RBF network short-time storm surge prediction method and system based on LM back propagation optimization
The RBF network optimized by LM back-propagation solves the problems of traditional RBF networks in handling non-static data and insufficient peak capture in storm surge prediction, achieving higher prediction accuracy and robustness.
Patent Information
- Application Number
- CN202510972793.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-15
- Publication Date
- 2025-09-12
AI Technical Summary
Traditional RBF networks have difficulty in processing non-static data and lack the ability to capture peak values, resulting in a decrease in the accuracy of storm surge prediction.
The RBF network based on LM back-propagation optimization is adopted. The center parameters are determined by K-means algorithm. Back-propagation optimization is performed in combination with Gaussian function and trainlm function to construct a dynamic radial basis function network for preprocessing and feature extraction of storm surge correlation data.
The model's prediction capability for non-static time series data has been improved, its ability to capture storm surge peaks has been enhanced, and its prediction accuracy and generalization capability have been improved.
Smart Images
Figure CN120632371A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of storm surge prediction, and in particular relates to a RBF network short-term storm surge prediction method and system based on LM back propagation optimization. Background Art
[0002] Storm surges are abnormal rises and falls in nearshore sea levels caused by strong atmospheric disturbances (such as typhoons and extratropical cyclones). Predicting these surges is crucial for disaster prevention and mitigation in coastal areas and for port operations. With the intensification of global climate change and the frequent occurrence of extreme weather events, the need for accurate prediction of short-term storm surges (typically those occurring within the next six or several hours) is becoming increasingly urgent. Currently, three main methods for predicting storm surge surges exist: empirical analysis, numerical calculation, and machine learning. However, these methods all have limitations. Empirical analysis relies on empirical evidence from historical disasters and subjective judgments of current meteorological conditions. While practical in localized regions or under specific weather conditions, its results are significantly affected by human factors and cannot meet the objectivity and repeatability requirements of modern, refined forecasting. Numerical calculation methods, by developing coupled ocean-atmosphere models and solving for surges based on physical laws such as fluid mechanics and continuity equations, offer theoretically high accuracy, but are sensitive to initial field errors and require significant computational resources, making efficient real-time computation difficult for short-term nowcasting.
[0003] In recent years, machine learning methods have become a research hotspot for storm surge prediction due to their powerful nonlinear fitting capabilities. Among them, radial basis function (RBF) neural networks, a typical feedforward neural network, have been widely used in storm surge prediction due to their simple structure, fast training speed, and outstanding local approximation properties.
[0004] However, the traditional RBF neural network has the following shortcomings in practical applications:
[0005] Difficulty processing non-static data: After training, traditional RBF networks have a fixed network structure (such as the number of neurons and the location of the center point), making them difficult to adapt to the non-static nature of data distribution, which changes over time. For storm surge forecasting, meteorological data often exhibits dynamic characteristics such as seasonality and interannual variations. Fixed-structured RBF networks struggle to effectively capture these variations, resulting in reduced prediction accuracy.
[0006] Inadequate peak capture capability: RBF networks use central points and extended parameters to locally fit data, and their responsiveness depends on the distribution and width of the central points. During storm surges, extreme peaks often exhibit sudden and nonlinear characteristics. Traditional RBF networks, with their fixed central points and non-dynamic parameter adjustments, struggle to effectively model extreme peak regions, resulting in significant peak prediction errors and impacting early warning effectiveness.
[0007] To address the above problems, we proposed a RBF network short-term storm surge prediction method and system based on LM back-propagation optimization. Summary of the Invention
[0008] The purpose of the present invention is to address the deficiencies of the existing technology and provide a RBF network short-term storm surge prediction method and system based on LM back propagation optimization, which solves the problems that traditional RBF networks are difficult to process non-static data and have insufficient peak capture ability.
