CN114912077A - Sea wave forecasting algorithm integrating random search and mixed decomposition error correction - Google Patents

Abstract

A wave forecasting algorithm integrating random search and mixed decomposition error correction relates to the technical field of artificial intelligence and comprises the following steps: step 1, acquiring buoy historical observation data based on buoy observation, integrating and processing ocean and meteorological data closely related to sea wave change, and constructing a sea wave forecasting database; step 2, analyzing sensitivity factors; step 3, constructing a sea wave forecasting model; and 4, acquiring a sea wave prediction error time sequence, and adding the prediction error and the initial prediction wave height to obtain the corrected prediction sea wave height. The method refers to the input of a numerical prediction model, carries out data driving by sea wave related elements at the prediction time, simultaneously carries out automatic optimization on the hyper-parameters of the model to optimize the model structure, and carries out error correction on sea wave prediction by combining a data decomposition algorithm and a deep learning mixed decomposition error correction model, so as to improve the sea wave prediction precision based on deep learning, improve the prediction timeliness and reduce the sea wave prediction time lag.

Description

Sea wave forecasting algorithm integrating random search and mixed decomposition error correction

Technical Field

The invention relates to the technical field of artificial intelligence, in particular to a sea wave forecasting algorithm integrating random search and mixed decomposition error correction.

Background

China is wide in region, coastlines are long, coastal regions are densely populated, activities on the sea are frequent, and the sea is also one of countries full of sea wave disasters. According to the '2020 Chinese Marine disaster Notification' issued by the national Marine agency in 2021, it is shown that [1 ]: sea wave disasters still form one of the main marine disasters in China, 8.32 million yuan of direct economic loss is caused by various marine disasters in China in 2020, and 6 people die (including missing). Wherein, the death (including the missing) of the personnel is all caused by the sea wave disaster. The method is characterized in that the disastrous wave process with the effective wave height of more than 4.0 meters (inclusive) is carried out 36 times in the offshore process, and wave disasters occur 8 times, so that 22 ships are damaged.

Therefore, timely and accurate sea wave forecasting has important significance in coastal engineering construction, offshore operation, ship navigation safety and other aspects. However, due to the complex wave origin and the strong non-linearity of the waves themselves, this makes accurate wave prediction difficult. The wave forecasting development is subjected to three stages of traditional wave forecasting, numerical wave forecasting and machine learning forecasting. The traditional sea wave forecasting method mainly comprises an empirical formula and a semi-empirical formula, and the traditional sea wave forecasting method is widely applied to an effective wave forecasting method, a PNJ spectrum forecasting method, a Wilson experience forecasting formula, an energy balance derived spectrum forecasting method proposed by the Saint Hospital, China and the like.

With the improvement of computer computing power and the continuous deepening of people's understanding of the mechanism of sea waves since the twentieth century, a numerical forecasting method of sea waves is proposed and paid attention to by people. Since the mid-fifties, the numerical prediction mode continuously breaks through the technical bottleneck, the numerical prediction mode is developed to the third-generation sea wave numerical mode so far, the problem that the wind field cannot be adapted to rapid change in the past is broken through, various wave generation and dissipation physical processes including the nonlinear action among waves are fully considered, and the numerical prediction mode becomes the mainstream mode of the current sea wave prediction. At present, the most widely used modes include WAM mode, swan (surface Wave modes) mode, and WAVEWATCHIII mode. The SWAN mode employs a fully-implicit calculation method and considers physical processes of wave propagation, diffraction, fragmentation, reflection and the like in shallow water, and is often applied to coastal areas with a resolution of 500 m to more than ten kilometers, while the WAM and WAVEWATCHIII use an explicit or semi-implicit difference technique and consider the processes of surge and wave dissipation, wind energy input and the like, and are generally applied to ocean scales of more than 20 kilometers.

In recent years, with the rapid development of machine learning methods, the method attracts the attention of numerous scholars at home and abroad, and provides a new idea for sea wave forecasting. Machine learning is a cross discipline that includes many disciplines, including mathematical statistics, higher mathematics, computers, and the like. The machine learning method has good self-adaptive learning and nonlinear mapping capabilities, does not need to solve the physical mechanism of the occurrence of things clearly, and is suitable for processing the nonlinear problems that the physical mechanism is complex and the causal relationship and the inference rule are difficult to determine. In addition, the types and the quantity of the oceanographic observation data and the oceanographic reanalysis data are greatly increased along with the scientific progress and the rapid development of the computer technology, and the data problem required by machine learning is well solved. The deep learning is taken as the most popular subdivision field of machine learning in the last decade, the method has strong advantages in the fields of voice recognition, image classification and the like, solves the problems that the machine learning is very difficult historically for a long time, and has great application potential in the ocean field.

