An evaporation waveguide height multi-step prediction method based on state anchoring gate residual

Abstract

The application discloses an evaporation duct height multi-step prediction method based on state anchoring gate residual, relates to the technical field of marine environment prediction, and comprises the following steps: obtaining and preprocessing historical evaporation duct height data and corresponding marine and atmospheric environmental variables to obtain a basic input sequence; setting a historical input window length as T, a prediction step as D, sliding and slicing the basic input sequence along a time axis, and constructing a historical input sequence and a corresponding future EDH prediction target; constructing multi-branch features of the historical input sequence, and respectively constructing state branch input, low-frequency background branch input and high-frequency disturbance branch input; the application constructs state branches, low-frequency background branches and high-frequency disturbance branches, respectively represents marine and atmospheric environmental states, EDH smooth trends and short-time local fluctuations, and independently encodes the branches in time sequence, so that mutual interference caused by direct mixing of different types of information is reduced, and the representation ability of the model to multi-scale time sequence change characteristics of EDH is improved.

Description

Technical Field

[0001] This invention relates to the field of marine environment prediction technology, and in particular to a multi-step prediction method for evaporation waveguide height based on state anchoring gated residuals. Background Technology

[0002] In marine channel modeling and marine electromagnetic propagation environment assessment, evaporation ducts are a crucial factor affecting the performance of marine radar, communication, and electronic reconnaissance systems. Evaporation Duct Height (EDH) is an important parameter describing the vertical range of near-surface evaporation ducts in the ocean, and its variation is closely related to near-surface meteorological conditions such as air pressure, sea surface temperature, air temperature, relative humidity, and wind speed. Due to the influence of complex air-sea interactions, EDH typically exhibits nonlinear, non-stationary, and short-term continuous variation characteristics, making short-term multi-step prediction highly challenging.

[0003] Existing methods for predicting evaporation duct height mainly include physical models, empirical models, statistical models, and prediction methods based on machine learning or deep learning. Among these, physical and empirical models typically calculate evaporation duct height based on near-sea meteorological parameters and atmospheric boundary layer theory, possessing a certain degree of physical interpretability. Data-driven methods can learn nonlinear relationships between historical evaporation duct heights and ocean-atmosphere environmental variables, improving prediction accuracy to some extent. However, in short-term, multi-step prediction scenarios, existing methods still have the following shortcomings: 1. Existing methods typically use historical evaporation waveguide height and air-sea environmental variables as unified inputs for modeling, failing to adequately distinguish between different types of information such as current environmental conditions, low-frequency background evolution, and high-frequency local disturbances. This makes it difficult for the model to simultaneously take into account both the overall trend of change and short-term fluctuation characteristics. 2. The height of the evaporation waveguide may fluctuate significantly in a short period of time. Existing multi-step prediction methods tend to smooth out such fluctuations, resulting in peak attenuation, lag in changes, and increased local prediction errors. 3. Existing methods mostly directly predict the absolute value of the evaporation waveguide height at future times, without fully utilizing the constraint effect of the latest observation state at the end of the historical input window on the short-term changes in the future, which is prone to state shifts in the continuous multi-step prediction process. 4. In the process of continuous multi-step prediction, prediction uncertainty accumulates, and the impact of local disturbances on different prediction step lengths varies. Existing methods lack the ability to adaptively coordinate between basic changes and local corrections, resulting in the prediction stability under longer prediction step lengths still needing to be improved.

[0004] Therefore, it is necessary to provide a short-time multi-step prediction method for evaporation waveguide height, which can construct hierarchical features of historical evaporation waveguide height and its corresponding air-sea environmental variables, respectively characterizing environmental state, low-frequency background evolution and high-frequency disturbance information. Combined with gated residual correction and state anchoring output, the method can improve the model's response to short-time local fluctuations and the stability of continuous multi-step prediction. Summary of the Invention

[0005] The purpose of this invention is to solve the problems in the prior art by proposing a multi-step prediction method for evaporation waveguide height based on state anchoring gated residuals.

[0006] A multi-step prediction method for evaporation waveguide height based on state-anchored gated residuals includes the following steps: S1: Acquire and preprocess historical evaporation waveguide height data and corresponding air-sea environmental variables to obtain the basic input sequence; S2: Let the length of the history input window be... The predicted step size is The basic input sequence is sliced ​​along the time axis to construct historical input sequences and their corresponding future EDH prediction targets. S3: Construct multi-branch features for the historical input sequence, and construct state branch input, low-frequency background branch input and high-frequency disturbance branch input respectively; S4: Input the state branch input, low-frequency background branch input, and high-frequency disturbance branch input into the corresponding timing encoders for timing encoding to obtain state representation, low-frequency background representation, and high-frequency disturbance representation; fuse the state representation and low-frequency background representation to obtain state-low-frequency enhanced representation; S5: Input the state-low frequency enhancement characterization and high frequency disturbance characterization into the basic change prediction module, the high frequency residual prediction module and the gate weight generation module respectively to obtain the basic change amount, the high frequency residual correction amount and the gate weight; use the gate weight to adjust the high frequency residual correction amount and fuse it with the basic change amount to obtain the initial future change prediction amount. S6: Using the EDH at the end of the historical input window as the current EDH state anchor point, the initial future change prediction is constrained and then added to the current EDH state anchor point to generate the future... EDH prediction sequence corresponding to each prediction step size; S7: Construct a training loss function based on the error between the EDH prediction sequence and the corresponding real prediction target, iteratively update the model parameters, and obtain the trained short-time multi-step prediction model for evaporation waveguide height. S8: Using the trained prediction model, the time before the time to be predicted is... The evaporation waveguide height data and its corresponding air-sea environmental variables are calculated forward according to steps S3 to S6 to obtain the future values ​​after the time to be predicted. The predicted height of the evaporation waveguide corresponding to each prediction step size.