[0009] The present invention is achieved by: a RBF network short-term storm surge prediction method based on LM back propagation optimization, the RBF network short-term storm surge prediction method based on LM back propagation optimization comprising:
[0010] S10, loading storm surge related data, preprocessing the storm surge related data, and outputting the preprocessed storm surge related data;
[0011] S20, RBF network initialization, using the K-means algorithm to determine the central parameters of the RBF network, using the Gaussian function as the radial basis function, calculating the hidden layer output, and obtaining the preliminary prediction results through linear combination;
[0012] S30, use the trainlm function to perform back propagation optimization on the RBF network, define the loss function of the RBF network, and output the converged RBF network;
[0013] S40, obtaining the pre-processed reduced-dimensionality feature data, using the reduced-dimensionality feature data as input, executing the RBF network, and outputting a storm surge short-term water increase prediction result.
[0014] Preferably, the time resolution of the storm surge data, meteorological data, and typhoon factor data is one hour, wherein the meteorological data includes air pressure, wind speed, and wind direction; the typhoon factor data includes the maximum wind speed at the typhoon center, the minimum air pressure, the invasion angle between the typhoon center and the tide station, and the shortest distance between the typhoon center and the tide station.
[0015] Preferably, the method for preprocessing storm surge related data includes:
[0016] S101, load storm surge related data, identify storm surge related data, and build data feature set M'i =[m' i1 、m' i2 、…、m' i8 、m' i9 、m' i10 ];
[0017] S102, normalizing the storm surge related data in the data feature set and outputting normalized feature data;
[0018] S103 , loading the normalized feature data, performing principal component analysis on the normalized feature data, retaining the principal component with 95% variance, and outputting the reduced-dimensional feature data after removing redundant information.
[0019] Preferably, construct a data feature set M' i =[m' i1 、m' i2 、…、m' i8 、m' i9 、m' i10 ], the data feature set M' i It contains ten types of characteristic data, including wind speed, wind direction, station air pressure, straight-line distance between the typhoon center and the monitoring station, intrusion angle, maximum wind speed at the typhoon center, minimum air pressure at the typhoon center, pressure gradient, wind speed to distance ratio, and storm surge water level increase in a short period of time.
[0020] Preferably, when normalizing the storm surge related data in the data feature set, the normalized feature data m 01 The normalized formula is expressed as:
[0021]
[0022] Among them, m 01 Represents the normalized feature data, max(m 01 )、min(m 01 ) represent the storm surge related data m in the data feature set 01 (i) The maximum and minimum values.
[0023] Preferably, when principal component analysis is performed on the normalized feature data, the projection direction with the largest variance of the normalized feature data is found based on the covariance matrix, where the covariance matrix is defined as:
[0024]
[0025] Assume that the normalized feature data matrix is M∈R n×d , n represents the data sample size, the feature dimension is d, C represents the symmetric semi-positive matrix, C ii Represents the diagonal elements, the variance of the i-th feature, and the off-diagonal elements Cij represents the covariance of feature i and feature j;
[0026] Perform eigendecomposition on the covariance matrix:
[0027] C=VAV T
[0028] Where V=[v1,v2,…,v d ] is the orthogonal eigenvector matrix, v i are unit vectors and are orthogonal to each other, A=diag(λ1,λ2,…,λ d ) is a diagonal matrix, λ1≥λ2≥…≥λ d ≥0 is the eigenvalue, representing the variance in the direction of each principal component;
[0029]
[0030] Here, k is the number of features remaining after principal component analysis selection.
[0031] Preferably, the radial basis function formula is as follows:
[0032]
[0033] Among them, c m is the center parameter, b m is the expansion parameter, m is the number of neural units;
[0034] The preliminary prediction results are obtained through linear combination. The preliminary prediction results are calculated as follows:
[0035]
[0036] Preferably, the loss function of the RBF network is a mean square error loss function, which calculates the mean square error by comparing the predicted result with the true value. The mean square error loss function is expressed as:
[0037]
[0038] Where θ=[w T ,c,b] T Contains adjustable parameters, is the error vector;
[0039] When using the trainlm function to perform back propagation optimization on the RBF network, the iterative formula is expressed as:
[0040] θ (k+1) =θ (k) -(J T J+μI) -1 J T e
[0041] Where J is the Jacobian matrix, μ>0 is the damping factor, and I is the unit matrix;
[0042] The actual parameter update formula is:
[0043] Δθ (k) =γΔθ (k-1) -(1-γ)(J T J+μI) -1 J T e
[0044] Among them, γ is the momentum coefficient, which reduces oscillation by accumulating historical update directions.
