Effective wave height prediction method based on feature engineering and brob length guidance mechanism

Abstract

The invention discloses a significant wave height prediction method based on feature engineering and a brother length guidance mechanism. The method comprises the following steps: acquiring a sea wave feature data set 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 into the sea wave feature data set to construct an enhanced feature data set; performing feature screening on features in the enhanced feature data set to obtain a screened enhanced feature data set; constructing a significant wave height prediction network model by adopting a broker length guidance mechanism based on a gating circulation unit; and inputting the screened enhanced feature data set into the significant wave height prediction network model for prediction to obtain a significant wave height prediction result. The prediction accuracy of the significant wave height can be improved.

Description

Technical Field

[0001] This application relates to the technical field of sea wave forecasting, and particularly to an effective wave height prediction method based on feature engineering and an elder brother guidance mechanism. Background Art

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

[0003] Some people combine the XgBoost and LSTM models to avoid the problems of model overfitting and data missing; the combination of signal decomposition methods and deep learning network models has been proven to effectively improve the correlation and reduce errors in more and more experiments. Some people combine the wavelet decomposition algorithm EMD with the LSTM model and prove its good superiority in predicting nonlinear and non-steady waves. A random decomposition wave method of DSD is also combined with LSTM, and better results than wavelet decomposition are obtained in multi-variable input prediction. However, existing deep learning methods still rarely pay attention to the problems of low computational efficiency caused by long-term forecasting and attenuation of long-sequence information. In recent years, data-based deep learning models such as LSTM and RNN have been widely used in single-point sea wave forecasting, but these methods have been proven to have high hysteresis and non-steadiness and poor generalization ability.

[0004] In summary, the existing related methods have relatively low prediction accuracy for the significant wave height. Summary of the Invention

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

[0006] An embodiment of this application provides an effective wave height prediction method based on feature engineering and an elder brother guidance mechanism. The method includes: Obtain a sea wave feature data set containing the original wave height time series signal; 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; 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; Perform feature screening on the features in the enhanced feature data set to obtain a screened enhanced feature data set; Based on the gated recurrent unit, construct an effective wave height prediction network model using the elder brother guidance mechanism; Input the filtered enhanced feature dataset into the significant wave height prediction network model for prediction to obtain the significant wave height prediction result.

[0007] In some embodiments, the significant wave height prediction network model constructed by adopting the elder brother guidance mechanism based on the gated recurrent unit includes: Adopt a plurality of cascaded gated recurrent units to construct an encoder and a decoder; Construct a significant wave height prediction network model according to the encoder and the decoder, wherein the encoder is used as the elder brother of the decoder to guide the decoder to make predictions.

[0008] In some embodiments, the inputting the filtered enhanced feature dataset into the significant wave height prediction network model for prediction to obtain the significant wave height prediction result includes: Process the filtered enhanced feature dataset through the plurality of gated recurrent units in the encoder to obtain the last hidden state of the encoder; Use the last hidden state of the encoder as the initial hidden state of the decoder, and use the last hidden state of the encoder as the semantic encoding learned by the encoder; Normalize the semantic encoding to obtain a weighted parameter; Perform a Hadamard product on the weighted parameter and the last temporal feature in the filtered enhanced feature dataset, and obtain the first input of the decoder through a linear layer operation; Input the first input of the decoder and the initial hidden state of the decoder into the decoder for prediction to obtain the significant wave height prediction result.

[0009] In some embodiments, the normalizing the semantic encoding to obtain a weighted parameter includes: ; ; Wherein, represents the semantic encoding, represents the last hidden state of the encoder, represents the weighted parameter, represents the normalization layer.

[0010] In some embodiments, the processing the filtered enhanced feature dataset through the plurality of gated recurrent units in the encoder includes: ; ; ; ; Among them, represents the encoder update gate at the th moment, represents the activation function, represents the weight matrix of the update gate, represents the encoder hidden state at the th moment, represents the encoder input feature, represents the bias vector of the update gate, represents the encoder reset gate, represents the weight matrix of the reset gate, represents the bias vector of the reset gate, represents the encoder candidate hidden state, represents the hyperbolic tangent function, represents the weight matrix of the encoder candidate hidden state, represents the bias vector of the encoder candidate hidden state, represents the Hadamard product.

[0011] In some embodiments, the operation of obtaining the first input of the decoder by performing a Hadamard product on the weighted parameter and the last temporal feature in the filtered enhanced feature dataset and then performing a linear layer operation includes: ; Among them, represents the first input of the decoder, represents the weight matrix, represents the weighted parameter, represents the Hadamard product, represents the last temporal feature in the filtered enhanced feature dataset, represents the bias vector, represents the linear layer operation.

[0012] In some embodiments, the operation of inputting 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 includes: ; ; ; ; ; ; Among them, represents the input of the decoder at the moment, represents the weight matrix, represents the weighting parameter, represents the Hadamard product, represents the last temporal feature in the filtered enhanced feature dataset, represents the bias vector, represents the input of the decoder at the represents the decoder update gate at the moment, represents the activation function, represents the weight matrix of the decoder update gate, represents the hidden state of the decoder at the moment, represents the bias vector of the update gate, represents the decoder reset gate, represents the weight matrix of the decoder reset gate, represents the bias vector of the reset gate, represents the decoder candidate hidden state, represents the hyperbolic tangent function, represents the weight matrix of the decoder candidate hidden state, represents the bias vector of the decoder candidate hidden state, represents the effective wave height prediction result at the moment.

[0013] In some embodiments, 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 includes: Performing modal decomposition on the original wave height time series signal by using ensemble empirical mode decomposition with adaptive noise to obtain a plurality of intrinsic mode functions and a plurality of residual terms.