[0007] Preferably, in step S1, the basic input sequence is specifically the first... The fundamental input vector at each time step:

[0008] in, Indicates air pressure. Indicates sea surface temperature. Indicates temperature, Indicates relative humidity. Indicates wind speed. Indicates the first EDH at any given moment; The preprocessing includes time sorting, anomaly cleaning, missing value handling, and time alignment.

[0009] Preferably, step S2 specifically includes: For the first Using each time point as a prediction sample at the end of the historical input window, a historical input sequence is constructed. :

[0010] And use its corresponding future EDH sequence as the prediction target. :

[0011] in, This represents the feature dimension of the basic input vector. Indicates the future number The actual EDH value corresponding to each prediction step size, and .

[0012] Preferably, in step S3, the state branch input Represented as:

[0013] in, Indicates air pressure. Indicates sea surface temperature. Indicates temperature, Indicates relative humidity. Indicates wind speed. Indicates the first EDH at any given moment; The low-frequency background branch input Represented as:

[0014] in, This represents the EDH rolling mean calculated within a preset rolling window. This represents the weighted average of the EDH index. This indicates the slope of EDH's change within the preset window. Indicates the phase of a time period; The construction of high-frequency perturbation branch input Represented as:

[0015] in, Indicates the first The difference characteristics of EDH at each time step relative to the previous time step. and These represent the positive and negative change components, respectively. This represents the standard deviation of EDH within a short time window. Indicates the first The deviation of EDH from the rolling mean background at each time point. This indicates the interaction characteristics between sea surface temperature difference and wind speed. This indicates the interaction characteristics between relative humidity and wind speed.

[0016] Preferably, in step S4, the temporal encoder includes a linear mapping layer, a one-dimensional temporal convolutional layer, a residual connection, a normalization layer, and a temporal aggregation layer. The temporal aggregation layer simultaneously uses the encoding features at the end of the historical input window and the average aggregated features of the encoding features at each time point within the historical window along the time dimension for weighted fusion. The state representation is fused with the low-frequency background representation, specifically as follows:

[0017] in, Indicates feature splicing, and For trainable parameters, Represents a non-linear activation function. For state representation, Characterized by low-frequency background. This is a state-low frequency enhancement characterization.

[0018] Preferably, step S5 specifically includes: The state-low frequency enhancement characterization Input the basic change prediction module to obtain the future The fundamental change of each prediction step size relative to the current state:

[0019] in, This represents the basic change prediction module. Indicates the future The basic change estimate of each prediction step size relative to the current state is used to characterize the main trend of future changes of EDH relative to the current state. Characterize the high-frequency perturbation branch encoding Input the high-frequency residual prediction module to obtain the future High-frequency residual correction amount corresponding to each prediction step size:

[0020] in, This indicates the high-frequency residual prediction module. This represents the high-frequency residual term used to correct for changes in the fundamental level; The state-low frequency enhancement characterization High-frequency perturbation branch coding characterization The data is concatenated and input into the gating weight generation module to obtain the gating weights corresponding to each future prediction step size:

[0021] in, This indicates the gating weight generation module. This represents the Sigmoid function. This represents the residual correction weights corresponding to different prediction step sizes, and the value of each weight ranges from 0 to 1; The high-frequency residual correction is adjusted step-by-step using the gating weights, and the base change is fused element-by-element with the high-frequency residual correction adjusted by the gating weights.

[0022] in, This indicates element-wise multiplication. This represents the initial predicted future changes obtained after fusion.

[0023] Preferably, step S6 specifically includes: Initial future change forecast Input the change magnitude constraint module to obtain the constrained future change prediction:

[0024] in, This indicates the preset scale parameter. Represents the hyperbolic tangent function, where the constrained future change prediction lies in ( , ) interval; Generate the future EDH prediction values ​​corresponding to each prediction step size:

[0025] in, This represents the current EDH state anchor point used to generate future prediction results. Indicates the future number The predicted change of each prediction step size relative to the current EDH state anchor point.

[0026] Preferably, in step S7, the training loss function uses the prediction step-size weighted Huber loss as the main loss:

[0027] in, For the first Loss weights for each prediction step size This represents the Huber loss function. For the first EDH prediction values ​​corresponding to each prediction step size For the first The true value of the EDH for each predicted step size.

[0028] Preferably, the training loss function further includes a long-term prediction auxiliary loss, which is applied to a preset set of long-term prediction step sizes. Construct a long-term prediction auxiliary loss:

[0029] The total training loss is obtained by weighting the main loss and the long-term prediction auxiliary loss:

[0030] in, This represents the weighting coefficient for the auxiliary loss in long-term forecasting.