[0045] On the other hand, the present invention also provides an RBF network short-term storm surge prediction system based on LM back propagation optimization, the system comprising:
[0046] A preprocessing module is used to load storm surge related data, preprocess the storm surge related data, and output the preprocessed storm surge related data;
[0047] The preliminary prediction module initializes the RBF network, uses the K-means algorithm to determine the central parameters of the RBF network, selects the Gaussian function as the radial basis function, calculates the hidden layer output, and obtains the preliminary prediction results through linear combination;
[0048] The water increase prediction module uses the trainlm function to perform back propagation optimization on the RBF network, defines the loss function of the RBF network, outputs the converged RBF network, obtains the preprocessed dimensionality reduction feature data, uses the dimensionality reduction feature data as input, executes the RBF network, and outputs the storm surge short-term water increase prediction results.
[0049] Compared with the prior art, the embodiments of the present application have the following beneficial effects:
[0050] In an embodiment of the present invention, a hybrid RBF network for non-static time series data prediction is constructed by combining a dynamic radial basis function network. The RBF network combines multiple features, multiple time series information, reasonable preprocessing, and optimized training strategies to improve the predictive ability and robustness of the model. The RBF network uses the trainlm training function to perform backpropagation optimization on the network parameters, thereby optimizing the problem of traditional RBF networks' insufficient capture of storm surge peaks. Compared with traditional RBF models, it can better cope with complex meteorological time series data, has higher accuracy and generalization ability, and performs better in predicting short-term storm surge water increase.
[0051] In an embodiment of the present invention, a method for preprocessing storm surge-related data uses normalization to uniformly map storm surge-related data of different dimensions and value ranges to the same interval, eliminating model bias caused by large numerical differences between features and improving the stability and convergence speed of model training. The model's ability to express nonlinear relationships is enhanced, and derived features such as pressure gradient and wind speed-to-distance ratio are introduced during the preprocessing process. These features can more accurately reflect typhoon intensity, path changes, and their impact on tide stations, thereby enhancing the model's ability to capture the nonlinear dynamic characteristics of storm surge water increase. BRIEF DESCRIPTION OF THE DRAWINGS
[0052] Figure 1 This is an architecture diagram of the RBF network short-term storm surge prediction system based on LM back propagation optimization provided by the present invention.
[0053] Figure 2 The training set prediction results of the RBF network short-term storm surge prediction method based on LM back propagation optimization provided by an embodiment of the present invention are shown.
[0054] Figure 3 The test set test results of the RBF network short-term storm surge prediction method based on LM back propagation optimization provided by an embodiment of the present invention are shown.
[0055] Figure 4 The figure shows the prediction results of the traditional RBF training set.
[0056] Figure 5 The figure shows the prediction results of the traditional RBF test set. DETAILED DESCRIPTION
[0057] Unless otherwise defined, all technical and scientific terms used herein have the same meanings as commonly understood by those skilled in the art to which this application belongs. The terms used in the specification of the application are only for the purpose of describing specific embodiments and are not intended to limit this application. The terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of this application or the above-mentioned drawings are used to distinguish different objects, not to describe a specific order.