[0014] In some embodiments, the feature screening of the features in the enhanced feature dataset to obtain a filtered enhanced feature dataset includes: Using the XGBoost regression model to rank the features in the enhanced feature dataset to obtain a ranked enhanced feature dataset; Sequentially inputting the features in the ranked enhanced feature dataset into the XGBoost regression model for prediction to obtain a current predicted value, and calculating the target mean absolute percentage error between the current predicted value and the true value; If the target mean absolute percentage error is smaller than the historical mean absolute percentage error, retain the features input into the XGBoost regression model until after multiple features are continuously input into the XGBoost regression model, and the calculated target mean absolute percentage error is equal to or greater than the historical mean absolute percentage error. End the iterative prediction to obtain the filtered enhanced feature dataset.

[0015] In some embodiments, the feature screening of the features in the enhanced feature dataset to obtain the filtered enhanced feature dataset includes: Use the XGBoost regression model to rank the features in the enhanced feature dataset to obtain the ranked enhanced feature dataset; Input the features in the ranked enhanced feature dataset into the XGBoost regression model in sequence for prediction to obtain the current predicted value, and calculate the target mean absolute percentage error between the current predicted value and the true value; If the target mean absolute percentage error is greater than or equal to the historical mean absolute percentage error, remove the features input into the XGBoost regression model until after multiple features are continuously input into the XGBoost regression model, and the calculated target mean absolute percentage error is greater than or equal to the historical mean absolute percentage error. End the iterative prediction to obtain the filtered enhanced feature dataset.

[0016] Compared with the prior art, the present application has the following beneficial effects: This method obtains a sea wave feature dataset containing the original wave height time series signal; performs modal decomposition on the original wave height time series signal to obtain a plurality of intrinsic mode functions and a plurality of residual terms; adds the plurality of intrinsic mode functions and the plurality of residual terms to the sea wave feature dataset to construct an enhanced feature dataset; performs feature screening on the features in the enhanced feature dataset to obtain the filtered enhanced feature dataset; constructs an effective wave height prediction network model based on a gated recurrent unit using the elder brother guidance mechanism; inputs the filtered enhanced feature dataset into the effective wave height prediction network model for prediction to obtain the effective wave height prediction result. In this way, through modal decomposition and feature screening, the model can better learn the non-linear relationship between beneficial features and prediction results, and then constructs an effective wave height prediction network model through the elder brother guidance mechanism, alleviates the error accumulation caused by the autoregression of the model, enhances the ability of the model to combat prediction lag, and thus improves the prediction accuracy of the effective wave height. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The above and / or additional aspects and advantages of the present application will become apparent and be readily understood from the following description of the embodiments in conjunction with the accompanying drawings, where: Figure 1It is a schematic flowchart of an embodiment of an effective wave height prediction method based on feature engineering and an elder brother guidance mechanism provided by the present application; Figure 2 It is a schematic structural diagram of a sequence-to-sequence neural network model based on an elder brother guidance mechanism in the best embodiment of an effective wave height prediction method based on feature engineering and an elder brother guidance mechanism provided by the present application; Figure 3 It is a schematic diagram showing the variation of MAE and MAPE with the number of features K in the best embodiment of an effective wave height prediction method based on feature engineering and an elder brother guidance mechanism provided by the present application; Figure 4 It is a schematic diagram of a SHAP value swarm plot in the best embodiment of an effective wave height prediction method based on feature engineering and an elder brother guidance mechanism provided by the present application. Detailed implementation manners

[0018] The embodiments of the present application will be described in detail below. Examples of the embodiments are shown in the drawings, where the same or similar reference numerals denote the same or similar elements or elements with the same or similar functions throughout. The embodiments described below with reference to the drawings are exemplary only for explaining the present application and should not be construed as limiting the present application.

[0019] In the description of the present application, if the first, second, etc. are described only for the purpose of distinguishing technical features, they should not be understood as indicating or implying relative importance or implicitly indicating the quantity of the indicated technical features or implicitly indicating the sequence relationship of the indicated technical features.

[0020] In the description of the present application, it should be understood that for the orientation description, such as up, down, etc., the orientation or positional relationship indicated is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present application and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as limiting the present application.

[0021] In the description of the present application, it should be noted that unless otherwise clearly defined, words such as setting, installing, connecting, etc. should be understood in a broad sense, and those skilled in the art can reasonably determine the specific meanings of the above words in the present application in combination with the specific content of the technical solution.

[0022] Since existing deep learning methods still rarely pay attention to the problems of low computational efficiency caused by long-term forecasting and attenuation of long-sequence information. In recent years, data-based deep learning models such as LSTM and RNN have been widely used in single-point sea wave forecasting, but these methods have been proven to have high hysteresis and non-steadiness, and poor generalization ability.

[0023] To solve the problem that the prediction accuracy of the existing related methods for significant wave height is relatively low, this application proposes a significant wave height prediction method based on feature engineering and an elder brother guidance mechanism.

[0024] Referring to Figure 1 , the flowchart of the significant wave height prediction method based on feature engineering and an elder brother guidance mechanism provided by the embodiments of this application. The significant wave height prediction method based on feature engineering and an elder brother guidance mechanism is applied to an electronic device, and the electronic device can be a server, a mobile terminal, etc. As Figure 1 shown, the significant wave height prediction method based on feature engineering and an elder brother guidance mechanism may include the following steps: Step S100, obtain a sea wave feature dataset containing an original wave height time series signal; Step S200, 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; Step S300, add the plurality of intrinsic mode functions and the plurality of residual terms to the sea wave feature dataset to construct an enhanced feature dataset; Step S400, perform feature screening on the features in the enhanced feature dataset to obtain a screened enhanced feature dataset; Step S500, based on a gated recurrent unit, construct a significant wave height prediction network model using an elder brother guidance mechanism; Step S600, input the screened enhanced feature dataset into the significant wave height prediction network model for prediction to obtain a significant wave height prediction result.