At present, sea wave forecasting of deep learning is mainly divided into two types: one method is to forecast by combining relevant factors of sea waves as model input, for example, Fan et al constructs a wave forecasting model based on LSTM by taking wave height, wind speed and wind direction as input. The results show that the LSTM algorithm is superior to BP neural networks, Extreme Learning Machines (ELMs), Support Vector Machines (SVMs), residual error networks (resnets) and Random Forest (RF) algorithms in terms of wave prediction. The other method is to combine a data decomposition method and take a decomposition subsequence of historical wave height as input, for example, Deka et al carries out wave height prediction based on a mixed wave height prediction model combining discrete wavelet decomposition and an artificial neural network ANN, and proves that the result is superior to the ANN model.

Although the traditional wave forecasting method plays a certain role in wave forecasting, certain limitations still exist in the universality and accuracy of forecasting areas and different wave processes. As the mainstream wave forecasting model at present, the numerical model is solved mainly based on an energy balance equation, and the method has the problems of large calculated amount, non-convergence and the like and is difficult to solve in a short time. Although deep learning is widely applied to sea wave forecasting, a deep learning model based on sea wave elements as input usually considers the elements as elements at historical time, and the relevance and the influence degree of the deep learning model are far lower than those at forecasting time, so that the forecasting precision and timeliness of the deep learning forecasting model based on the method are lower than those of a forecasting model driven by numerical storm elements at forecasting time. Although the nonlinearity of sea waves is reduced by the deep learning forecasting model based on the data decomposition method, the influence of other elements cannot be considered, and the accuracy is insufficient only based on the self rule of the historical sea waves. Meanwhile, many hyper-parameters in the deep neural network model need to be adjusted, such as the number of model layers, the number of units, an activation function, an optimizer, a learning rate and the like, and the hyper-parameters cannot be continuously optimized iteratively through the training process of the neural network but need to be set manually. Currently, in deep learning research of sea wave forecasting, adjustment of hyper-parameters is mostly selected manually. The method for manually adjusting the hyper-parameters wastes a large amount of time and cannot give full play to the performance of the model. However, the related research of the sea wave forecasting and correcting model based on deep learning is still very limited, and the main application field of the sea wave forecasting and correcting model is also the error correction of the data forecasting model.

Disclosure of Invention

The invention provides a sea wave forecasting algorithm integrating random search and mixed decomposition error correction, which aims to refer to the input of a numerical forecasting model, carry out data drive on sea wave related elements at the forecasting time, simultaneously carry out automatic optimization on a model hyper-parameter so as to optimize a model structure, and carry out error correction on sea wave forecasting by combining a data decomposition algorithm and a deeply-learned mixed decomposition error correction model so as to improve the sea wave forecasting precision based on deep learning, improve the forecasting timeliness and reduce the sea wave forecasting time lag.

In order to achieve the purpose, the technical scheme of the invention is as follows:

a wave forecasting algorithm integrating random search and mixed decomposition error correction comprises the following steps: step 1, acquiring buoy historical observation data based on buoy observation, integrating and processing ocean and meteorological data closely related to sea wave change, and constructing a sea wave forecasting database; step 2, sensitivity factor analysis is carried out, and input parameters and step length of the sea wave forecasting model are determined; step 3, constructing a sea wave forecasting model; and step 4, obtaining a sea wave prediction error time sequence, and adding the prediction error and the initial prediction wave height to obtain the corrected prediction sea wave height.

Preferably, in the step 1, the wind speed, the wind direction, the air temperature, the air pressure, the temperature, the humidity, the wave height, the wave direction and the wave period are used as input parameters to be evaluated of the sea wave height forecasting model. .

Preferably, the step 2 comprises the following specific steps: removing the data abnormal value by using a Hampel filter, and defining an outlier as an element which is within a window length specified by the window and has a difference with a local median by more than three times and locally converted MAD; filling missing values by using a cubic polynomial interpolation value, performing correlation analysis on related elements of sea waves by using a Pearson correlation coefficient, and normalizing input parameters with different orders of magnitude; determining the optimal input characteristic and the corresponding step length through an input step length test;

the calculation formula of the Pearson correlation coefficient is as follows:

in formula (1): n is the sequence length, o _i Is the wave height sequence value and is the wave height sequence value,

is the average value of the wave height sequence, u _i In order to input the parameters to be evaluated,

the average value of the input variables to be evaluated;

the calculation formula for normalizing the input parameters with different orders of magnitude is as follows:

Data _nor ＝(Data _nor,max -Data _nor,min )·(Data-Data _min )/(Data _max -Data _min )+Data _nor,min (2)

in the formula (2): data _nor For the sequences after normalization, Data _nor，min To normalize the interval upper bound, Data _nor，min For lower limit of normalization interval, Data is the sequence before normalization _max To normalize the maximum of the pre-sequence, Data _min Is the minimum value before normalization.