[0058] In view of the problems that traditional RBF networks have difficulty in processing non-static data and insufficient peak capture capabilities, we propose a RBF network short-term storm surge prediction method and system based on LM back-propagation optimization. In short, when implementing the method, the storm surge related data is pre-processed, the RBF network is initialized, the center parameters of the RBF network are determined using the K-means algorithm, the RBF network is back-propagated and optimized using the trainlm function, and the loss function of the RBF network is defined. The converged RBF network is output, and the dimensionality-reduced feature data is used as input to execute the RBF network and output the storm surge short-term water increase prediction results. In an embodiment of the present invention, a hybrid RBF network for non-static time series data prediction is constructed by combining a dynamic radial basis function network. The RBF network combines multiple features, multiple time series information, reasonable pre-processing and optimized training strategies, thereby improving the prediction ability and robustness of the model. The RBF network uses the trainlm training function to back-propagate and optimize the network parameters, thereby optimizing the problem of the traditional RBF network's insufficient capture of storm surge peaks. Compared with the traditional RBF model, it can better cope with complex meteorological time series data, has higher accuracy and generalization ability, and performs better in predicting short-term storm surge water increase.
[0059] Example 1
[0060] The embodiment of the present invention provides a RBF network short-term storm surge prediction method based on LM back propagation optimization, and the RBF network short-term storm surge prediction method based on LM back propagation optimization specifically includes:
[0061] S10, loading storm surge related data, preprocessing the storm surge related data, and outputting the preprocessed storm surge related data;
[0062] It should be noted that the time resolution of the storm surge data, meteorological data, and typhoon factor data is one hour. Among them, the meteorological data includes air pressure, wind speed, and wind direction. The typhoon factor data includes the maximum wind speed of the typhoon center, the lowest air pressure, the invasion angle between the typhoon center and the tide station, and the shortest distance between the typhoon center and the tide station.
[0063] In an embodiment of the present invention, the method for preprocessing storm surge related data includes:
[0064] S101, load storm surge related data, identify storm surge related data, and build data feature set M' i =[m' i1 、m' i2 、…、m' i8 、m' i9 、m' i10 ];
[0065] Among them, construct the data feature set M'i =[m' i1 、m' i2 、…、m' i8 、m' i9 、m' i10 ], the data feature set is the collection of meteorological data, typhoon factor data, and storm surge water increase data in the i-th 24-hour period, and the data feature set M' i It contains ten types of characteristic data, including wind speed, wind direction, station air pressure, straight-line distance between the typhoon center and the monitoring station, intrusion angle, maximum wind speed in the typhoon center, minimum air pressure in the typhoon center, pressure gradient, wind speed to distance ratio, and storm surge water increase in a short period of time. Among them, the pressure gradient and wind speed to distance ratio are newly created features. The pressure gradient reflects the rate of change of air pressure over time, and the wind speed / distance ratio reflects the intensity and influence of the typhoon.
[0066] S102, normalizing the storm surge related data in the data feature set and outputting normalized feature data;
[0067] When normalizing the storm surge related data in the data feature set, the normalized feature data m 01 The normalized formula is expressed as:
[0068]
[0069] Among them, the wind speed data m during the test period is 01 For example, m 01 Represents the normalized feature data, max(m 01 )、min(m 01 ) represent the storm surge related data m in the data feature set 01 (i) The maximum and minimum values.
[0070] S103 , loading the normalized feature data, performing principal component analysis on the normalized feature data, retaining the principal component with 95% variance, and outputting the reduced-dimensional feature data after removing redundant information.
[0071] In an embodiment of the present invention, a method for preprocessing storm surge-related data uses normalization to uniformly map storm surge-related data of different dimensions and value ranges to the same interval, eliminating model bias caused by large numerical differences between features and improving the stability and convergence speed of model training. The model's ability to express nonlinear relationships is enhanced, and derived features such as pressure gradient and wind speed-to-distance ratio are introduced during the preprocessing process. These features can more accurately reflect typhoon intensity, path changes, and their impact on tide stations, thereby enhancing the model's ability to capture the nonlinear dynamic characteristics of storm surge water increase.