[0025] In this embodiment, by obtaining a sea wave feature dataset containing an original wave height time series signal; performing modal decomposition on the original wave height time series 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; based on a gated recurrent unit, constructing a significant wave height prediction network model using an elder brother guidance mechanism; inputting the screened enhanced feature dataset into the significant wave height prediction network model for prediction to obtain a significant wave height prediction result. In this way, through modal decomposition and feature screening, the model can better learn the non-linear relationship between beneficial features and prediction results, and then construct a significant wave height prediction network model through an elder brother guidance mechanism, alleviating the error accumulation caused by the autoregression of the model and enhancing the ability of the model to combat prediction lag, thereby improving the prediction accuracy of significant wave height.

[0026] The above-mentioned obtaining of the sea wave feature dataset containing an original wave height time series signal may be by obtaining single-point sea wave forecast data, and the single-point sea wave forecast data may include data such as an original wave height time series signal, and this embodiment does not make specific limitations.

[0027] The above-mentioned modal decomposition of the original wave height time series signal can be performed by using complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) on the original wave height time series signal, or by using other modal decomposition techniques well-known to those skilled in the art to perform modal decomposition on the original wave height time series signal. This embodiment does not make specific limitations.

[0028] In some embodiments, based on the gated recurrent unit, an effective wave height prediction network model is constructed by adopting a big brother guidance mechanism, including: Using multiple cascaded gated recurrent units to construct an encoder and a decoder; According to the encoder and the decoder, an effective wave height prediction network model is constructed, where the encoder is used as the big brother of the decoder to guide the decoder to make predictions.

[0029] In this embodiment, by using multiple cascaded gated recurrent units, an encoder and a decoder are constructed; according to the encoder and the decoder, an effective wave height prediction network model is constructed, where the encoder is used as the big brother of the decoder to guide the decoder to make predictions. In this way, the encoder is used as the big brother of the decoder, and the decoder is guided by the encoder to make predictions, that is, the encoder and the decoder perform parallel predictions through the big brother guidance mechanism, which can reduce error accumulation and thus improve the prediction accuracy of the effective wave height.

[0030] 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: Processing the filtered enhanced feature dataset through multiple gated recurrent units in the encoder to obtain the last hidden state of the encoder; Taking the last hidden state of the encoder as the initial hidden state of the decoder, and taking the last hidden state of the encoder as the semantic encoding learned by the encoder; Normalizing the semantic encoding to obtain a weighted parameter; Performing a Hadamard product on the weighted parameter and the last time series feature in the filtered enhanced feature dataset, and after linear layer operations, obtaining the first input of the decoder; Inputting 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.

[0031] In this embodiment, the filtered enhanced feature dataset is processed by multiple 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 temporal feature in the filtered enhanced feature dataset are subjected to a Hadamard product, and after a 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. In this way, at the decoding stage of the decoder, the encoder receives the input data at the end of the encoder again, and gives the prediction output (that is, using the last hidden state of the encoder as the semantic encoding learned by the encoder) to the decoder as input to alleviate the error accumulation caused by autoregressive prediction and better combat the lag of prediction.

[0032] In some embodiments, normalizing the semantic encoding to obtain a weighting parameter includes: ; ; where, represents the semantic encoding, represents the last hidden state of the encoder, represents the weighting parameter, represents the normalization layer.

[0033] In some embodiments, processing the filtered enhanced feature dataset by multiple gated recurrent units in the encoder includes: ; ; ; ; where, represents the encoder update gate at the -th moment, represents the activation function, represents the weight matrix of the update gate, represents the encoder hidden state at the -th moment, represents the encoder input feature, represents the bias vector of the update gate, represents the encoder reset gate, represents the weight matrix of the reset gate, represents the bias vector of the reset gate, Denotes the encoder candidate hidden state, Denotes the hyperbolic tangent function, Denotes the weight matrix of the encoder candidate hidden state, Denotes the bias vector of the encoder candidate hidden state, Denotes the Hadamard product.

[0034] In some embodiments, the weighted parameter is subjected to a Hadamard product with the last temporal feature in the filtered enhanced feature dataset, and after a linear layer operation, the first input to the decoder is obtained, including: ; Wherein, Denotes the first input to the decoder, Denotes the weight matrix, Denotes the weighted parameter, Denotes the Hadamard product, Denotes the last temporal feature in the filtered enhanced feature dataset, Denotes the bias vector, Denotes the linear layer operation.

[0035] In some embodiments, the first input to 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, including: ; ; ; ; ; ; Wherein, Denotes the input to the decoder at the th moment, Denotes the weight matrix, Denotes the weighted parameter, Denotes the Hadamard product, Denotes the last temporal feature in the filtered enhanced feature dataset, Denotes the bias vector, Denotes the input to the decoder at the th moment, Denotes the decoder update gate at the th moment, Denotes the activation function, Denotes the weight matrix of the update gate, Denotes the hidden state of the decoder at the th moment, Represents the bias vector for updating the gate, Represents the decoder reset gate, Represents the weight matrix of the reset gate, Represents the bias vector of the reset gate, Represents the decoder candidate hidden state, Represents the hyperbolic tangent function, Represents the weight matrix of the decoder candidate hidden state, Represents the bias vector of the decoder candidate hidden state, Represents the Effective significant wave height prediction result at time

[0036] 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: The original wave height time series signal is subjected to modal decomposition by using the complete ensemble empirical mode decomposition with adaptive noise to obtain a plurality of intrinsic mode functions and a plurality of residual terms.