Preferably, in step 2, the step of determining the input parameters of the wave height prediction model is as follows:

(a) selecting a parameter with highest correlation as a variable to be evaluated according to a correlation coefficient result, selecting 1-k advanced step length variables to be evaluated and the effective wave height of the past moment as input, and inputting the input into a model;

(b) training a deep learning model, and evaluating a forecast result;

(c) if the model precision is improved after k variables to be evaluated are added, the variables to be evaluated are reserved, and the precision is reduced until k +1 step lengths are input;

(d) reserving the last optimized variable, and adding a new variable with the highest correlation in the remaining variables as a variable to be evaluated;

(e) repeating the steps (a) to (d) until all the variables to be evaluated are tested;

(f) and obtaining the input parameters of the wave height forecasting model according to the test result tested in the step.

Preferably, the step 3 comprises the following specific steps: the method comprises the steps of adopting a long-short term memory neural network as a forecasting model, dividing a training set and a verification set, adopting a random search algorithm to optimize hyper-parameters of a sea wave forecasting model, and outputting a forecasting result.

Preferably, in the step 3, the long and short term memory neural network solves the problems of gradient disappearance and gradient explosion through a gate type unit structure, and the problem is solved through an input gate (i) _t ) Forget gate (f) _t ) Output gate (o) _t ) The memory unit is used for controlling information to selectively enter the memory unit for storage or forgetting; each gate performs control output through an activation function Sigmoid;

the Sigmoid function is defined as:

i _t ＝σ(W _i ·[S _t-1 ，x _t ]+b _i ) (4)

f _t ＝σ(W _f ·[S _t-1 ，x _t ]+b _f ) (5)

o _t ＝σ(W _o ·[S _t-1 ，x _t ]+b _o ) (6)

in formulae (3) to (6), S _t-1 Representing the output at time t-1, x _t Denotes the input at the current time t, σ denotes the Sigmoid activation function, W _f 、W _i 、W _o Respectively, the weight parameters of the forgetting gate, the input gate and the output gate, b _f 、b _i 、b _o The offset coefficients of the forgetting gate, the input gate and the output gate are respectively.

Preferably, in step 3, a random search algorithm is used to perform hyper-parameter adjustment to optimize the model structure including the number of neural network layers and the number of neurons, and the search strategy is as follows:

(a) defining a search space;

(b) for the hyper-parameter with the search range being an interval, randomly sampling according to the given interval; for a hyper-parameter whose search range is a limited number of lists, equal probability sampling is performed in a given list;

(c) traversing the niter group sampling results obtained in the step (b); if the given search ranges are lists, the samples are not put back for niter times;

(d) and comparing the values of the target functions of the points meeting the constraint conditions one by one, discarding the combination with large error, reserving the combination with small error, and finally obtaining the approximate solution of the optimal solution.

Preferably, the step 4 comprises the following specific steps: obtaining a sea wave prediction error time sequence, decomposing the error time sequence based on improved adaptive noise ensemble empirical mode decomposition (ICEEMDAN), optimizing a mixed error correction model structure by adopting a random search algorithm, outputting a prediction error after training and verification, and adding the prediction error and an initial prediction wave height to obtain the corrected prediction sea wave height.

Preferably, in the step 4, the neural network prediction value is set as

The measured data at the corresponding time is

Prediction error sequence E _i Expressed as:

in formula (7):

is a predicted value of the wave height prediction model,

for measured data at the corresponding time, E _i Is a prediction error;

carrying out pure randomness inspection, namely white noise inspection on the error sequence to ensure that the error time sequence has a self rule; the autocorrelation coefficients are used for checking, if the sequence is white noise, the approximate mean value of the autocorrelation coefficients of the delay nonzero period of the sequence is 0; for time series { X _t In respect of x _t And x _t-k The autocorrelation coefficient between them is defined as:

in the formula (8), Cov represents covariance, and Var represents variance. When k is 0, ρ ₀ ＝1；X _t And X _t-k The meaning of the autocorrelation coefficients of (a) is to give the correlation of the function itself.

