CN-121980242-A - Effective wave height real-time prediction method and device based on dynamic residual error correction

CN121980242ACN 121980242 ACN121980242 ACN 121980242ACN-121980242-A

Abstract

The invention provides an effective wave height real-time prediction method and device based on dynamic residual error correction, and relates to the field of ocean engineering. The method comprises the specific steps of obtaining wave height data, carrying out multidimensional feature screening, constructing an integrated filter integrating L1 trend filtering and variation modal decomposition, utilizing sea image optimization algorithm optimizing parameters, introducing causal sliding window extracting features to construct time sequence input tensors, adding noise for standardization, inputting a stacked bidirectional long-short-term memory network based on an attention mechanism to obtain basic prediction values, calculating manifold coherent structures, PID dynamics and physical statistics features, cascading the basic prediction features to construct comprehensive element feature vectors, constructing LightGBM architecture, fitting prediction residual errors after optimizing by the sea image optimization algorithm, finally, carrying out linear reconstruction based on dynamic safety threshold constraint, and outputting real-time correction results. The method and the device remarkably improve the real-time precision and the robustness of effective wave height prediction by utilizing the depth mining error evolution rule of manifold geometric features.

Inventors

LI JING
LU HAOYUN
Jiao Jiange

Assignees

中国计量大学

Dates

Publication Date: 20260505
Application Date: 20260123

Claims (9)