[0031] Compared with existing technologies, the advantages of this invention are: 1. This invention constructs state branches, low-frequency background branches, and high-frequency disturbance branches to represent the atmospheric-ocean environment state, EDH smoothing trend, and short-term local fluctuations, respectively. Each branch is independently time-series encoded to reduce mutual interference caused by the direct mixing of different types of information and improve the model's ability to represent the multi-scale time-series changes of EDH.

[0032] 2. This invention generates residual correction quantities based on high-frequency disturbance characterization, and uses gated weights to control the degree of participation of these quantities in different prediction steps, thereby achieving adaptive adjustment of the high-frequency correction intensity. At the same time, it uses state-low frequency enhancement to characterize the predicted basic change quantity, thereby depicting the main evolution direction of EDH and avoiding insufficient or excessive introduction of high-frequency disturbances.

[0033] 3. This invention uses the end of the historical window, EDH, as the state anchor point to generate prediction results, so that the output of each step is based on the latest observation, reducing the overall offset in multi-step prediction; at the same time, it introduces amplitude constraints to limit abnormal changes, improves prediction stability, and makes the predicted value more consistent with the current actual state. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the overall architecture of the present invention.

[0035] Figure 2 This is a schematic diagram of the multi-branch feature encoding and representation fusion process in this invention. Detailed Implementation

[0036] To facilitate understanding of this application and to make the above-mentioned objectives, features and advantages of this application more apparent, the specific embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0037] Reference Figure 1-2 As shown, a multi-step prediction method for evaporation waveguide height based on state-anchored gated residuals includes the following steps: S1: Acquire and preprocess historical evaporation waveguide height (EDH) data and corresponding air-sea environmental variables. Read historical EDH data recorded in chronological order and its corresponding air-sea environmental variables. These variables include air pressure, sea surface temperature, air temperature, relative humidity, and wind speed. Perform time sorting, anomaly cleaning, missing value handling, and time alignment on the raw data to obtain the basic input sequence.

[0038] No. The fundamental input vector at each time step is represented as: .

[0039] in, Indicates air pressure. Indicates sea surface temperature. Indicates temperature, Indicates relative humidity. Indicates wind speed. Indicates the first EDH at a given moment.

[0040] In some implementations, abnormal data are identified and corrected, and missing data is supplemented to ensure the continuity and consistency of evaporation waveguide height data with the corresponding ocean-atmosphere environmental variables in the time dimension.

[0041] S2: Construct the historical input sequence and the future multi-step prediction target. Let the historical input window length be... The predicted step size is Slide the basic data sequence obtained from S1 along the time axis. For the data sequence starting from the first time step... Using the predicted samples at the end of the historical input window as time points, construct the historical input sequence:

[0042] And its corresponding future EDH sequence is used as the prediction target:

[0043] in, This represents the feature dimension of the basic input vector. Indicates the future number The actual EDH value corresponding to each prediction step size, and When the time interval for the basic data is 1 hour, The number of times after the end of the corresponding historical input window The actual value of EDH per hour. As input for subsequent multi-branch feature construction, It serves as the true prediction target during model training and is used to evaluate the model's prediction results.

[0044] During the model training phase, the multiple prediction samples are divided into training set, validation set and test set according to the time sequence corresponding to each prediction sample, so as to avoid leakage of time information between different datasets.

[0045] S3: Construct multi-branch features for the historical input sequence. Based on the historical input sequence, construct state branches, low-frequency background branches, and high-frequency disturbance branches respectively. Specifically, extract ocean-atmosphere environmental variables and historical EDH from the historical input sequence to form the state branch input; calculate smoothing statistical features and time periodic features based on historical EDH to form the low-frequency background branch input; and calculate difference features, local fluctuation features, background deviation features, and ocean-atmosphere interaction features based on historical EDH to form the high-frequency disturbance branch input.

[0046] For any branch Construct the corresponding branch input sequence:

[0047] in, , , These represent the state branch, the low-frequency background branch, and the high-frequency disturbance branch, respectively. When... When used as a superscript, it indicates a high-frequency perturbation branch; when When used as a prediction step size index, it indicates the future step size index. Each prediction step size.

[0048] Furthermore, step S3 specifically includes the following steps: S31: Constructing the state branch input. Extract air pressure, sea surface temperature, air temperature, relative humidity, wind speed, and EDH from historical observation sequences to form the single-time-input of the state branch:

[0049] And according to the historical window length Construct the state branch input sequence: .

[0050] S32: Construct the low-frequency background branch input. Based on the historical EDH sequence, calculate one or more of the following: EDH rolling mean, exponentially weighted mean, low-frequency change slope, and time periodic features, to form the single-moment input of the low-frequency background branch:

[0051] in, This represents the EDH rolling mean calculated within a preset rolling window. This represents the weighted average of the EDH index. This indicates the slope of EDH's change within the preset window. Indicates the phase of the time period.

[0052] Also based on the historical window length Constructing the low-frequency background branch input sequence: .

[0053] S33: Construct the high-frequency perturbation branch input. Based on the historical EDH sequence, calculate one or more of the following: difference features, positive and negative variation components, short window standard deviation, background deviation, and air-sea interaction features, to form the single-time-step input of the high-frequency perturbation branch:

[0054] in, Indicates the first The difference characteristics of EDH at each time step relative to the previous time step. and These represent the positive and negative change components, respectively. This represents the standard deviation of EDH within a short time window. Indicates the first The deviation of EDH from the rolling mean background at each time point. This branch is used to characterize short-term fluctuations and rapid local changes in EDH. The air-sea interaction features include the interaction features between air-sea temperature difference and wind speed, and the interaction features between relative humidity and wind speed, respectively expressed as:

[0055]

[0056] According to historical window length Constructing the high-frequency perturbation branch input sequence: .