Preferably, in the step 4, the improved adaptive noise ensemble empirical mode decomposition (ICEEMDAN) is an improved algorithm of the Empirical Mode Decomposition (EMD); EMD is proposed by Huang et al in 1998, is an algorithm for carrying out adaptive decomposition according to data, the method does not need to manually divide decomposition layer numbers, but carries out adaptive decomposition, is convenient for researchers to research, has the precision subjected to experimental verification of prediction in fields such as weather and the like, and can decompose original data into a plurality of Intrinsic Mode Functions (IMFs) and a residual error term through EMD decomposition; the decomposition steps are as follows:

(a) solving upper and lower envelope lines according to upper and lower extreme points of the original sequence;

(b) obtaining the mean value envelope curve by solving the mean value of the upper envelope curve and the lower envelope curve;

(c) subtracting the mean envelope curve from the original data to obtain a middle signal;

(d) judging whether the intermediate signal meets 1) whether the number of the extreme points and the number of the zero crossing points are equal or the phase difference is not more than one at most in the whole data section; 2) at any time, the average value of an upper envelope formed by the local maximum value points and a lower envelope formed by the local minimum value points is zero, namely the upper envelope and the lower envelope are locally symmetrical relative to a time axis;

(e) if the two conditions in (d) are not satisfied, repeating (a) to (d) based on the intermediate signal; if the condition is met, the signal is regarded as an IMF component, the IMF component is subtracted from the original signal to serve as a new original sequence, and the steps (a) to (d) are repeated to obtain IMFs 2 and … until the remaining terms contain less than 2 extreme values and further decomposition is not needed;

let x be the original signal, E _k (. h) is the kth IMF component obtained after EMD decomposition of the signal, w (i) is unit variance white noise with a mean value of zero, and M (. cndot.) is the local mean value of the signal;

the calculation steps for improving adaptive noise ensemble empirical mode decomposition (icemdan) are as follows:

(a) performing EMD operation on the L mixed original signals with the white noise superimposed:

x ⁱ ＝x+β ₀ E ₁ (w(i))，i＝1，...，L (9)

β ₀ ＝λ ₀ std(x)/std(E ₁ (w (i)) (10) in the formula (10), std (. cndot.) represents the calculation standard deviation, and λ is recommended ₀ A value of 0.2;

the first remainder calculation formula is:

in formula (11): r represents residual components, IMF represents a natural mode function, and n represents the number of IMFs;

(b) the modal component 1 IMF may be expressed as follows:

IMF ₁ (t)＝x-r ₁ (t) (12)

in formula (12): x is the original signal, r ₁ (t) a first remainder;

(c) calculate the kth remainder:

in formula (13): r is _k (t) is the kth remainder

(d) Calculate the kth IMF component:

IMF _k (t)＝r _k-1 (t)-r _k (t)，k＝2，3，... (14)

(e) repeating (c) and (d) until the remainder does not require further decomposition, the decomposition result of the original sequence being expressed as follows:

in formula (15): x (t) is the original signal, IMF is the modal component, r _n (t) is the remainder.

The invention discloses a sea wave forecasting algorithm integrating random search and mixed decomposition error correction, which has the beneficial effects that:

1. the invention adopts sensitivity factor analysis and random search algorithm to determine the input parameters and the model structure of the forecasting model, thereby realizing timely and accurate rapid wave height forecasting of the sea waves.

2. The invention provides a sea wave prediction error correction technology based on hybrid decomposition, thereby reducing the time lag of the model, and improving the prediction precision and the prediction timeliness.

Drawings

Fig. 1 is a technical route diagram of the present invention.

Detailed Description

In the following, embodiments of the present invention are described in detail in a stepwise manner, which is merely a preferred embodiment of the present invention and is not intended to limit the scope of the present invention, and any modifications, equivalents, improvements, etc. made within the spirit and principle of the present invention should be included in the scope of the present invention.

In the description of the present invention, it should be noted that the terms "upper", "lower", "left", "right", "top", "bottom", "inner", "outer", and the like indicate orientations and positional relationships based on the orientations and positional relationships shown in the drawings, and are only used for describing the present invention and simplifying the description, but do not indicate or imply that the referred device or element must have a specific orientation and a specific orientation configuration and operation, and thus, the present invention is not to be construed as being limited thereto.

Please refer to fig. 1:

a wave forecasting algorithm integrating random search and mixed decomposition error correction comprises the following steps: step 1, acquiring buoy historical observation data based on buoy observation, integrating and processing ocean and meteorological data closely related to sea wave change, and constructing a sea wave forecasting database; step 2, sensitivity factor analysis is carried out, and input parameters and step length of the sea wave forecasting model are determined; (ii) a Step 3, constructing a sea wave forecasting model; and 4, acquiring a sea wave prediction error time sequence to obtain the corrected predicted sea wave height.