[0072] In the embodiment of the present invention, principal component analysis is performed on the normalized feature data to retain the principal components that can explain 95% of the variance, reduce the number of features, remove redundant information, reduce model complexity, and thus improve model performance.
[0073] When performing principal component analysis on the normalized feature data, the projection direction with the largest variance of the normalized feature data is found based on the covariance matrix, where the covariance matrix is defined as:
[0074]
[0075] Assume that the normalized feature data matrix is M∈R n×d , n represents the data sample size, the feature dimension is d, C represents the symmetric semi-positive matrix, C ii Represents the diagonal elements, the variance of the i-th feature, and the off-diagonal elements C ij represents the covariance of feature i and feature j;
[0076] Perform eigendecomposition on the covariance matrix:
[0077] C=VAV T
[0078] Where V=[v1,v2,…,v d ] is the orthogonal eigenvector matrix, v i are unit vectors and are orthogonal to each other, A=diag(λ1,λ2,…,λ d ) is a diagonal matrix, λ1≥λ2≥…≥λ d ≥0 is the eigenvalue, representing the variance in the direction of each principal component;
[0079]
[0080] Here, k is the number of features remaining after principal component analysis selection.
[0081] S20, RBF network initialization, using the K-means algorithm to determine the central parameters of the RBF network, using the Gaussian function as the radial basis function, calculating the hidden layer output, and obtaining the preliminary prediction results through linear combination;
[0082] In the embodiment of the present invention, the radial basis function formula is as follows:
[0083]
[0084] Among them, c m is the center parameter, b m is the expansion parameter, m is the number of neural units;
[0085] The preliminary prediction results are obtained through linear combination. The preliminary prediction results are calculated as follows:
[0086]
[0087] S30, use the trainlm function to perform back propagation optimization on the RBF network, define the loss function of the RBF network, and output the converged RBF network;
[0088] It should be noted that the loss function of the RBF network is the mean square error loss function. The mean square error loss function calculates the mean square error between the predicted result and the true value. The mean square error loss function is expressed as:
[0089]
[0090] Where θ=[w T ,c,b] T Contains adjustable parameters, is the error vector;
[0091] When using the trainlm function to perform back propagation optimization on the RBF network, the trainlm algorithm combines the advantages of the gradient descent method and the Gauss-Newton method. The iterative formula is expressed as:
[0092] θ (k+1) =θ (k) -(J T J+μI) -1 J T e
[0093] Where J is the Jacobian matrix, μ>0 is the damping factor (used to control the step direction), and I is the identity matrix (used to ensure the matrix is reversible);
[0094] The actual parameter update formula can be modified as follows:
[0095] Δθ (k) =γΔθ (k-1) -(1-γ)(J T J+μI) -1 J T e
[0096] Among them, γ is the momentum coefficient, which reduces oscillation by accumulating historical update directions
[0097] S40, obtaining the pre-processed reduced-dimensionality feature data, using the reduced-dimensionality feature data as input, executing the RBF network, and outputting a storm surge short-term water increase prediction result.
[0098] In an embodiment of the present invention, a hybrid RBF network for non-static time series data prediction is constructed by combining a dynamic radial basis function network. The RBF network combines multiple features, multiple time series information, reasonable preprocessing, and optimized training strategies to improve the predictive ability and robustness of the model. The RBF network uses the trainlm training function to perform backpropagation optimization on the network parameters, thereby optimizing the problem of traditional RBF networks' insufficient capture of storm surge peaks. Compared with traditional RBF models, it can better cope with complex meteorological time series data, has higher accuracy and generalization ability, and performs better in predicting short-term storm surge water increase.