[0057] In some implementations, multiple time windows of different lengths are used to calculate the corresponding low-frequency background features and high-frequency disturbance features to characterize the smooth changes and short-term local fluctuations of EDH at different time scales. The low-frequency background branch and high-frequency disturbance branch are functional divisions based on the feature change scale and characterization effect, and are not limited to low-frequency components and high-frequency components in the frequency domain sense.

[0058] S4: Perform temporal coding and feature fusion on the multi-branch input. The three types of branch input sequences from S3 are input into their respective temporal encoders to obtain the corresponding state representation, low-frequency background representation, and high-frequency perturbation representation. Then, the state representation and low-frequency background representation are concatenated and fused using a nonlinear mapping to obtain the state-low-frequency enhanced representation. Finally, the obtained state-low-frequency enhanced representation... Used for subsequent prediction of fundamental changes, and characterization of high-frequency disturbances. Used for subsequent high-frequency residual correction.

[0059] Furthermore, step S4 specifically includes the following steps: S41: Perform feature mapping on the input features of each branch. For any branch... The input sequence is fed into the corresponding linear mapping layer, and the branch input features at each time step are mapped to a latent feature space of a preset dimension, and nonlinear activation processing is performed; in some implementations, random deactivation processing is further performed during the model training phase, and then the initial hidden layer feature sequence is obtained:

[0060] in, Indicates branch The corresponding feature mapping operation, This represents the initial hidden layer feature sequence after mapping.

[0061] The feature mapping is used to transform input features with different dimensions and physical meanings in different branches into a unified latent feature space, while maintaining the temporal order of the input sequences of each branch.

[0062] S42: Extract and aggregate the temporal features of each branch. Input the initial hidden layer feature sequence into the corresponding temporal feature extraction layer to extract the temporal dependencies and local change patterns between adjacent moments within the historical input window. Then, obtain the corresponding branch temporal representations through residual connections and temporal aggregation.

[0063] in, Indicates branch The corresponding timing encoder, This represents the corresponding branch timing representation. After the timing encoding, state representations are obtained respectively. Low-frequency background characterization and high-frequency disturbance characterization .

[0064] In some implementations, the temporal encoder includes a linear mapping layer, a one-dimensional temporal convolutional layer, residual connections, a normalization layer, and a temporal aggregation layer. The temporal aggregation simultaneously utilizes features from the end of the historical input window and overall features within the historical window to take into account both current state information and historical change information.

[0065] S43: Fusion of state representation and low-frequency background representation. The state representation and low-frequency background representation are combined using feature concatenation and nonlinear mapping to obtain the state-low-frequency enhanced representation:

[0066] in, Indicates feature splicing, and For trainable parameters, This represents a non-linear activation function.

[0067] The state-low frequency enhancement characterization Used to jointly characterize the atmospheric-sea environment state, historical evaporation waveguide height level, smoothing trend, and periodic changes, and for predicting fundamental changes in S5; the high-frequency perturbation characterization It remains independent for high-frequency residual correction and gated weight generation in S5.

[0068] S5: Gated residual prediction is performed based on multi-branch representation to obtain the initial future change prediction. The state-low frequency enhanced representation obtained in S4 is then used. and high-frequency disturbance characterization Inputs are made to the basic change prediction module, the high-frequency residual prediction module, and the gating weight generation module to obtain the basic change amount, the high-frequency residual correction amount, and the gating weight. The high-frequency residual correction amount is adjusted using the gating weight, and then fused with the basic change amount to obtain the initial future change prediction amount. The high-frequency residual correction amount is not the error between the predicted true value and the predicted result, but rather an additional correction amount generated relative to the basic change amount.

[0069] Furthermore, step S5 specifically includes the following steps: S51: Generate the basic change quantity based on the state-low frequency enhancement characterization. (This is the process of generating the state-low frequency enhancement characterization obtained in step S4.) Input the basic change prediction module to obtain the future Estimates of the fundamental change in each prediction step size relative to the current state:

[0070] in, This represents the basic change prediction module. Indicates the future The basic change estimate of the prediction step size relative to the current state is used to characterize the main trend of future changes of EDH relative to the current state.

[0071] S52: Generate high-frequency residual correction based on the high-frequency perturbation characterization. Then, encode the high-frequency perturbation branch characterization obtained in step S4. Input the high-frequency residual prediction module to obtain the future High-frequency residual correction amount corresponding to each prediction step size:

[0072] in, This indicates the high-frequency residual prediction module. This represents the high-frequency residual term used to correct for changes in the fundamental quantity.

[0073] S53: Generate gating weights based on the state-low frequency enhancement representation and the high frequency disturbance representation. (The state-low frequency enhancement representation...) High-frequency perturbation branch coding characterization The data is concatenated and input into the gating weight generation module to obtain the gating weights corresponding to each future prediction step size:

[0074] in, This indicates the gating weight generation module. This represents the Sigmoid function. This represents the residual correction weights corresponding to different prediction step sizes, and the value of each weight ranges from 0 to 1. The gating weights are used to control the contribution ratio of high-frequency residual corrections in the final change prediction.