[0037] In this embodiment, by subjecting the original wave height time series signal to modal decomposition by using the complete ensemble empirical mode decomposition with adaptive noise, the intrinsic mode functions (i.e., IMFs) that sequentially separate signals from high frequency to low frequency can be obtained, that is, the detail components are obtained. Through noise-assisted decomposition and multi-scale integration operations, the frequency domain completeness and low reconstruction error characteristics of each IMF are ensured. Later, these detail components are added to the original feature dataset, which can enhance the feature dataset and lay a good data foundation for accurately predicting the effective significant wave height in the later stage.

[0038] In some embodiments, feature screening is performed on the features in the enhanced feature dataset to obtain a screened enhanced feature dataset, including: The XGBoost regression model is used to rank the features in the enhanced feature dataset to obtain a ranked enhanced feature dataset; The features in the ranked enhanced feature dataset are sequentially input into the XGBoost regression model for prediction to obtain the current predicted value, and the target mean absolute percentage error between the current predicted value and the true value is calculated; 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 the target mean absolute percentage error calculated after continuously inputting multiple features into the XGBoost regression model is equal to or greater than the historical mean absolute percentage error, and the iterative prediction is ended to obtain a screened enhanced feature dataset.

[0039] In this embodiment, the features in the enhanced feature dataset are sorted by using the XGBoost regression model to obtain the sorted enhanced feature dataset; the features in the sorted enhanced feature dataset are sequentially input into the XGBoost regression model for prediction to obtain the current predicted value, and the target mean absolute percentage error between the current predicted value and the true value is calculated; 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 after multiple features are continuously input into the XGBoost regression model, and the calculated target mean absolute percentage error is equal to or greater than the historical mean absolute percentage error, then the iterative prediction ends, and the filtered enhanced feature dataset is obtained. In this way, by using the XGBoost regression model to select the modes beneficial to prediction and include them in the feature subset (i.e., obtaining the filtered enhanced feature dataset), the effective wave height prediction network model can better learn the non-linear relationship between the beneficial features and the prediction results, thereby improving the prediction accuracy of the effective wave height.

[0040] In some embodiments, feature screening is performed on the features in the enhanced feature dataset to obtain the filtered enhanced feature dataset, including: The features in the enhanced feature dataset are sorted by using the XGBoost regression model to obtain the sorted enhanced feature dataset; The features in the sorted enhanced feature dataset are sequentially input into the XGBoost regression model for prediction to obtain the current predicted value, and the target mean absolute percentage error between the current predicted value and the true value is calculated; If the target mean absolute percentage error is greater than or equal to the historical mean absolute percentage error, the features input into the XGBoost regression model are removed until after multiple features are continuously input into the XGBoost regression model, and the calculated target mean absolute percentage error is greater than or equal to the historical mean absolute percentage error, then the iterative prediction ends, and the filtered enhanced feature dataset is obtained.

[0041] In this embodiment, the features in the enhanced feature dataset are sorted by using the XGBoost regression model to obtain the sorted enhanced feature dataset; the features in the sorted enhanced feature dataset are sequentially input into the XGBoost regression model for prediction to obtain the current predicted value, and the target mean absolute percentage error between the current predicted value and the true value is calculated; if the target mean absolute percentage error is greater than or equal to the historical mean absolute percentage error, the features input into the XGBoost regression model are removed until the target mean absolute percentage error calculated after continuously inputting multiple features into the XGBoost regression model is greater than or equal to the historical mean absolute percentage error, and the iterative prediction is ended to obtain the filtered enhanced feature dataset. In this way, by excluding the features in the enhanced feature dataset that affect the learning of the significant wave height prediction network model, the significant wave height prediction network model can better learn the non-linear relationship between the beneficial features and the prediction results, thereby improving the prediction accuracy of the significant wave height.

[0042] For the convenience of those skilled in the art to understand, the following provides a set of optimal embodiments: Ocean waves are a complex wave phenomenon generated by wind in the ocean. Among them, the significant wave height (SWH) is an important parameter for describing waves and also the key prediction target for single-point ocean wave forecasting. Its properties can be characterized by multiple ocean parameters: significant wave height, mean wave height, mean wave period, etc. Among them, the significant wave height (SWH) is the most important parameter in ocean waves, and the accurate prediction of SWH is also the main research direction of ocean forecasting.

[0043] Timely and accurate ocean wave forecasting not only plays an important role in marine activities such as fishery, 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 recurrent neural networks (RNN) have been widely used in single-point ocean wave forecasting, but these methods have been proven to have high lag and non-steadiness, 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 non-stationarity and high-noise characteristics of ocean waves themselves all affect the accuracy of ocean wave forecasting.

[0044] To address the above limitations, this embodiment proposes a sequence-to-sequence model based on feature selection and the big brother guidance mechanism (FE-BG-Seq2Seq) (i.e., the significant wave height prediction network model). First, the original wave height signal is decomposed by complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), and the detailed components of each order approximation are incorporated into the wave features. Then, a feature selection process based on XGBoost is used to screen out the features that are beneficial to the model prediction. Finally, the future wave height is predicted by a sequence-to-sequence model based on the big brother guidance mechanism. The sequence-to-sequence model in this embodiment alleviates the error accumulation caused by autoregression in the sequence-to-sequence model based on the big brother guidance mechanism, and enhances the model's ability to combat prediction lag. Comparing the sequence-to-sequence model in this embodiment with the CEEMDAN-BiDLSTM model in the existing technology with better current effects, experiments prove that the sequence-to-sequence model in this embodiment improves the prediction ability and accuracy for long time series.