In the step 1, wind speed, wind direction, air temperature, air pressure, temperature, humidity, wave height, wave direction and wave period are used as input parameters to be evaluated of the sea wave height forecasting model

The step 2 comprises the following specific steps: removing the data abnormal value by using a Hampel filter, and defining an outlier as an element which is within a window length specified by the window and has a difference with a local median by more than three times and locally converted MAD; filling missing values by applying cubic polynomial interpolation, wherein the principle is an iterative algorithm for gradually approximating the minimum point of the function to be solved by the minimum point of a cubic curve so as to solve the approximate minimum point of the function to be solved; performing correlation analysis on related elements of the sea waves by adopting a Pearson correlation coefficient, and normalizing input parameters with different orders of magnitude; determining the optimal input characteristics and the corresponding step length through an input step length test;

the calculation formula of the Pearson correlation coefficient is as follows:

in formula (1): n is the sequence length, o _i Is the wave height sequence value and is the wave height sequence value,

is the average value of the wave height sequence, u _i In order to input the parameters to be evaluated,

the average value of the input variables to be evaluated;

the calculation formula for normalizing the input parameters with different orders of magnitude is as follows:

Data _nor ＝(Data _nor,max -Data _nor,min )·(Data-Data _min )/(Data _max -Data _min )+Data _nor,min (2)

in formula (2): data _nor For the sequences after normalization, Data _nor，max To normalize the upper bound of the interval, Data _nor，min For lower limit of normalization interval, Data is the sequence before normalization _max To normalize the maximum of the pre-sequence, Data _min Is the minimum value before normalization.

In the step 2, the step of determining the wave height forecasting model input parameters is as follows:

(a) according to the correlation coefficient result, selecting a parameter with highest correlation as a variable to be evaluated, selecting 1-k advanced step length variables to be evaluated and the effective wave height at the past moment as input, and inputting the input into a wave height forecasting model;

(b) training a deep learning model, and evaluating a forecast result;

(c) if the model precision is improved after k variables to be evaluated are added, the variables to be evaluated are reserved, and the precision is reduced until k +1 step lengths are input;

(d) reserving the last optimized variable, and adding a new variable with the highest correlation in the remaining variables as a variable to be evaluated;

(e) repeating the steps (a) to (d) until all the variables to be evaluated are tested;

(f) and obtaining the input parameters of the wave height forecasting model according to the test result tested in the step.

The step 3 comprises the following specific steps: and (3) adopting a long-short term memory neural network as a forecasting model, dividing a training set and a verification set, optimizing the hyper-parameters of the sea wave forecasting model by adopting a random search algorithm, and outputting a forecasting result.

In the step 3, the long-term and short-term memory neural network is used as a variant of the cyclic neural network, and the problems of gradient disappearance and gradient explosion in the cyclic neural network are solved through the unique gate-type unit structure; the gate represents a way to filter information into the memory unit, through the input gate (i) _t ) Forget gate (f) _t ) Output gate (o) _t ) The memory unit is used for controlling information to selectively enter the memory unit for storage or forgetting; each gate performs control output through an activation function Sigmoid; the Sigmoid function is defined as:

i _t ＝σ(W _i ·[S _t-1 ，x _t ]+b _i ) (4)

f _t ＝σ(W _f ·[s _t-1 ，x _t ]+b _f ) (5)

o _t ＝σ(W _o ·[S _t-1 ，x _t ]+b _o ) (6)

in formulae (3) to (6), S _t-1 Representing the output at time t-1, x _t Representing the input at the current time t, sigma represents the Sigmoid activation function, W _f 、W _i 、W _o Respectively, the weight parameters of the forgetting gate, the input gate and the output gate, b _f 、b _i 、b _o The offset coefficients of the forgetting gate, the input gate and the output gate are respectively.

In the step 3, a random search algorithm is adopted to perform hyper-parameter adjustment to optimize the model structure including the number of neural network layers and the number of neurons, and the search strategy is as follows:

(a) defining a search space;

(b) for the hyper-parameter with the search range being an interval, randomly sampling according to the given interval; for a hyper-parameter whose search range is a limited number of lists, equal probability sampling is performed in a given list;

(c) traversing the niter group sampling results obtained in the step (b); if the given search ranges are all lists, the samples are not put back for niter times;

(d) and comparing the values of the target functions of the points meeting the constraint conditions one by one, discarding the combination with large error, reserving the combination with small error, and finally obtaining the approximate solution of the optimal solution.