[0075] S54: The initial future change prediction is obtained by fusing the basic change and the gated residual correction. The high-frequency residual correction is adjusted step-by-step using the gated weights, and the basic change is then fused element-by-element with the gated residual correction.

[0076] in, This indicates element-wise multiplication. This represents the initial predicted future change obtained after fusion. When the gating weight corresponding to a certain prediction step size is small, the prediction result is mainly generated based on the basic change; when the gating weight is large, the correction effect of the high-frequency residual correction on the basic change is strengthened. Thus, the model can adaptively adjust the correction intensity of short-term local disturbances according to different historical samples and different prediction step sizes. The initial predicted future change is an unconstrained change, used for subsequent scale constraints and state anchoring output.

[0077] S6: Generate future multi-step EDH prediction results based on the current EDH state anchor point. Using the EDH at the end of the historical input window as the current EDH state anchor point, apply amplitude constraints to the initial future change predictions obtained in S5. Then, add the constrained future change predictions to the current EDH state anchor point to obtain the future step-by-step predictions. EDH prediction values ​​for each prediction step size.

[0078] Furthermore, step S6 specifically includes the following steps: S61: Determine the current EDH state anchor point Extract the EDH at the end of the historical input window and use it as the anchor point of the current EDH state:

[0079] in, EDH represents the time at the end of the historical input window. This represents the current EDH state anchor point used to generate future prediction results.

[0080] By setting the current EDH state anchor point, the model learns the change in each future prediction step size relative to the current EDH value based on the current EDH level, rather than directly generating the absolute value of the future EDH.

[0081] S62: Apply amplitude constraints to the initial future change forecast. The initial future change forecast obtained from S5 Input the change magnitude constraint module to obtain the constrained future change prediction:

[0082] in, This indicates the preset scale parameter. This represents the hyperbolic tangent function.

[0083] The change amplitude constraint module first determines the change scale parameter. The initial future change forecasts are scaled and normalized, then smoothly mapped to ( ) using the hyperbolic tangent function. The interval is 1,1), and based on the scaling parameter. Restoring the magnitude of change so that the constrained future change forecast is located at ( , The constraint, while approximately preserving small predicted changes, smooths and suppresses abnormally large predicted changes, thereby improving the stability of future multi-step prediction results.

[0084] Specifically, the constrained future change prediction sequence is represented as follows:

[0085] in, Indicates the future number The predicted change of each prediction step size relative to the current EDH state anchor point.

[0086] S63: Generate a sequence of future multi-step evaporation waveguide height predictions Add the constrained future change corresponding to each prediction step size to the current EDH state anchor point to obtain the future step size. EDH prediction values ​​corresponding to each prediction step size:

[0087] From this, we can obtain the future. EDH prediction sequence corresponding to each time point:

[0088] By using the state anchoring method, the output of each prediction step is referenced to the EDH at the end of the historical input window, thereby maintaining the continuity of short-term EDH changes and reducing the overall offset risk caused by directly predicting the future absolute height.

[0089] S7: Construct the training loss function and complete model training. Based on the EDH prediction sequence obtained in S6. With the corresponding true prediction target The training loss function is constructed based on the error between the parameters, and the model parameters are iteratively updated based on the training loss function.

[0090] In some implementations, the prediction step-weighted Huber loss is used as the main loss:

[0091] in, For the first Loss weights for each prediction step size This represents the Huber loss function.

[0092] Furthermore, for the preset set of long-term prediction step sizes Construct a long-term prediction auxiliary loss:

[0093] When using the aforementioned long-term prediction auxiliary loss, the main loss and the long-term prediction auxiliary loss are weighted and combined to obtain the total training loss:

[0094] in, This represents the weighting coefficient for the auxiliary loss in long-term forecasting.

[0095] Based on the total training loss, the trainable parameters in the prediction model are iteratively updated through backpropagation and optimization algorithms; training stops when the preset training termination condition is met, and the model parameters corresponding to the optimal prediction performance on the validation set are retained, thus obtaining the trained short-time multi-step prediction model for evaporation waveguide height.

[0096] S8: Perform short-time multi-step prediction of the evaporation waveguide height using the trained model. Obtain the length before the time to be predicted as... The evaporation waveguide height data and its corresponding air-sea environmental variables are used to construct state branch input, low-frequency background branch input, and high-frequency disturbance branch input according to the method described in step S3, and then input into the trained short-time multi-step prediction model of the evaporation waveguide height. The prediction model is then used for forward calculation according to the methods described in steps S4 to S6 to obtain the future value after the time to be predicted. The predicted height of the evaporation waveguide corresponding to each prediction step size.

[0097] This invention does not merely perform general neural network fitting on historical meteorological variables. Instead, based on the characteristic that short-term changes in evaporation duct height simultaneously involve smooth background evolution and local high-frequency disturbances, it divides the input information into a state branch, a low-frequency background branch, and a high-frequency disturbance branch. Furthermore, it uses state-low-frequency enhancement to characterize the predicted fundamental change relative to the current state, and uses high-frequency disturbances to characterize the generated residual correction. Gated weights control the degree of participation of the residual correction in different prediction steps. Finally, it uses the end of the historical window (EDH) as the state anchor point to generate future multi-step prediction results. This allows the model to maintain the overall trend while improving its responsiveness to short-term local fluctuations and reducing the state shift caused by directly predicting absolute values.

