the encoder and decoder predict in parallel through the brother guidance mechanism, reducing error accumulation.

[0210] In the experiment of the embodiment, the feature selection process is first proved to significantly improve MAE, MAPE, R2 and RMSE and other indicators by comparing the best parameters Seq2Seq and XB-Seq2Seq in the transverse direction, and then the long-term prediction ability is proved to be enhanced and the prediction accuracy under extreme SWH samples can be maintained by comparing the CEEMDAN-BiDLSTM and the model of the embodiment.

[0211] In addition, the model interpretability is enhanced by using SHAP in the embodiment, SHAP well verifies the correctness of the previous feature selection process and the features that have gain on the model prediction, and the deep learning black box model FE-BG-Seq2Seq of the embodiment model is analyzed from the inside.

[0212] The embodiments of the application are described in detail above in combination with the drawings, but the application is not limited to the above embodiments, and various changes can be made within the knowledge range of ordinary skilled in the art without departing from the purpose of the application.

Claims

1. An effective wave height prediction method based on feature engineering and elder guiding mechanism, characterized in that, The method comprises: obtaining a sea wave feature data set containing an original wave height time series signal; modal decomposition is performed on the original wave height time series signal to obtain a plurality of intrinsic mode functions and a plurality of residual terms; the plurality of intrinsic mode functions and the plurality of residual terms are added to the sea wave feature data set to construct an enhanced feature data set; feature screening is performed on the features in the enhanced feature data set to obtain a screened enhanced feature data set; an effective wave height prediction network model is constructed based on a gated recurrent unit and using an elder guidance mechanism, wherein a plurality of cascaded gated recurrent units are used to construct an encoder and a decoder; an effective wave height prediction network model is constructed according to the encoder and the decoder, wherein the encoder is used as the elder of the decoder to guide the decoder to make a prediction; the screened enhanced feature data set is input into the effective wave height prediction network model for prediction to obtain an effective wave height prediction result, specifically as follows: the screened enhanced feature data set is processed by a plurality of gated recurrent units in the encoder to obtain the last hidden state of the encoder; the last hidden state of the encoder is used as the initial hidden state of the decoder, and the last hidden state of the encoder is used as the semantic encoding learned by the encoder; the semantic encoding is normalized to obtain a weighting parameter; the weighting parameter and the last time series feature in the screened enhanced feature data set are subjected to Hadamard product, and after linear layer operation, the first input of the decoder is obtained; the first input of the decoder and the initial hidden state of the decoder are input into the decoder for prediction to obtain an effective wave height prediction result. 2.The method of claim 1, wherein, The normalization of the semantic encoding to obtain a weighting parameter comprises: ； ； wherein, denotes the semantic encoding, denotes the last hidden state of the encoder, denotes the weighting parameter, denotes the normalization layer. 3.The effective wave height prediction method based on feature engineering and elder brother guidance mechanism according to claim 1, characterized in that, The processing of the screened enhanced feature data set by the plurality of gated recurrent units in the encoder comprises: ； ； ； ； in, Indicates the The encoder updates the gate at time , represents the activation function, represents the weight matrix of the update gate, Indicates the The encoder hidden state at time t, represents the encoder input features, represents the bias vector of the update gate, Indicates the encoder reset gate, represents the weight matrix of the reset gate, represents the bias vector for resetting the gate, represents the encoder candidate hidden state, represents the hyperbolic tangent function, The weight matrix representing the candidate hidden states of the encoder, The bias vector representing the candidate hidden state of the encoder, represents the Hadamard product. 4.The method of claim 1, wherein, The Hadamard product of the weighting parameter and the last time series feature in the screened enhanced feature data set, and the linear layer operation to obtain the first input of the decoder, comprises: ； wherein, denotes a first input of the decoder, denotes a weight matrix, denotes a weighting parameter, denotes a Hadamard product, denotes the last temporal feature of the filtered enhanced feature dataset, denotes a bias vector, denotes a linear layer operation.

5. The effective wave height prediction method based on feature engineering and brother guidance mechanism according to claim 1 is characterized in that: The input of the first input of the decoder and the initial hidden state of the decoder into the decoder for prediction to obtain an effective wave height prediction result comprises: ； ； ； ； ； ； wherein, represents the input of the decoder at the time step, represents a weight matrix, represents a weight parameter, represents a Hadamard product, represents the last time-series feature in the filtered enhanced feature dataset, represents a bias vector, represents the input of the decoder at the time step, represents the input of the decoder at the time step, represents an activation function, represents a weight matrix of the update gate, represents the decoder hidden state at the time step, represents a bias vector of the update gate, represents the decoder reset gate, represents a weight matrix of the reset gate, represents a bias vector of the reset gate, represents the decoder candidate hidden state, represents a hyperbolic tangent function, represents a weight matrix of the decoder candidate hidden state, represents a bias vector of the decoder candidate hidden state, represents the significant wave height prediction result at the time step. 6.The method of claim 1, wherein, The modal decomposition of the original wave height time series signal to obtain a plurality of intrinsic mode functions and a plurality of residual terms comprises: Adaptive noise complete ensemble empirical mode decomposition is used to perform modal decomposition on the original wave height time series signal to obtain a plurality of intrinsic mode functions and a plurality of residual terms.

7. The method of claim 1, wherein the method is characterized by, The feature screening of the features in the enhanced feature data set to obtain a screened enhanced feature data set comprises: An XGBoost regression model is used to sort the features in the enhanced feature data set to obtain a sorted enhanced feature data set; the features in the sorted enhanced feature data set are sequentially input into the XGBoost regression model for prediction to obtain a current prediction value, and the target mean absolute percentage error between the current prediction value and the true value is calculated; If the target mean absolute percentage error is smaller than the historical mean absolute percentage error, the feature input into the XGBoost regression model is retained until after a plurality of features are continuously input into the XGBoost regression model, the calculated target mean absolute percentage error is equal to or greater than the historical mean absolute percentage error, the iterative prediction is ended, and a screened enhanced feature dataset is obtained. 8.The method of claim 1, wherein, The feature screening on the features in the enhanced feature dataset to obtain a screened enhanced feature dataset comprises: The features in the enhanced feature dataset are sorted by using an XGBoost regression model to obtain a sorted enhanced feature dataset; The features in the sorted enhanced feature dataset are sequentially input into the XGBoost regression model for prediction to obtain a current prediction value, and a target mean absolute percentage error between the current prediction value and an actual value is calculated; If the target mean absolute percentage error is greater than or equal to the historical mean absolute percentage error, the feature input into the XGBoost regression model is removed until after a plurality of features are continuously input into the XGBoost regression model, the calculated target mean absolute percentage error is greater than or equal to the historical mean absolute percentage error, the iterative prediction is ended, and a screened enhanced feature dataset is obtained.

Images