The step 4 comprises the following specific steps: obtaining a sea wave prediction error time sequence, decomposing the error time sequence based on improved adaptive noise ensemble empirical mode decomposition (ICEEMDAN), optimizing a mixed error correction model structure by adopting a random search algorithm, outputting a prediction error after training and verification, and adding the prediction error and an initial prediction wave height to obtain the corrected prediction sea wave height.

In the step 4, the neural network prediction value is set as

The measured data at the corresponding time is

Prediction error sequence E _i Expressed as:

in formula (7):

is a predicted value of the wave height prediction model,

for measured data at the corresponding time, E _i Is a prediction error;

carrying out pure randomness inspection, namely white noise inspection on the error sequence to ensure that the error time sequence has a self rule; checking by using an autocorrelation coefficient, and if the sequence is white noise, approximating the autocorrelation coefficient of the delay non-zero period of the sequence to be subject to the average value of 0; forTime series { X _t In terms of x _t And x _t-k The autocorrelation coefficient between is defined as:

in the formula (8), Cov represents covariance, and Var represents variance. When k is 0, ρ ₀ ＝1；X _t And X _t-k The meaning of the autocorrelation coefficients of (a) is to give the correlation of the function itself.

In the step 4, the improved adaptive noise ensemble empirical mode decomposition (ICEEMDAN) is an improved algorithm of the Empirical Mode Decomposition (EMD). EMD was proposed in 1998 by Huang et al as an algorithm that performs adaptive decomposition from data. The method does not need to manually divide the number of decomposition layers, but adopts self-adaptive decomposition, thereby facilitating the research of scholars, and the precision of the method is also verified by the experiments of prediction in the fields of weather and the like. Through EMD decomposition, raw data can be decomposed into a number of Intrinsic Mode Functions (IMFs) and a residual term. The decomposition steps are as follows:

(a) solving upper and lower envelope lines according to upper and lower extreme points of the original sequence;

(b) obtaining the mean value envelope curve by solving the mean value of the upper envelope curve and the lower envelope curve;

(c) subtracting the mean envelope curve from the original data to obtain a middle signal;

(d) judging whether the intermediate signal meets 1) whether the number of the extreme points and the number of the zero crossing points are equal or the phase difference is not more than one at most in the whole data section; 2) at any time, the average value of an upper envelope formed by the local maximum value points and a lower envelope formed by the local minimum value points is zero, namely the upper envelope and the lower envelope are locally symmetrical relative to a time axis;

(e) if the two conditions in (d) are not satisfied, repeating (a) to (d) based on the intermediate signal; if the condition is met, the signal is regarded as an IMF component, the IMF component is subtracted from the original signal to serve as a new original sequence, and the steps (a) to (d) are repeated to obtain IMFs 2 and … until the remaining terms contain less than 2 extreme values and further decomposition is not needed;

let x be the original signal, E _k (. h) is the kth IMF component obtained after EMD decomposition of the signal, w (i) is unit variance white noise with a mean value of zero, and M (. cndot.) is the local mean value of the signal; the calculation steps for improving adaptive noise ensemble empirical mode decomposition (icemdan) are as follows:

(a) performing EMD operation on the L mixed original signals with the white noise superimposed:

x ⁱ ＝x+β ₀ E ₁ (w(i))，i＝1，...L (9)

β ₀ ＝λ ₀ std(x)/std(E ₁ (w(i))) (10)

in the formula (10), std (. cndot.) represents the calculation standard deviation, and λ is recommended ₀ A value of 0.2;

the first remainder calculation formula is:

in formula (11): r represents residual components, IMF represents a natural mode function, and n represents the number of IMFs;

(b) the modal component 1 IMF may be expressed as follows:

IMF ₁ (t)＝x-r ₁ (t) (12)

in formula (12): x is an original signal;

(c) calculate the kth remainder:

in formula (13): r is _k (t) is the kth remainder;

(d) calculate the kth IMF component:

IMF _k (t)＝r _k-1 (t)-r _k (t)，k＝2，3，... (14)

(e) repeating (c) and (d) until the remainder does not require further decomposition, the decomposition result of the original sequence being expressed as follows:

in formula (15): x (t) is the original signal, IMF is the modal component, r _n (t) is the remainder.

Claims (10)

1. A sea wave forecasting algorithm integrating random search and mixed decomposition error correction is characterized in that: the method comprises the following steps: step 1, acquiring buoy historical observation data based on buoy observation, integrating and processing ocean and meteorological data closely related to sea wave change, and constructing a sea wave forecasting database; step 2, sensitivity factor analysis is carried out, and input parameters and step length of the sea wave forecasting model are determined; step 3, constructing a sea wave forecasting model; and 4, acquiring a sea wave prediction error time sequence, and adding the prediction error and the initial prediction wave height to obtain the corrected prediction sea wave height.