[0098] Example 1 This embodiment uses historical EDH data recorded hourly and its corresponding marine-atmospheric environmental variables as an example. These variables include air pressure, sea surface temperature, air temperature, relative humidity, and wind speed. This embodiment uses data from 48 historical time points to predict the EDH for the next 6 time points, i.e., the length of the historical input window. =48, prediction step size =6. Specifically, it includes the following steps.

[0099] Step 1: Acquire and preprocess historical data Read the historical EDH data recorded in chronological order and its corresponding ocean-atmosphere environmental variables. Perform time sorting, outlier processing, missing value processing and time alignment on the raw data to make the evaporation waveguide height data correspond one-to-one with the corresponding ocean-atmosphere environmental variables in the time dimension, and obtain the basic input sequence.

[0100] In this embodiment, the first The basic input vector at each moment includes air pressure, sea surface temperature, air temperature, relative humidity, wind speed, and EDH, namely:

[0101] in, Indicates air pressure. Indicates sea surface temperature. Indicates temperature, Indicates relative humidity. Indicates wind speed. Indicates the first EDH at a given moment.

[0102] Step 2: Constructing short-time multi-step prediction samples A sliding window approach is used to construct model samples. Specifically, the basic input sequence of 48 consecutive time steps is used as the historical input, and the EDH of the next 6 consecutive time steps is used as the prediction target.

[0103] for For each sample, the historical input sequence is represented as:

[0104] The prediction target is expressed as:

[0105] in, Input to the model, The EDH values ​​are the actual values ​​corresponding to the next 6 prediction steps. Based on the chronological order of the prediction samples, the constructed prediction samples are divided into training, validation, and test sets.

[0106] Step 3: Construct multi-branch input features Based on the historical input sequence of each prediction sample State branch input, low-frequency background branch input, and high-frequency disturbance branch input are constructed respectively.

[0107] The state branch input consists of air pressure, sea surface temperature, air temperature, relative humidity, wind speed, and EDH, which are used to characterize the air-sea environment state and historical EDH levels.

[0108] The low-frequency background branch input consists of the EDH rolling mean, exponentially weighted mean, low-frequency change slope, and time periodic features, which are used to characterize the smooth trend and periodic changes of EDH.

[0109] The high-frequency disturbance branch input consists of EDH differential features, positive and negative variation components, short-window standard deviation, background deviation, and air-sea interaction features, used to characterize short-term local fluctuations in EDH. Among them, the air-sea interaction features include the interaction features between sea surface temperature difference and wind speed, and the interaction features between relative humidity and wind speed.

[0110] After constructing the input features for each branch, the corresponding standardization parameters are determined based on the input features of each branch in the training set, and the same standardization parameters are used to standardize the corresponding input features of the branches in the training set, validation set, and test set.

[0111] Step 4: Perform the forward computation process of the model The standardized state branch input, low-frequency background branch input, and high-frequency disturbance branch input are input into the corresponding time-series encoders to obtain state representation, low-frequency background representation, and high-frequency disturbance representation. For each branch, the encoded features at the end of the historical input window are extracted, and the encoded features at each time point within the historical input window are averaged and aggregated along the time dimension. Then, the end-time features are weighted and fused with the averaged aggregated features to obtain the state representation, low-frequency background representation, and high-frequency disturbance representation.

[0112] Subsequently, the state representation is fused with the low-frequency background representation to obtain the state-low-frequency enhanced representation, and the state-low-frequency enhanced representation is input into the basic change prediction module to obtain the basic change amount of the next 6 prediction steps relative to the current state.

[0113] The high-frequency disturbance characterization is input into the high-frequency residual prediction module to obtain the high-frequency residual correction amount corresponding to the next 6 prediction steps. Simultaneously, gating weights corresponding to the next 6 prediction steps are generated based on the state-low-frequency enhancement characterization and the high-frequency disturbance characterization. The high-frequency residual correction amount is adjusted using these gating weights and fused with the basic change amount to obtain the initial future change prediction amount.

[0114] Finally, the EDH at the end of the historical input window is used as the current EDH state anchor point. After applying scale constraints to the initial future change prediction, it is added to the current EDH state anchor point to obtain the EDH prediction sequence for the next 6 prediction steps corresponding to the current training sample.

[0115] Step 5: Model Training and Testing The model is input with samples from the training set. The training loss is calculated based on the EDH prediction sequence and the corresponding EDH true value output by the model. The model parameters are updated through backpropagation and optimization algorithms. The model performance is evaluated and the model parameters are selected using the validation set to obtain the trained model. The prediction performance of the trained model on data not used in the training is evaluated using the test set.

[0116] In this embodiment, the model employs a direct multi-step prediction approach, outputting the EDH prediction sequence corresponding to the next 6 prediction steps in one go. Simultaneously, the AdamW optimization algorithm is used to update the model parameters, and an early stopping strategy is employed to control the training process based on the validation set RMSE, retaining the model parameters corresponding to the minimum validation set RMSE. Evaluation metrics include Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Coefficient of Determination (R²). 2 .