1. The method for predicting the effective wave height in real time based on dynamic residual correction is characterized by comprising the following steps of: configuring a plurality of auxiliary features; Acquiring historical data for wave height prediction, preprocessing the historical data, and acquiring feature data corresponding to each auxiliary feature; Carrying out correlation analysis and nonlinear importance evaluation on the characteristic data, screening and determining input characteristics serving as a basic prediction model; Constructing an integrated filter, performing parameter optimization, simulating a real-time filtering environment by using a causal sliding window mechanism, and constructing a time sequence input tensor containing a trend component and a multi-scale detail component based on the input characteristics; The processed time sequence input tensor is added with noise for standardization and then is input into a two-way long-short-term memory network based on an attention mechanism, and a basic prediction model is obtained through training; Carrying out dynamic phase space reconstruction on the wave height sequence based on a sliding window, extracting the structural characteristics of the flow shape coherent, and cascading the predicted sequence characteristics, PID dynamic characteristics and physical statistical characteristics of a basic prediction model to construct a comprehensive element characteristic vector; Calculating a residual matrix of the basic prediction model on the verification set, taking the comprehensive element feature vector as input, taking the residual matrix as a target, and training LightGBM the model to obtain a dynamic residual correction model; in an online prediction stage, real-time observation data are received, model input is synchronously updated by utilizing a causal sliding window mechanism, and a corresponding wave height prediction result is generated by combining a basic prediction value output by a basic prediction model and a residual error compensation value output by a dynamic residual error correction model.
2. The method of claim 1, wherein the preprocessing includes outlier rejection, time-series re-indexing of raw data, vector reconstruction of the ring features, and interpolation padding of the linear features.
3. The method of claim 1, wherein the input data screening step comprises: calculating the spearman correlation coefficient between each auxiliary feature and the target wave height to obtain corresponding correlation; utilizing a random forest model to learn a nonlinear mapping relation between the auxiliary features and the target wave height, and evaluating the parameter prediction contribution degree; and screening and determining the input characteristics of the basic prediction model through the correlation and the contribution degree corresponding to each auxiliary characteristic.
4. The method of claim 1, wherein the step of constructing an integrated filter and a time-ordered input tensor comprises: Constructing an integrated filter model fused with an L1 trend filtering algorithm and a variation modal decomposition algorithm; Initializing a search space of a sea image optimization algorithm, and setting smoothing parameters of L1 trend filtering, penalty factors of variation modal decomposition, modal numbers and weights of an integrated filter as variables to be optimized; A causal sliding window mechanism is introduced to simulate a real-time filtering environment, and data in a window are independently decomposed and a last time step component is intercepted in an iterative optimization process; Executing population iteration, determining the optimal hysteresis steps of the filtering sequence and the original sequence, and carrying out translation correction on the filtering sequence according to the optimal hysteresis steps; and outputting a global optimal parameter combination by taking the loss of the verification set as an fitness function, extracting a trend component and a multi-scale detail component based on the combination, and splicing with the input characteristic sequence determined by screening to construct a time sequence input tensor.
5. The method of claim 1, wherein the basic predictive model training process step comprises: based on the input characteristic sequence and the filtering sequence, constructing a standardized time sequence input tensor, and injecting Gaussian noise; Inputting the noise-added standardized time sequence input tensor into a stacked bidirectional long-short-term memory network to capture bidirectional time sequence dependence, and carrying out self-adaptive weighting on a hidden state sequence by using an attention mechanism to generate a feature vector which aggregates global context information; and outputting a prediction result by using the feature vector, and iteratively updating parameters based on the prediction deviation until the model converges to obtain a basic prediction model.
6. The method of claim 1, wherein the steps of extracting manifold coherent structural features and constructing a composite meta-feature vector comprise: Performing dynamic phase space reconstruction on the wave height sequence based on the sliding window to generate a track flow reflecting sea state evolution, and tracking the extinction process of homomorphism characteristics by using simplex complex filtering to generate a persistence graph; calculating manifold coherent structural features reflecting the structural evolution complexity based on the persistence graph; extracting a prediction sequence of a basic prediction model and a first-order differential trend of a historical observation value, and combining the prediction sequence and the first-order differential trend to form a basic feature vector; And carrying out numerical value splicing on the basic feature vector, the PID feature vector, the physical statistics feature vector and the manifold coherent structure feature vector to construct the comprehensive element feature vector.
7. The method of claim 6, wherein the training step of the dynamic residual correction model comprises: calculating the output deviation of the basic prediction model on the verification set, and constructing a residual target matrix with the number of lines as the number of samples and the number of columns as the prediction step length; Adopting a multi-output LightGBM regression architecture for configuring Huber robust loss functions, and iteratively optimizing model super-parameters by utilizing a sea image algorithm; And performing supervision training based on the optimal parameters by taking the comprehensive meta-characteristics as input and column vectors of the residual matrix as labels to obtain a dynamic residual correction model.
8. The method according to claim 1, wherein the specific implementation step of the online prediction phase comprises: adding real-time observation data to a buffer sequence, decomposing and intercepting last bit components by using a causal sliding window mechanism to construct an input tensor, and inputting a basic model to obtain a reference predicted value P base ; Constructing a comprehensive meta-feature vector based on the updated residual sequence and the manifold coherent structural features, inputting a dynamic residual correction model and outputting an initial error compensation value delta P raw ; Calculating a dynamic safety threshold L according to the reference predicted value, and limiting the initial error compensation value in a safety interval [ -L, L ] after the initial error compensation value is scaled by a damping coefficient to obtain a final compensation value delta P final , and outputting a final result through linear superposition, wherein the calculation formulas of the dynamic safety threshold and the final predicted result are as follows: L=max(γ·P base , ) ΔP final =min(max(η·ΔP raw ,-L),L) P final =P base +ΔP final wherein, gamma is the relative threshold coefficient, For the minimum absolute threshold, η is the corrected damping coefficient, and P final is the final output corrected predicted result.
9. An effective wave height real-time prediction device based on dynamic residual correction, which is characterized by comprising: the data processing module is used for acquiring and preprocessing the multidimensional marine environment parameters to obtain a normalized data set; The feature screening module is used for screening and determining the input features of the basic prediction model through the Style-Kalman correlation coefficient and the random forest nonlinear importance score of the normalized data set; the self-adaptive decomposition module is used for optimizing signal decomposition parameters of the effective wave height, carrying out endpoint processing on a decomposition mode based on a causal mechanism, and constructing a time sequence input tensor containing a trend component and a multi-scale detail component by combining input characteristics; and the basic prediction module is used for inputting the noise-added standardized time sequence input tensor into the stacked two-way long-short-term memory network based on the attention mechanism. Training to obtain a basic prediction model by capturing bidirectional time sequence dependence and global context characteristics; The dynamic feature construction module is used for receiving the prediction sequence of the basic prediction model and constructing a basic feature vector, extracting manifold coherent structural features of the wave height sequence, combining residual PID dynamic features and physical statistical features, and splicing and constructing a comprehensive meta feature vector; The dynamic residual error correction module is used for constructing a multi-output LightGBM model group by taking the comprehensive meta-characteristic as an input and verifying set residual error as a target; The real-time correction output module is used for receiving the real-time observation data, carrying out linear fusion on the reference wave height predicted value and the error compensation value generated in real time, and generating a corresponding wave height predicted result after dynamic safety threshold constraint processing.