2. A hybrid stochastic search and hybrid decomposition error corrected ocean wave prediction algorithm as claimed in claim 1 wherein: in the step 1, wind speed, wind direction, air temperature, air pressure, temperature, humidity, wave height, wave direction and wave period are used as input parameters to be evaluated of the sea wave height forecasting model.

3. A hybrid stochastic search and hybrid decomposition error corrected ocean wave prediction algorithm as claimed in claim 2 wherein: the step 2 comprises the following specific steps: removing the data abnormal value by using a Hampel filter, and defining an outlier as an element which is within a window length specified by the window and has a difference with a local median by more than three times and locally converted MAD; filling missing values by using a cubic polynomial interpolation value, performing relevance analysis on sea wave related elements by using a Pearson correlation coefficient, and normalizing input parameters with different orders of magnitude; determining the optimal input characteristic and the corresponding step length through an input step length test;

the calculation formula of the Pearson correlation coefficient is as follows:

in formula (1): n is the sequence length, o _i Is the wave height sequence value and is the wave height sequence value,

is the average value of the wave height sequence, u _i In order to input the parameters to be evaluated,

the average value of the input variables to be evaluated;

the calculation formula for normalizing the input parameters with different orders of magnitude is as follows:

Data _nor ＝(Data _nor,max -Data _nor,min )·(Data-Data _min )/(Data _max -Data _min )+Data _nor,min (2)

in formula (2): data _nor For the sequences after normalization, Data _nor，max To normalize the interval upper bound, Data _nor，min For lower limit of normalization interval, Data is the sequence before normalization _max To normalize the maximum of the pre-sequence, Data _min Is the minimum value before normalization.

4. A hybrid stochastic search and hybrid decomposition error corrected ocean wave prediction algorithm as claimed in claim 3 wherein: in the step 2, the step of determining the wave height forecasting model input parameters is as follows:

(a) according to the correlation coefficient result, selecting a parameter with highest correlation as a variable to be evaluated, selecting 1-k advanced step length variables to be evaluated and the effective wave height at the past moment as input, and inputting the input into a wave height forecasting model;

(b) training a deep learning model, and evaluating a forecast result;

(c) if the model precision is improved after k variables to be evaluated are added, the variables to be evaluated are reserved, and the precision is reduced until k +1 step lengths are input;

(d) reserving the last optimized variable, and adding a new variable with the highest correlation in the remaining variables as a variable to be evaluated;

(e) repeating the steps (a) to (d) until all the variables to be evaluated are tested;

(f) and obtaining the input parameters of the wave height forecasting model according to the test result tested in the step.

5. A hybrid stochastic search and hybrid decomposition error corrected ocean wave prediction algorithm as claimed in claim 4 wherein: the step 3 comprises the following specific steps: and (3) adopting a long-short term memory neural network as a forecasting model, dividing a training set and a verification set, optimizing the hyper-parameters of the sea wave forecasting model by adopting a random search algorithm, and outputting a forecasting result.

6. A hybrid stochastic search and hybrid decomposition error corrected ocean wave prediction algorithm as claimed in claim 5 wherein: in the step 3, the long and short term memory neural network solves the problems of gradient disappearance and gradient explosion through a gate type unit structure and passes through an input gate (i) _t ) Forget gate (f) _t ) Output gate (o) _t ) The memory unit is used for controlling information to selectively enter the memory unit for storage or forgetting; each gate performs control output through an activation function Sigmoid;

the Sigmoid function is defined as:

i _t ＝σ(W _i ·[S _t-1 ，x _t ]+b _i ) (4)

f _t ＝σ(W _f ·[S _r-1 ，x _t ]+b _f ) (5)

o _t ＝σ(W _o ·[s _t-1 ，x _t ]+b _o ) (6)

in formulae (3) to (6), S _t-1 Representing the output at time t-1, x _t Representing the input at the current time t, sigma represents the Sigmoid activation function, W _f 、W _i 、W _o Respectively, the weight parameters of the forgetting gate, the input gate and the output gate, b _f 、b _i 、b _o The offset coefficients of the forgetting gate, the input gate and the output gate are respectively.