[0045] The technical solution of this embodiment specifically includes the following content: 1. Problem restatement.

[0046] The input data of the single-point ocean wave forecast (i.e., the ocean wave feature data set) can be expressed as , and the output data can be expressed as , where is the number of ocean wave features, is the predicted significant wave height, is the input time step, is the output time step. The sequence-to-sequence model in this embodiment is trained multiple times through the first time steps and the -dimensional ocean wave features to obtain a trained sequence-to-sequence model, and then the trained sequence-to-sequence model is used to predict the time step of the ocean wave significant wave height data.

[0047] 2. Introduction to CEEMDAN decomposition.

[0048] Complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) is an adaptive noise-assisted improved algorithm of empirical mode decomposition. CEEMDAN adaptively injects white noise modulated by scale during the hierarchical decomposition process and uses the ensemble average strategy to effectively overcome the defects of mode mixing and residual noise interference in traditional methods. This method iteratively separates the intrinsic mode functions IMF of the signal from high frequency to low frequency, and through noise-assisted decomposition and multi-scale integration operations, ensures the frequency domain completeness and low reconstruction error characteristics of each IMF.

[0049] 3. Introduction to Seq2Seq.

[0050] Seq2Seq is a sequence-to-sequence learning method based on gated recurrent units (GRUs) for handling variable-length outputs and is widely used in natural language processing applications such as translation tasks and dialogue generation. It maps the input sequence to a fixed-dimensional vector using a multi-layer GRU structure or other Encoder structures to capture semantic information in the specified dimension, and then decodes the input sequence using another GRU or other Decoder structure and autoregressively generates an output sequence of a specified length.

[0051] 4. Introduction to XGBoost.

[0052] XGBoost is an efficient machine learning algorithm based on the gradient boosting framework that optimizes the objective function by iteratively adding weak learners (usually CART regression trees).

[0053] The objective function for the th iteration of XGBoost can be defined as: (1); where is the feature vector, represents the label, represents the model complexity, and the regularization term therein can be further written as: (2); where is the number of leaf nodes, is the leaf node weight, represents the complexity control coefficient, i.e., the minimum gain threshold for splitting leaf nodes, represents the L2 regularization coefficient of the leaf weights (output values).

[0054] When splitting a node into left and right child nodes , the gain of the objective function for the th iteration can be expressed as: (3); where is the gain of the decision tree, are the first and second derivatives of the left and right child nodes respectively, is the L2 regularization factor, is the complexity control coefficient.

[0055] 5. Introduction to SHAP.

[0056] SHAP is a model interpretation method based on game theory that can quantify the contribution of each feature to the model's prediction. In deep learning, due to the complexity and high non-linearity of the model, it is challenging to explain its decision logic. SHAP provides a mathematically rigorous and intuitive way to understand the internal mechanism of black-box models by randomly masking features and observing the prediction changes. The following is the calculation formula of SHAP: (4); Where S represents each feature, F represents the set of sample points of all features, represents the model output after filling non-S features with background data. By visualizing the Shapley graph, the contribution weight of each feature to the prediction result can be intuitively characterized, which includes positive and negative impacts and the degree of impact, including being able to clearly see the distribution of sample points, represents the factorial operator.

[0057] 6. Feature gain.

[0058] To capture the non-stationarity and multi-scale characteristics of the significant wave height (SWH), in this embodiment, CEEMDAN decomposition is used to perform adaptive mode 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 sea wave feature data set) to construct an enhanced feature data set.

[0059] 7. Feature selection.

[0060] 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 the features that can improve the prediction accuracy in advance and eliminate redundant features. 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 features to the model still needs to be tested. Therefore, this embodiment designs an algorithm for the feature selection process.

[0061] Specifically, by selecting the mean absolute percentage error (MAPE) as the evaluation index, the features in the enhanced feature data set 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 MAPE optimal value, 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.

[0062] 8. Sequence-to-sequence neural network model based on the big brother guidance mechanism (i.e., the significant wave height prediction network model).

[0063] Sequence-to-sequence network models are widely used in time series prediction, especially in the case where the input and output sequences have different lengths. The GRU neural network model can effectively transmit the important information of the input sequence through the gating mechanism and suppress the propagation of invalid information. At the same time, it is simpler and more efficient than the LSTM neural network model because GRU does not need to store cell information separately. In this embodiment, a prediction model suitable for significant wave height (SWH) is constructed. Both the encoder and decoder of the model in this embodiment use GRU neural network units. The encoder in this embodiment will fully learn the information of the input sequence and transmit the learned information to the decoder. Different from the conventional sequence-to-sequence network, the big brother guidance mechanism adopted in this embodiment enables the encoder to make predictions together with the decoder, referring to Figure 2 .

[0064] Specifically, given the input sequence , where is the length of the input sequence, and at the same time, the GRU hidden layer state is initialized. The encoder GRU unit processes the input sequence sequentially and updates the hidden state according to formulas (5) to (8): (5); (6); (7); (8); Where, represents the encoder update gate at the th moment, represents the activation function, represents the weight matrix of the update gate, represents the encoder hidden state at the th moment, represents the encoder input feature, represents the bias vector of the update gate, represents the encoder reset gate, represents the weight matrix of the reset gate, represents the bias vector of the reset gate, represents the encoder candidate hidden state, represents the hyperbolic tangent function, represents the weight matrix of the encoder candidate hidden state, represents the bias vector of the encoder candidate hidden state, represents the Hadamard product.