[0099] Example 2
[0100] The embodiment of the present invention provides a RBF network short-term storm surge prediction system based on LM back propagation optimization. Figure 1 The following figure shows the architecture of the RBF network short-term storm surge prediction system based on LM back propagation optimization. The RBF network short-term storm surge prediction system based on LM back propagation optimization specifically includes:
[0101] The preprocessing module 100 is used to load storm surge related data, preprocess the storm surge related data, and output the preprocessed storm surge related data;
[0102] The preliminary prediction module 200 initializes the RBF network, uses the K-means algorithm to determine the central parameters of the RBF network, selects the Gaussian function as the radial basis function, calculates the hidden layer output, and obtains the preliminary prediction result through linear combination;
[0103] The water increase prediction module 300 uses the trainlm function to perform back propagation optimization on the RBF network, defines the loss function of the RBF network, outputs the converged RBF network, and obtains the pre-processed dimensionality reduction feature data. With the dimensionality reduction feature data as input, the RBF network is executed to output the short-term storm surge water increase prediction result.
[0104] It should be noted that the preprocessing module 100, the preliminary prediction module 200, and the water increase prediction module 300 can be connected by Bluetooth, DTU or 5G communication. The RBF network short-time storm surge prediction system based on LM back propagation optimization in the embodiment of the present invention corresponds to the above-mentioned RBF network short-time storm surge prediction method based on LM back propagation optimization, and will not be repeated here.
[0105] Example 3
[0106] In this embodiment of the present invention, storm surge data from 14 typhoons (approximately 14 days) collected from observation stations are collected. Table 1 shows the typhoon storm surge data. The 14 typhoons are randomly arranged, and the first 13 typhoons are used as a training set. The peak value of the last typhoon is used as a test set. A model (RBF network) is constructed to predict the future height of a single storm surge. Gaussian radial basis functions are used to process time series data. It should be noted that in this embodiment, the model refers to an RBF network model optimized based on LM backpropagation.
[0107] Table 1
[0108]
[0109] In this embodiment, the factors input to the model are wind speed, wind direction, site air pressure, the straight-line distance between the typhoon center and the monitoring station, the intrusion angle, the maximum air pressure at the typhoon center, and the maximum wind speed at the typhoon center, and the target prediction factor is the water surge height. After data preprocessing, a 3-step delay step is set. The number of neurons in the RBF model is set to 100, the initial expansion parameter is 2, and the training target error is set to 0.01. First, build the RBF network, and then use the trainlm function for reverse fine-tuning. Among them, the maximum number of training times of the trainlm function is set to 8 times, the training target error is 0.001, the learning rate is 0.01, and the momentum parameter is set to 0.9. The experimental results are expressed using root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R 2 ) and correlation coefficient (CC).
[0110] Among them, the RMSE formula is:
[0111]
[0112] The MAE formula is:
[0113]
[0114] R 2 The formula is:
[0115]
[0116] The CC formula is:
[0117]
[0118] By comparing the RBF network short-term storm surge prediction method based on LM back propagation optimization provided by the embodiment of the present invention with the traditional RBF method, the prediction accuracy can be significantly improved in the short-term storm surge water increase prediction. Figure 2 The training set prediction results of the RBF network short-term storm surge prediction method based on LM back propagation optimization provided by an embodiment of the present invention are shown. Figure 3 The test results of the test set of the RBF network short-term storm surge prediction method based on LM back propagation optimization provided by the embodiment of the present invention are shown. Figure 4 Shows the prediction results of the traditional RBF training set, Figure 5 The traditional RBF test set prediction results are shown.
[0119] Table 2 shows the comparison of the prediction results of the present method and the traditional RBF training set, and Table 3 shows the comparison of the prediction results of the present method and the traditional RBF test set.
[0120] Table 2 Comparison of prediction results of this method and traditional RBF training set
[0121]
[0122] Table 3 Comparison of prediction results of this method and traditional RBF test set
[0123]
[0124] From Tables 2 and 3, we can see that both from the perspective of model learning and fitting ability on the training set and from the perspective of model generalization and application ability on the test set, the RBF network short-term storm surge prediction method based on LM back propagation optimization has achieved a significant reduction in RMSE and MAE, two indicators that measure the size of the prediction error, and has a significant impact on R 2 Significant improvements were achieved in both ΔC and CC, two indicators that measure the quality of model fit and correlation. Therefore, this method has achieved a qualitative leap and significant improvement in the prediction accuracy of short-term storm surge water increase forecasting compared to the traditional RBF method.