[0117] Step Six: Prediction Results With a historical input window length of 48 and a prediction step size of 6, the input features of each branch in the test set were input into the trained model and the comparison model. The comparison models included the extreme gradient boosting model XGBoost, the long short-term memory network LSTM, the Transformer model, and the DLinear model. The specific results are shown in the table below: Table 1. Multi-model prediction results of Example 1

[0118] The method in this embodiment achieved an MAE of 0.7128, an MSE of 0.9374, an RMSE of 0.9682, and an R...2 The prediction result of 0.9056 indicates that the method in this embodiment can achieve high-precision short-time multi-step prediction for EDH. Under the same historical input window length and prediction step size settings, compared with the comparative model, the method in this embodiment can reduce the prediction error, demonstrating that multi-branch feature construction, gated residual correction, and state-anchored output are beneficial to improving the performance of short-time multi-step prediction for EDH.

[0119] Further testing showed that replacing the multi-branch input with a single hybrid input, or removing the high-frequency gated residual correction and state anchoring output, increased the model prediction error, indicating that the above-mentioned technical modules help improve the short-time multi-step prediction performance of evaporating waveguide height.

[0120] Example 2 This embodiment modifies the prediction step size setting based on Embodiment 1 to illustrate the applicability of the method of the present invention over a longer prediction range.

[0121] The difference between this embodiment and Embodiment 1 is that the length of the history input window is still set to... =48, prediction step size is from =6 adjusted to =12, meaning that EDH data from 48 historical time points and their corresponding air-sea environmental variables are used to predict EDH for the next 12 time points. Except for the prediction step size and model output dimension, the data preprocessing, basic input sequence construction, multi-branch feature construction, temporal coding, gated residual prediction, and state anchoring output processes in this embodiment are the same as in Embodiment 1.

[0122] In this embodiment, for the first For each sample, the historical input sequence is still represented as:

[0123] The prediction target is expanded from the next 6 prediction steps in Example 1 to the next 12 prediction steps, as shown below:

[0124] Therefore, the basic change prediction module outputs the basic change amount relative to the current state for the next 12 prediction steps, the high-frequency residual prediction module outputs the high-frequency residual correction amount corresponding to the next 12 prediction steps, and the gating weight generation module outputs the gating weight corresponding to the next 12 prediction steps. Accordingly, the basic change amount, the high-frequency residual correction amount, the gating weight, and the initial future change prediction amount are all 12-dimensional vectors, corresponding to the 1st to 12th prediction steps, respectively.

[0125] After the model outputs the initial future change prediction, it still uses the EDH at the end of the historical input window as the current EDH state anchor point, applies a scale constraint to the initial future change prediction, and adds it to the current EDH state anchor point to obtain the EDH prediction sequence corresponding to the next 12 prediction steps.

[0126] With a historical input window length of 48 and a prediction step size of 12, the input features of each branch in the test set are input into the trained model and the comparison model to obtain the EDH prediction sequence corresponding to the next 12 prediction steps. The mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), and coefficient of determination (R²) are then used. 2 The prediction results were evaluated. The specific results are shown in the table below: Table 2. Multi-model prediction results of Example 2

[0127] The method in this embodiment achieved an MAE of 0.9675, an MSE of 1.6620, an RMSE of 1.2892, and an R... 2 The prediction result is 0.8327. Compared with the above comparative model, the method in this embodiment can reduce the prediction error, indicating that even when the prediction step size is expanded to 12, the method can still achieve continuous multi-step prediction of EDH and adapt to short-term prediction tasks with a longer prediction range.

[0128] As is known from common technical knowledge, this invention can be implemented through other embodiments that do not depart from its spirit or essential characteristics. Therefore, the disclosed embodiments described above are merely illustrative in all respects and are not the only ones. All modifications within the scope of this invention or its equivalents are included in this invention.

Claims

1. A multi-step prediction method for evaporation waveguide height based on state-anchored gated residuals, characterized in that, Includes the following steps: S1: Acquire and preprocess historical evaporation waveguide height data and corresponding air-sea environmental variables to obtain the basic input sequence; S2: Let the length of the history input window be... The predicted step size is The basic input sequence is sliced ​​along the time axis to construct historical input sequences and their corresponding future EDH prediction targets; S3: Construct multi-branch features for the historical input sequence, and construct state branch input, low-frequency background branch input and high-frequency disturbance branch input respectively; S4: Input the state branch input, low-frequency background branch input, and high-frequency disturbance branch input into the corresponding timing encoders for timing encoding to obtain state representation, low-frequency background representation, and high-frequency disturbance representation; fuse the state representation and low-frequency background representation to obtain state-low-frequency enhanced representation; S5: Input the state-low frequency enhancement characterization and high frequency disturbance characterization into the basic change prediction module, the high frequency residual prediction module and the gate weight generation module respectively to obtain the basic change amount, the high frequency residual correction amount and the gate weight; use the gate weight to adjust the high frequency residual correction amount and fuse it with the basic change amount to obtain the initial future change prediction amount. S6: Using the EDH at the end of the historical input window as the current EDH state anchor point, the initial future change prediction is constrained and then added to the current EDH state anchor point to generate the future... EDH prediction sequence corresponding to each prediction step size; S7: Construct a training loss function based on the error between the EDH prediction sequence and the corresponding real prediction target, iteratively update the model parameters, and obtain the trained short-time multi-step prediction model for evaporation waveguide height. S8: Using the trained prediction model, the time before the time to be predicted is... The evaporation waveguide height data and its corresponding air-sea environmental variables are calculated forward according to steps S3 to S6 to obtain the future values ​​after the time to be predicted. The predicted height of the evaporation waveguide corresponding to each prediction step size.