[0065] In the decoding stage, the autoregressive mechanism is used to gradually predict the future time series. To retain the sequence continuity between the encoder and the decoder, the last hidden state of the encoder As the initial hidden state of the decoder , to ensure the continuity of the hidden layer, i.e.: (9); At the same time, take the last hidden state of the encoder As the semantic encoding obtained after the encoder learns the temporal information of the input sequence , and normalize it through the Softmax layer to obtain the weighted parameters with the global information of the input sequence : (10); (11); Weighted parameters Perform the Hadamard product operation with the last value in the input sequence and pass it through a linear layer to obtain the first input of the decoder, i.e.: (12); 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 required prediction sequence length (i.e., the output time step): (13); (14); (15); (16); (17); (18); Among them, Represents the input of the decoder at the th moment, Represents the weight matrix, Represents the weighted parameters, Represents the Hadamard product, Represents the last temporal feature in the filtered enhanced feature dataset, Represents the bias vector, Represents the input of the decoder at the th moment, Represents the decoder update gate at the th moment, Represents the decoder hidden state at the th moment, Represents the decoder reset gate, Represents the decoder candidate hidden state The weight matrix representing the decoder candidate hidden state, The bias vector representing the decoder candidate hidden state, Denote the Effective significant wave height prediction result at time

[0066] In the model of this embodiment, the encoder is regarded as the older brother of the decoder. When the decoder makes each step of prediction, it is "guided". Since the encoder has fully learned the global information of the input sequence, guiding the decoder will improve the performance of the final model. Compared with the traditional autoregressive model, for each autoregressive input of the decoder based on the older brother guidance mechanism, the global information of the input sequence given by the encoder will be fully utilized, which will alleviate the accumulation of prediction errors caused by autoregression. At the same time, in this embodiment, the use of the older brother guidance mechanism enables the single-item GRU to receive not only the information transmitted by the hidden layer as input data, which can enrich the input knowledge of the GRU, thereby improving the prediction ability of the model of this embodiment.

[0067] To better illustrate the effect of the technical solution of this embodiment, the following experiments were carried out in this embodiment: This embodiment uses four metrics: 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, and MAPE measures the relative error. Relatively speaking, MAE and MAPE are not easily affected by extreme values, and RMSE is prone to amplify errors after squaring and is more sensitive to outlier data. Measures the correlation of this prediction. The closer it is to 1, the higher the credibility of this prediction. The following are the calculation formulas for the four metrics: (19); (20); (21); (22); Among them, is the true value, is the predicted value, is the average value of the true values, is the average value of the predicted values.

[0068] (1) Data preprocessing.

[0069] Ocean buoys are one of the most effective ways to obtain wave data in a small sea area. Buoy data is usually used for studying single-point wave forecasting. By deploying numerous ocean buoys globally, a large amount of ocean data can be provided stably for a long time. In this embodiment, the data of four stations, namely 41008, 41010, 46080, and 46013, are selected as examples, and their time resolution reaches 1 hour. In this embodiment, this long time series data is segmented into required samples by using the sliding method, trained according to the ratios of 0.8, 0.1, and 0.1 for the training set, validation set, and test set respectively, and trained in a certain sample batch during this process.

[0070] Since buoys work in an unstable ocean environment, data loss often occurs, which affects the training accuracy of the model. Therefore, in this embodiment, a large number of consecutive missing values are deleted, and the linear interpolation method is used to process single missing values. This filling method is mainly based on the observation of past and future data: (23); Among them, is the filled value, and are respectively the first non-missing values before and after the missing value, 、 and respectively represent the missing value, and of the time positions. This approach effectively compensates for the defects of the data itself in terms of space and time.

[0071] By normalizing the data, the problems caused by dimension and numerical differences are eliminated, the process of the optimization algorithm gradient descent is optimized, the convergence speed is increased, and the model robustness is enhanced. The normalization method is as follows: (24); Here is the normalized feature data, is the maximum value in the feature, is the minimum value in the feature. Through this processing, the feature is mapped to the 0 - 1 interval to improve the convergence speed.

[0072] (2)Feature selection and evaluation.

[0073] Table 1 shows the forecasting performances of two models in 24 hours

[0074] In this experiment, in this embodiment, XGBoost is used to screen 19 features in total, and the features with little gain in the prediction accuracy of the model are removed. By using the XGBoost regressor to rank the feature importance, the importance of each modal component can be clearly seen. The experimental results of this embodiment show that the modal component GST has the highest importance, WSPD has the second highest importance, and then APD.

[0075] Figure 3 It depicts the changing trends of the mean absolute error (MAE) and the mean absolute percentage error (MAPE) after the features are input in order of feature importance. At the initial moment, the MAE value decreases significantly as the number of features increases. When the number of input features reaches the extreme points of the MAE and MAPE of the model, and when the number of features is too large, the error further increases, showing a trend of overfitting. Therefore, these 11 features are selected for model training in the experiment.

[0076] Due to space limitations, only the comparison of the models before and after feature selection for 24 hours is presented here. As can be seen from Table 1, the prediction accuracy of the model after adding feature engineering has a significant improvement compared to before adding. Among them, Seq2Seq is before adding, and XB-Seq2Seq is after adding feature engineering. The main reason is that by separating different modalities of features and screening out the modalities that have a gain in model prediction, the stacking effect is clearly depicted, and each feature modality affects the model individually, reducing the ambiguity of full-feature prediction.