[0125] In summary, the present invention provides a short-term storm surge prediction method and system based on RBF network optimized by LM back propagation. In an embodiment of the present invention, a hybrid RBF network for non-static time series data prediction is constructed in combination with a dynamic radial basis function network. The RBF network combines multiple features, multiple time series information, reasonable preprocessing and optimized training strategies to improve the prediction ability and robustness of the model, and the RBF network uses the trainlm training function to perform back propagation optimization on the network parameters, thereby optimizing the problem of insufficient capture of storm surge peaks by traditional RBF networks. Compared with traditional RBF models, it can better cope with complex meteorological time series data, has higher accuracy and generalization ability, and has better performance in short-term storm surge water increase prediction.
[0126] In an embodiment of the present invention, a method for preprocessing storm surge-related data uses normalization to uniformly map storm surge-related data of different dimensions and value ranges to the same interval, eliminating model bias caused by large numerical differences between features and improving the stability and convergence speed of model training. The model's ability to express nonlinear relationships is enhanced, and derived features such as pressure gradient and wind speed-to-distance ratio are introduced during the preprocessing process. These features can more accurately reflect typhoon intensity, path changes, and their impact on tide stations, thereby enhancing the model's ability to capture the nonlinear dynamic characteristics of storm surge water increase.
[0127] It should be noted that for the aforementioned embodiments, for simplicity of description, they are all expressed as a series of action combinations. However, those skilled in the art should be aware that the present invention is not limited by the order of the actions described, because according to the present invention, certain steps may be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present invention.
[0128] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the scope of protection of the invention. Obviously, the embodiments described are only some embodiments of the present invention, rather than all embodiments. Based on these embodiments, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in this field can still combine, add, delete or make other adjustments to the features in the various embodiments of the present invention according to the circumstances without conflict, without making creative work, so as to obtain different other technical solutions that do not deviate from the concept of the present invention in essence, and these technical solutions also fall within the scope of protection of the present invention.
Claims
1. The RBF network short-term storm surge prediction method based on LM back propagation optimization is characterized by: The method comprises: S10, loading storm surge related data, preprocessing the storm surge related data, and outputting the preprocessed storm surge related data; S20, RBF network initialization, using the K-means algorithm to determine the central parameters of the RBF network, using the Gaussian function as the radial basis function, calculating the hidden layer output, and obtaining the preliminary prediction results through linear combination; S30, use the trainlm function to perform back propagation optimization on the RBF network, define the loss function of the RBF network, and output the converged RBF network; S40, obtaining the pre-processed reduced-dimensionality feature data, using the reduced-dimensionality feature data as input, executing the RBF network, and outputting a storm surge short-term water increase prediction result.
2. The RBF network short-term storm surge prediction method based on LM back propagation optimization according to claim 1 is characterized in that: The time resolution of the storm surge data, meteorological data, and typhoon factor data is one hour. The meteorological data includes air pressure, wind speed, and wind direction. The typhoon factor data includes the maximum wind speed at the typhoon center, the minimum air pressure, the invasion angle between the typhoon center and the tide station, and the shortest distance between the typhoon center and the tide station.
3. The RBF network short-term storm surge prediction method based on LM back propagation optimization according to claim 1 is characterized in that: The method for preprocessing storm surge related data includes: S101, load storm surge related data, identify storm surge related data, and build data feature set M' i =[m' i1 、m' i2 、…、m' i8 、m' i9 、m' i10 ]; S102, normalizing the storm surge related data in the data feature set and outputting normalized feature data; S103 , loading the normalized feature data, performing principal component analysis on the normalized feature data, retaining the principal component with 95% variance, and outputting the reduced-dimensional feature data after removing redundant information.