Description

Effective wave height real-time prediction method and device based on dynamic residual error correction Technical Field The application relates to the technical field of effective wave height prediction, in particular to an effective wave height real-time prediction method and device based on dynamic residual error correction. Background The China is accelerating the mass utilization of the propulsion ocean energy under the driving of global energy low-carbon transformation and 'double-carbon' targets. The effective wave height of the sea surface is a key parameter for describing the sea state, and has important significance for wave energy capture research. Currently, the prediction of effective wave height mainly includes numerical simulation and data driving models. The numerical simulation method is limited by the complexity of solving the physical equation, the calculation time consumption is long, and the timeliness requirement of real-time scheduling is difficult to meet. While most data-driven models represented by deep learning adopt an offline training mode, model parameters are difficult to change once established, and the static mapping mechanism causes the model to lack real-time self-adaptive adjustment capability when facing the transient mutation of the marine environment, so that systematic prediction deviation is easy to generate. In the aspect of data preprocessing, signal processing methods such as Variational Modal Decomposition (VMD) and the like are mostly adopted in the prior art to extract multi-scale characteristics of waves. However, VMDs essentially belong to a global optimization algorithm that relies on the full amount of data. In an online real-time prediction scenario, direct decomposition can lead to lack of boundary constraints at the signal end due to failure to acquire future data, thereby causing serious end point divergence effects and modal aliasing. The distortion of the real-time characteristic signal seriously affects the stability and accuracy of prediction. On the other hand, aiming at correction of effective wave height prediction deviation, the existing residual error correction mechanism is mostly limited to utilization of low-order statistical rules or time sequence autocorrelation of residual error sequences, a high-dimensional manifold structure implicit in the wave field evolution process is generally ignored, and deep analysis of residual errors cannot be carried out from the view angles of high-dimensional geometric space and dynamic system evolution. Due to the absence of multidimensional dynamic information, when the correction model is used for dealing with unstable and complex sea conditions, the residual evolution mode under the unstable sea conditions is difficult to accurately identify, and robust dynamic compensation cannot be realized. Disclosure of Invention The invention mainly aims to provide an effective wave height real-time prediction method and device based on dynamic residual error correction, which are used for solving the technical problem that the effective real-time self-adaptive compensation cannot be realized in online real-time application due to the fact that the existing prediction correction strategy ignores decomposition causal constraint and high-dimensional error dynamics characteristics. In a first aspect, the present invention provides an effective wave height real-time prediction method based on dynamic residual correction, including: the system acquires multidimensional ocean environment parameters, and preprocesses historical characteristic data to obtain an effective wave height data set; carrying out correlation analysis and nonlinear importance evaluation on the effective wave height data set, screening and determining input characteristics of a basic prediction model; Constructing an integrated filter, performing parameter optimization, simulating a real-time filtering environment by using a causal sliding window mechanism, and constructing a time sequence input tensor containing a trend component and a multi-scale detail component based on the input characteristics; normalizing the time sequence input tensor processed by the optimizing parameters, inputting the normalized time sequence input tensor into a two-way long-short-term memory network based on an attention mechanism, and training to obtain a basic prediction model (M1); Carrying out dynamic phase space reconstruction on the wave height sequence based on a sliding window, extracting the structural characteristics of the flow shape coherent, and cascading the predicted sequence characteristics, PID dynamic characteristics and physical statistical characteristics of a basic prediction model to construct a comprehensive element characteristic vector; Calculating a residual matrix of the basic prediction model on the verification set, taking the comprehensive element feature vector as input and taking the residual matrix as a target, and training LightGBM a mode