[0077] (3) Performance evaluation.

[0078] Parameters are crucial for the training effect of deep learning models. The main parameters of the innovative model FE-BG-Seq2Seq include: seqlength, hiddensize, batchsize, nhead, numlayers, and learningrate. Among them, in this embodiment, a medium-sized batchsize = 256 and a relatively small num_layer = 1 are uniformly set to balance model accuracy and model training speed. Through a series of parameter experiments, this embodiment determines the parameter settings of seqlength = 128, hiddensize = 72, nhead = 8, and learningrate = 0.002.

[0079] Table 2 shows the performance comparison results of each model for 24 hours

[0080] The comparative models selected in this experiment are Seq2Seq, XB-Seq2Seq, and CEEMDAN-BiDLSTM models. These are all models in the prior art and will not be specifically described in this embodiment. Among them, XB-Seq2Seq integrates the feature selection methods of XGBoost and Seq2Seq, and CEEMDAN-BiDLSTM integrates the CEEMDAN decomposition algorithm and the BiDLSTM model. This model has been proven to have good performance in the time series prediction of non-linear and non-steady ocean waves in recent research.

[0081] Table 3 shows the performance comparison results of each model for 48 hours.

[0082] Tables 2 and 3 list the prediction results of the significant wave height of the ocean waves for four models at 24 hours and 48 hours, using long time series data (for example, data from 2005 to 2023 is selected for Station 41008). From the vertical comparison, it can be seen that the prediction effects of the four models at 24 hours are better than those at 48 hours, proving that the four models have good prediction capabilities for short-term predictions within 24 hours. From the horizontal comparison, the model (FE-BG-Seq2Seq) of this embodiment has achieved the best results, followed by CEEMDAN-BiDLSTM. Taking MAE as an example, for the 24-hour forecast, the errors of the model in this embodiment at each station are reduced by 15%, 18%, 17%, and 21% compared with CEEMDAN-BiDLSTM respectively. For the 48-hour forecast, the errors at each station are reduced by 22%, 24%, 22%, and 25% compared with CEEMDAN-BiDLSTM respectively. This shows that the model in this embodiment has a relatively large improvement compared with CEEMDAN-BiDLSTM after adding the elder brother guidance mechanism.

[0083] (4) Model verification.

[0084] Feature analysis is very important for understanding the internal mechanism of deep learning models. Therefore, in this experiment, this embodiment innovatively uses SHAP values to analyze the network model constructed based on the elder brother guidance mechanism proposed in this embodiment, so as to quantitatively and visually display the influence of each ocean wave feature on the prediction result. This can also further verify the correctness of the feature selection process in this embodiment for screening features and the correctness of the model for assigning weights to different features.

[0085] Since SHAP can only recognize two dimensions, batchsize and hiddensize, this embodiment aggregates the seqlength dimension and takes the average over the time steps. This embodiment uses the training set as the background sample and the test set as the test instance, and draws a representative SHAP contribution diagram (swarm plot).

[0086] Figure 4 This is the SHAP value swarm plot drawn by the model of this embodiment. 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 corresponding feature of the sample point. The size of the abscissa represents the magnitude of the SHAP influence, 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 feature importance decreases from top to bottom in the SHAP plot, and the displayed results are consistent with the feature importance ranking in the feature selection process. Moreover, the sample points corresponding to features such as GST, WSPD, and APD are densely distributed on the positive half-axis, indicating that these features have an obvious positive correlation with the significant wave height, which is consistent with the results obtained in the feature selection evaluation. All of these further verify the correctness of the model in this embodiment.

[0087] Compared with the prior art, the method of this embodiment has the following advantages: Previous models such as ARIMA, LSTM, and Seq2Seq often perform short-term prediction based on short sequence data, and have defects such as hysteresis and non-steadiness, making it difficult to capture the global multi-dimensional connections of the data. The FE-BG-Seq2Seq model proposed in this embodiment integrates feature engineering, the elder brother guidance mechanism, and the sequence-to-sequence main model. It can not only effectively transmit the information of the input sequence through the sequence-to-sequence network, but also make the encoder and decoder predict in parallel through the elder brother guidance mechanism to reduce error accumulation.

[0088] In the experiment of this embodiment, by horizontally comparing the best parameter Seq2Seq with XB-Seq2Seq first, it is proved that the feature selection process can significantly improve indicators such as MAE, MAPE, R2, and RMSE. Then, by comparing CEEMDAN-BiDLSTM with the model of this embodiment, it is proved that its long-term prediction ability is enhanced and the prediction accuracy rate under extreme SWH samples can be maintained.

[0089] In addition, this embodiment also uses SHAP to enhance the interpretability of the model. SHAP can well verify the correctness of the previous feature selection process and the features that are beneficial to the model prediction, and internally analyzes the black box model of the deep learning model FE-BG-Seq2Seq in this embodiment.

[0090] The above has described the embodiments of the present application in detail with reference to the accompanying drawings, but the present application is not limited to the above embodiments. Within the knowledge scope of those of ordinary skill in the art to which the present application pertains, various changes can be made without departing from the purpose of the present application.

Claims

1. An effective significant wave height prediction method based on feature engineering and the elder brother guidance mechanism, characterized in that, The method includes: Obtaining a sea wave feature dataset containing the original wave height time series signal; Performing modal decomposition on the original wave height time series 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; Based on a gated recurrent unit, constructing an effective wave height prediction network model using the elder brother guidance mechanism; Inputting the screened enhanced feature dataset into the effective wave height prediction network model for prediction to obtain an effective wave height prediction result.