4. The RBF network short-term storm surge prediction method based on LM back propagation optimization according to claim 3 is characterized in that: Construct data feature set M' i =[m' i1 、m' i2 、…、m' i8 、m' i9 、m' i10 ], the data feature set M' i It contains ten types of characteristic data, including wind speed, wind direction, station air pressure, straight-line distance between the typhoon center and the monitoring station, intrusion angle, maximum wind speed at the typhoon center, minimum air pressure at the typhoon center, pressure gradient, wind speed to distance ratio, and storm surge water level increase in a short period of time.
5. The RBF network short-term storm surge prediction method based on LM back propagation optimization according to claim 3 is characterized in that: When normalizing the storm surge related data in the data feature set, the normalized feature data m 01 The normalized formula is expressed as: Among them, m 01 Represents the normalized feature data, max(m 01 )、min(m 01 ) represent the storm surge related data m in the data feature set 01 (i) The maximum and minimum values.
6. The RBF network short-term storm surge prediction method based on LM back propagation optimization according to claim 5 is characterized in that: When performing principal component analysis on the normalized feature data, the projection direction with the largest variance of the normalized feature data is found based on the covariance matrix, where the covariance matrix is defined as: Assume that the normalized feature data matrix is M∈R n×d , n represents the data sample size, the feature dimension is d, C represents the symmetric semi-positive matrix, C ii Represents the diagonal elements, the variance of the i-th feature, and the off-diagonal elements C ij represents the covariance of feature i and feature j; Perform eigendecomposition on the covariance matrix: C=VAV T Where V=[v1,v2,…,v d ] is the orthogonal eigenvector matrix, v i are unit vectors and are orthogonal to each other, A=diag(λ1,λ2,…,λ d ) is a diagonal matrix, λ1≥λ2≥…≥λ d ≥0 is the eigenvalue, representing the variance in the direction of each principal component; Here, k is the number of features remaining after principal component analysis selection.
7. The RBF network short-term storm surge prediction method based on LM back propagation optimization according to claim 1, characterized in that: The radial basis function formula is as follows: Among them, c m is the center parameter, b m is the expansion parameter, m is the number of neural units; The preliminary prediction results are obtained through linear combination. The preliminary prediction results are calculated as follows:
8. The RBF network short-term storm surge prediction method based on LM back propagation optimization according to claim 7, characterized in that: The loss function of the RBF network is the mean square error loss function. The mean square error loss function calculates the mean square error between the predicted result and the true value. The mean square error loss function is expressed as: Where θ=[w T ,c,b] T Contains adjustable parameters, is the error vector; When using the trainlm function to perform back propagation optimization on the RBF network, the iterative formula is expressed as: i (k+1) =θ (k) -(J T J+μI) -1 J T e Where J is the Jacobian matrix, μ>0 is the damping factor, and I is the unit matrix; The actual parameter update formula is: Dth (k) =γΔθ (k-1) -(1-c)(J T J+μI) -1 J T e Among them, γ is the momentum coefficient, which reduces oscillation by accumulating historical update directions.
9. An RBF network short-term storm surge prediction system based on LM back-propagation optimization, used to implement the RBF network short-term storm surge prediction method based on LM back-propagation optimization according to any one of claims 1 to 8, characterized in that: The system comprises: A preprocessing module is used to load storm surge related data, preprocess the storm surge related data, and output the preprocessed storm surge related data; The preliminary prediction module initializes the RBF network, uses the K-means algorithm to determine the central parameters of the RBF network, selects the Gaussian function as the radial basis function, calculates the hidden layer output, and obtains the preliminary prediction results through linear combination; The water increase prediction module uses the trainlm function to perform back propagation optimization on the RBF network, defines the loss function of the RBF network, outputs the converged RBF network, obtains the preprocessed dimensionality reduction feature data, uses the dimensionality reduction feature data as input, executes the RBF network, and outputs the storm surge short-term water increase prediction results.