2. The multi-step prediction method for evaporation waveguide height based on state-anchored gated residuals according to claim 1, characterized in that: In step S1, the basic input sequence is specifically the first... The fundamental input vector at each time step: in, Indicates air pressure. Indicates sea surface temperature. Indicates temperature, Indicates relative humidity. Indicates wind speed. Indicates the first EDH at any given moment; The preprocessing includes time sorting, anomaly cleaning, missing value handling, and time alignment.

3. The multi-step prediction method for evaporation waveguide height based on state-anchored gated residuals according to claim 1, characterized in that: Step S2 specifically includes: For the first Using each time point as a prediction sample at the end of the historical input window, a historical input sequence is constructed. : And use its corresponding future EDH sequence as the prediction target. : in, This represents the feature dimension of the basic input vector. Indicates the future number The actual EDH value corresponding to each prediction step size, and .

4. The multi-step prediction method for evaporation waveguide height based on state-anchored gated residuals according to claim 1, characterized in that: In step S3, the state branch input Represented as: in, Indicates air pressure. Indicates sea surface temperature. Indicates temperature, Indicates relative humidity. Indicates wind speed. Indicates the first EDH at any given moment; The low-frequency background branch input Represented as: in, This represents the EDH rolling mean calculated within a preset rolling window. This represents the weighted average of the EDH index. This indicates the slope of EDH's change within the preset window. Indicates the phase of a time period; The construction of high-frequency perturbation branch input Represented as: in, Indicates the first The difference characteristics of EDH at each time step relative to the previous time step. and These represent the positive and negative change components, respectively. This represents the standard deviation of EDH within a short time window. Indicates the first The deviation of EDH from the rolling mean background at each time point. This indicates the interaction characteristics between sea surface temperature difference and wind speed. This indicates the interaction characteristics between relative humidity and wind speed.

5. The multi-step prediction method for evaporation waveguide height based on state-anchored gated residuals according to claim 1, characterized in that: In step S4, the temporal encoder includes a linear mapping layer, a one-dimensional temporal convolutional layer, a residual connection, a normalization layer, and a temporal aggregation layer. The temporal aggregation layer simultaneously uses the encoding features at the end of the historical input window and the average aggregated features of the encoding features at each time point within the historical window along the time dimension for weighted fusion. The state representation is fused with the low-frequency background representation, specifically as follows: in, Indicates feature splicing, and For trainable parameters, Represents a non-linear activation function. For state representation, Characterized by low-frequency background. This is a state-low frequency enhancement characterization.

6. The multi-step prediction method for evaporation waveguide height based on state-anchored gated residuals according to claim 1, characterized in that: Step S5 specifically includes: The state-low frequency enhancement characterization Input the basic change prediction module to obtain the future The fundamental change of each prediction step size relative to the current state: in, This represents the basic change prediction module. Indicates the future The basic change estimate of each prediction step size relative to the current state is used to characterize the main trend of future changes of EDH relative to the current state. Characterize the high-frequency perturbation branch encoding Input the high-frequency residual prediction module to obtain the future High-frequency residual correction amount corresponding to each prediction step size: in, This indicates the high-frequency residual prediction module. This represents the high-frequency residual term used to correct for changes in the fundamental level; The state-low frequency enhancement characterization High-frequency perturbation branch coding characterization The data is concatenated and input into the gating weight generation module to obtain the gating weights corresponding to each future prediction step size: in, This indicates the gating weight generation module. This represents the Sigmoid function. This represents the residual correction weights corresponding to different prediction step sizes, and the value of each weight ranges from 0 to 1; The high-frequency residual correction is adjusted step-by-step using the gating weights, and the base change is fused element-by-element with the high-frequency residual correction adjusted by the gating weights. in, This indicates element-wise multiplication. This represents the initial predicted future changes obtained after fusion.

7. The multi-step prediction method for evaporation waveguide height based on state-anchored gated residuals according to claim 1, characterized in that: Step S6 specifically includes: Initial future change forecast Input the change magnitude constraint module to obtain the constrained future change prediction: in, This indicates the preset scale parameter. Represents the hyperbolic tangent function, and the constrained future change prediction lies in ( , ) interval; Generate the future EDH prediction values ​​corresponding to each prediction step size: in, This represents the current EDH state anchor point used to generate future prediction results. Indicates the future number The predicted change of each prediction step size relative to the current EDH state anchor point.

8. The multi-step prediction method for evaporation waveguide height based on state-anchored gated residuals according to claim 1, characterized in that: In step S7, the training loss function uses the prediction step-size weighted Huber loss as the main loss: in, For the first Loss weights for each prediction step size This represents the Huber loss function. For the first EDH prediction values ​​corresponding to each prediction step size For the first The true value of the EDH for each predicted step size.

9. The multi-step prediction method for evaporation waveguide height based on state-anchored gated residuals according to claim 8, characterized in that: The training loss function also includes a long-term prediction auxiliary loss, which is applied to a preset set of long-term prediction step sizes. Construct a long-term prediction auxiliary loss: The total training loss is obtained by weighting the main loss and the long-term prediction auxiliary loss: in, This represents the weighting coefficient for the auxiliary loss in long-term forecasting.

Images