2. The effective wave height prediction method based on feature engineering and elder brother guidance mechanism according to claim 1, wherein The constructing of the effective wave height prediction network model based on a gated recurrent unit and using the elder brother guidance mechanism includes: Using a plurality of cascaded gated recurrent units to construct an encoder and a decoder; According to the encoder and the decoder, constructing an effective wave height prediction network model, where the encoder is used as the elder brother of the decoder to guide the decoder to perform prediction.

3. The effective wave height prediction method based on feature engineering and the elder brother guidance mechanism according to claim 2, wherein The inputting of the screened enhanced feature dataset into the effective wave height prediction network model for prediction to obtain an effective wave height prediction result includes: Processing the screened enhanced feature dataset through the plurality of gated recurrent units in the encoder to obtain the last hidden state of the encoder; Taking the last hidden state of the encoder as the initial hidden state of the decoder, and taking the last hidden state of the encoder as the semantic encoding learned by the encoder; Normalizing the semantic encoding to obtain a weighted parameter; Performing a Hadamard product on the weighted parameter and the last time series feature in the screened enhanced feature dataset, and after linear layer operation, obtaining the first input of the decoder; Inputting 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.

4. The effective wave height prediction method based on feature engineering and elder brother guidance mechanism according to claim 3, characterized in that, The normalizing of the semantic encoding to obtain a weighted parameter includes: ； ； Among them, represents semantic encoding, represents the last hidden state of the encoder, represents the weighting parameter, represents the normalization layer.

5. The effective wave height prediction method based on feature engineering and the elder brother guidance mechanism according to claim 3, characterized in that The processing of the screened enhanced feature dataset through the plurality of gated recurrent units in the encoder includes: ； ； ； ； Among them, represents the encoder update gate at the moment, represents the activation function, represents the weight matrix of the update gate, represents the moment encoder hidden state, represents the encoder input feature, represents the bias vector of the update gate, represents the encoder reset gate, represents the weight matrix of the reset gate, represents the bias vector of the reset gate, represents the encoder candidate hidden state, represents the hyperbolic tangent function, represents the weight matrix of the encoder candidate hidden state, represents the bias vector of the encoder candidate hidden state, represents the Hadamard product.

6. The effective wave height prediction method based on feature engineering and the elder brother guidance mechanism according to claim 3, characterized in that The performing of the Hadamard product on the weighted parameter and the last time series feature in the screened enhanced feature dataset, and after linear layer operation, obtaining the first input of the decoder includes: ； Among them, represents the first input of the decoder, represents the weight matrix, represents the weighting parameter, represents the Hadamard product, represents the last temporal feature in the filtered enhanced feature dataset, represents the bias vector, represents the linear layer operation.

7. The effective wave height prediction method based on feature engineering and the elder brother guidance mechanism according to claim 3, characterized in that The inputting 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 includes: ； ； ； ； ； ； Among them, represents the input of the decoder at the moment, represents the weight matrix, represents the weighting parameter, represents the Hadamard product, represents the last temporal feature in the filtered enhanced feature dataset, represents the bias vector, represents the input of the decoder at the represents the decoder update gate at the moment, represents the activation function, represents the weight matrix of the update gate, represents the decoder hidden state at the moment, represents the bias vector of the update gate, represents the decoder reset gate, represents the weight matrix of the reset gate, represents the bias vector of the reset gate, represents the decoder candidate hidden state, represents the hyperbolic tangent function, represents the weight matrix of the decoder candidate hidden state, represents the bias vector of the decoder candidate hidden state, represents the effective wave height prediction result at the moment.

8. The effective wave height prediction method based on feature engineering and elder brother guidance mechanism according to claim 1, characterized in that, The performing of modal decomposition on the original wave height time series signal to obtain a plurality of intrinsic mode functions and a plurality of residual terms includes: Performing modal decomposition on the original wave height time series signal using adaptive noise complete ensemble empirical mode decomposition to obtain a plurality of intrinsic mode functions and a plurality of residual terms.

9. The effective wave height prediction method based on feature engineering and the elder brother guidance mechanism according to claim 1, wherein The performing of feature screening on the features in the enhanced feature dataset to obtain a screened enhanced feature dataset includes: Using an XGBoost regression model to rank the features in the enhanced feature dataset to obtain a ranked enhanced feature dataset; Input the features in the sorted enhanced feature dataset into the XGBoost regression model in sequence for prediction to obtain the current predicted value, and calculate the target mean absolute percentage error between the current predicted value and the true value; If the target mean absolute percentage error is smaller than the historical mean absolute percentage error, retain the features input into the XGBoost regression model until after continuously inputting multiple features into the XGBoost regression model, the calculated target mean absolute percentage error is equal to or greater than the historical mean absolute percentage error, end the iterative prediction, and obtain the filtered enhanced feature dataset.

10. The effective wave height prediction method based on feature engineering and elder brother guidance mechanism according to claim 1, wherein The feature screening of the features in the enhanced feature dataset to obtain the filtered enhanced feature dataset includes: Use the XGBoost regression model to sort the features in the enhanced feature dataset to obtain the sorted enhanced feature dataset; Input the features in the sorted enhanced feature dataset into the XGBoost regression model in sequence for prediction to obtain the current predicted value, and calculate the target mean absolute percentage error between the current predicted value and the true value; If the target mean absolute percentage error is greater than or equal to the historical mean absolute percentage error, remove the features input into the XGBoost regression model until after continuously inputting multiple features into the XGBoost regression model, the calculated target mean absolute percentage error is greater than or equal to the historical mean absolute percentage error, end the iterative prediction, and obtain the filtered enhanced feature dataset.

Images