7. A hybrid stochastic search and hybrid decomposition error corrected ocean wave prediction algorithm as claimed in claim 6 wherein: in the step 3, a random search algorithm is adopted to perform hyper-parameter adjustment to optimize the model structure including the number of neural network layers and the number of neurons, and the search strategy is as follows:

(a) defining a search space;

(b) for the hyper-parameter with the search range being an interval, randomly sampling according to the given interval; for a hyper-parameter whose search range is a limited number of lists, equal probability sampling is performed in a given list;

(c) traversing the niter group sampling results obtained in the step (b); if the given search ranges are all lists, the samples are not put back for niter times;

(d) and comparing the values of the target functions of the points meeting the constraint conditions one by one, discarding the combination with large error, reserving the combination with small error, and finally obtaining the approximate solution of the optimal solution.

8. A hybrid stochastic search and hybrid decomposition error corrected ocean wave prediction algorithm as claimed in claim 7 wherein: the step 4 comprises the following specific steps: obtaining a sea wave forecast error time sequence, decomposing the error time sequence based on improved adaptive noise set empirical mode decomposition, optimizing a mixed error correction model structure by adopting a random search algorithm, outputting a forecast error after training and verification, and adding the forecast error and an initial forecast wave height to obtain a corrected forecast sea wave height.

9. A hybrid stochastic search and hybrid decomposition error corrected ocean wave prediction algorithm as claimed in claim 8 wherein: in the step 4, the neural network prediction value is set as

The measured data at the corresponding time is

Prediction error sequence E _i Expressed as:

in formula (7):

is a predicted value of the wave height prediction model,

for measured data at the corresponding time, E _i Is a prediction error;

carrying out pure randomness inspection, namely white noise inspection on the error sequence to ensure that the error time sequence has a self rule; the autocorrelation coefficients are used for checking, if the sequence is white noise, the approximate mean value of the autocorrelation coefficients of the delay nonzero period of the sequence is 0; for time series { X _t In terms of x _t And x _t-k The autocorrelation coefficient between them is defined as:

in the formula (8), Cov represents covariance, and Var represents variance. When k is 0, ρ ₀ ＝1；X _t And X _t-k The meaning of the autocorrelation coefficients of (a) is to give the correlation of the function itself.

10. A hybrid stochastic search and hybrid decomposition error corrected ocean wave prediction algorithm as claimed in claim 9 wherein: in the step 4, the empirical mode decomposition of the improved adaptive noise set is an improved algorithm of empirical mode decomposition, and the decomposition step of empirical mode decomposition is as follows:

(a) solving upper and lower envelope lines according to upper and lower extreme points of the original sequence;

(b) obtaining the mean value envelope curve by solving the mean value of the upper envelope curve and the lower envelope curve;

(c) subtracting the mean envelope curve from the original data to obtain a middle signal;

(d) judging whether the intermediate signal meets 1) whether the number of the extreme points and the number of the zero crossing points are equal or the phase difference is not more than one at most in the whole data section; 2) at any time, the average value of an upper envelope formed by the local maximum value points and a lower envelope formed by the local minimum value points is zero, namely the upper envelope and the lower envelope are locally symmetrical relative to a time axis;

(e) if the two conditions in (d) are not satisfied, repeating (a) to (d) based on the intermediate signal; if the condition is met, the signal is regarded as an IMF component, the IMF component is subtracted from the original signal to serve as a new original sequence, and the steps (a) to (d) are repeated to obtain IMFs 2 and … until the remaining terms contain less than 2 extreme values and further decomposition is not needed;

let x be the original signal, E _k (. h) is the kth IMF component obtained after EMD decomposition of the signal, w (i) is unit variance white noise with a mean value of zero, and M (. cndot.) is the local mean value of the signal;

the calculation steps for improving the empirical mode decomposition of the adaptive noise set are as follows:

(a) performing EMD operation on the L mixed original signals with the white noise superimposed:

x ⁱ ＝x+β ₀ E ₁ (w(i))，i＝1，...，L (9)

β ₀ ＝λ ₀ std(x)/std(E ₁ (w(i))) (10)

in the formula (10), std (. cndot.) represents the calculation standard deviation, and λ is recommended ₀ A value of 0.2；

The first remainder calculation formula is:

in formula (11): r represents residual components, IMF represents a natural mode function, and n represents the number of IMFs;

(b) the modal component 1 IMF may be expressed as follows:

IMF ₁ (t)＝x-r ₁ (t) (12)

in formula (12): x is the original signal, r ₁ (t) a first remainder;

(c) calculate the kth remainder:

in formula (13): r is a radical of hydrogen _k (t) is the kth remainder;

(d) calculate the kth IMF component:

IMF _k (t)＝r _k-1 (t)-r _k (t)，k＝2，3，... (14)

(e) repeating (c) and (d) until the remainder does not require further decomposition, the decomposition result of the original sequence being expressed as follows:

in formula (15): x (t) is the original signal, IMF is the modal component, r _n (t) is the remainder.

Images